Word position effects in speech perception Thesis submitted for the degree of Doctor of Philosophy School of Psychology and Clinical Language Sciences
Luca Cilibrasi November 2015
Declaration I confirm that this is my own work and use of all material from other sources has been properly and fully acknowledged
_____________________ Luca Cilibrasi
TABLE OF CONTENTS
CHAPTER 1. INTRODUCTION
CHAPTER 2. THEORETICAL FRAMEWORK AND HYPOTHESES
2. 1 THE INFORMATION PROCESSING MODEL OF RAMUS ET AL. (2010) 2. 1. 1 SURFACE AND UNDERLYING REPRESENTATIONS 2. 1. 2 INFORMATION PROCESSING MODEL 2. 1. 3 ON THE PHONOLOGY OF PERCEPTION: THE PERCEPTION OF SPECIFIED CONTRASTS 2. 1. 4 BOUND MORPHEMES AND SUBLEXICAL LEVEL 2. 2 WORD BEGINNINGS AND WORD ENDINGS 2. 2. 1 WORD BEGINNINGS 2. 2. 2 WORD ENDINGS 2. 2. 3 THE DIFFERENCE BETWEEN WORD BEGINNINGS AND WORD ENDINGS 2. 3 THE IMPORTANCE OF USING MINIMAL PAIRS FOR THE INVESTIGATION OF PERCEPTION OF
25 25 31 35 37 39 39 41 44
PHONOLOGICAL CONTRASTS 2. 3. 1 INTRODUCTION 2. 3. 2 MINIMAL PAIRS IN RESEARCH 2. 3. 3 MINIMAL PAIRS IN SECOND LANGUAGE TEACHING 2. 3. 4 MINIMAL PAIRS IN SPEECH AND LANGUAGE THERAPY 2.4 THE CONCEPT OF SALIENCY 2. 5 HYPOTHESES
45 45 49 53 54 56 58
CHAPTER 3. WORD BEGINNINGS MATTER
3. 1 SUMMARY 3. 2 INTRODUCTION 3. 3 HYPOTHESIS 3. 4 METHODS 3. 4. 1 METHODS: EXPERIMENT I - ITALIAN ADULTS 3. 4. 2 METHOD: EXPERIMENT II - ITALIAN CHILDREN 3. 4. 3 METHOD: EXPERIMENT III – ENGLISH ADULTS 3. 5 RESULTS 3. 5. 1 RESULTS: EXPERIMENT I 3. 5. 2 RESULTS: EXPERIMENT II
59 60 63 65 65 71 75 76 76 79
3. 5. 3 RESULTS: EXPERIMENT III 3. 6 DISCUSSION 3. 7 CONCLUSION
86 90 96
CHAPTER 4. WORD ENDINGS ARE OPTIONALLY SALIENT
4. 1 INTRODUCTION 4. 2 BACKGROUND AND HYPOTHESIS 4.2 METHODS 4. 2. 1.METHOD: EXPERIMENT IV - LEXICAL AND MORPHOSYNTACTIC MINIMAL PAIRS 4. 2. 2 METHOD: EXPERIMENT V - MMN AND LEXICAL/MORPHOSYNTACTIC PAIRS 4. 2. 3 METHOD: EXPERIMENT VI - NONWORDS AND MORPHEME STRIPPING 4. 3 RESULTS 4. 3. 1 RESULTS: EXPERIMENT IV 4. 3. 2 RESULTS: EXPERIMENT V 4. 3. 3 RESULTS: EXPERIMENT VI 4. 4 DISCUSSION
98 99 108 108 109 112 119 119 120 127 131
CHAPTER 5. IMPLICATIONS FOR THE ASSESSMENT OF SPECIFIC LANGUAGE IMPAIRMENT AND DYSLEXIA 139 5. 1 SUMMARY 5. 2 READING AND MODELS OF READING 5. 3 WHEN READING IS IMPAIRED: DYSLEXIA 5. 3. 1 THEORIES FOR DYSLEXIA 5. 3. 2 DYSLEXIA AND PHONOLOGY 5. 4 A FREQUENTLY CO-MORBID DISORDER: SPECIFIC LANGUAGE IMPAIRMENT (SLI) 5. 4. 1 INTRODUCTION 5. 4. 2 SLI AND MORPHOLOGY 5. 4. 3 SLI AS A MORPHO-PHONOLOGICAL DISORDER 5. 4. 4 PHONOLOGY AS A BOOTSTRAP FOR THE MORPHOLOGICAL DEFICIT 5. 4. 5 CLINICAL MARKERS FOR SLI 5.5 ANALYSES OF CNREP AND TEGI 5. 5. 1 CNREP AND WORD POSITION EFFECTS 5. 5. 2 TEGI AND WORD POSITION EFFECTS 5. 6 CLINICAL DATA ANALYSIS 5. 7 CONCLUSION
139 139 149 149 151 153 153 154 155 157 158 160 160 162 167 176
CHAPTER 6. DISCUSSION
6. 1 SUMMARY 6. 2 CRITICAL EVALUATION OF STIMULI AND METHODS 6. 3 MAIN FINDINGS FROM THE PHD 6. 4 THEORETICAL CONTRIBUTION 6. 4. 1 FINDINGS CONSISTENT WITH OUR HYPOTHESES 6. 4. 2 FINDINGS NOT CONSISTENT WITH OUR HYPOTHESES
178 178 183 186 186 188
6. 5 IMPLICATIONS OF OUR FINDINGS FOR RESEARCH 6. 6 IMPLICATIONS FOR CLINICAL PRACTICE 6. 7 LIMITATIONS AND FUTURE DIRECTIONS 6. 8 FINAL REFLECTIONS
189 190 192 195
APPENDIX 1 APPENDIX 2 APPENDIX 3 INTRODUCTION DESCRIPTION OF BAAYEN’S MODEL SIMULATION WITH OUR STIMULI INSIGHTS FROM THE MODEL POTENTIAL ISSUE CONCLUSION APPENDIX 4
231 235 242 242 242 243 245 246 246 247
LIST OF FIGURES Figure 1.1 The model of phonology used as a reference in this PhD (Ramus et al., 2010). ......................................................................................................................... 18 Figure 1.2 Schematic representation of the set of experiments conducted in this PhD. .................................................................................................................................... 24 Figure 2.1 Generative and Optimality theories gap with perception. ........................ 29 Figure 2.2 Information Processing Model according to Ramus and colleagues in its first version (Ramus et al., 2001). .............................................................................. 33 Figure 2.3 An up to date version of the Information Processing Model of Ramus et al. (2010). ................................................................................................................... 34 Figure 2.4 Prototypical syllabic structure. ................................................................. 44 Figure 2.5 Schematic representation of the Information Processing Model proposed by Ramus et al. (2010). Input sublexical representations, circled in red, are the focus of the research conducted in this PhD. ....................................................................... 58 Figure 3.1 Mean comparison for Italian adults’ accuracy.......................................... 78 Figure 3.2 Scatterplot of the correlation between accuracy in the perception task and performance in the reading task in seconds. .............................................................. 81 Figure 3.3 Scatterplot of the correlation between accuracy in the production task and performance in the reading task in seconds. .............................................................. 82 Figure 3.4 Production: comparison of the means in the four conditions. .................. 83 Figure 3.5. Perception: comparison of the means across stressed and unstressed, and initial and medial clusters........................................................................................... 84 Figure 3.6 Analysis of reaction times. ....................................................................... 89 Figure 4.1 Reaction times for the discrimination of elements in lexical and morphosyntactic minimal pairs. ............................................................................... 119 Figure 4.2 Example of sound wave. The flat line in the end of the first word indicates the beginning of the plosive consonant and thus the disambiguating point. ............ 121 Figure 4.3 Average of peaks. ................................................................................... 124 Figure 4.4 Averages of mean amplitudes. ................................................................ 125 Figure 4.5 Example of MMN over time in the frontal region with morphosyntactic minimal pairs............................................................................................................ 126 Figure 4.6 ERP components elicited by the four conditions in the 300ms following the disambiguating point, averaged in the cluster of frontal electrodes. .................. 127 Figure 4.7 Reaction times for the discrimination of elements with and without morphological information (+ control). ................................................................... 129 Figure 5.1 The dual route cascaded model of reading aloud by Coltheart et al. (2001) .................................................................................................................................. 145 Figure 5.2 Pictures of this type are used in the Rice & Wexler (2001) assessment to elicit words ending in the phoneme /s/..................................................................... 163 Figure 5.3 Pictures of this type are used in the Rice & Wexler (2001) assessment to elicitate third person verbs ending in the phoneme /s/. ............................................ 164 Figure 5.4 Accuracy (number of errors) in 5 syllable words. .................................. 172 Figure 5.5 Baddeley’s (1992) graphic representation of the phonological loop. ..... 176 Figure 7.1 Activation of the pairs cared/cares and side/size. ................................... 245
LIST OF TABLES Table 3.1 Example of stimuli for experiment 1. The four conditions are presented together with their contrasting items. ......................................................................... 69 Table 3.2 Example of stimuli for the production task in experiment 2. The four conditions are presented. ............................................................................................ 73 Table 3.3 Example of stimuli for the perceptual task. ............................................... 75 Table 3.4 Example of stimuli for the English perceptual task. .................................. 76 Table 3.5 Descriptive Statistics for experiment 1, including min and max values for accuracy in the four conditions, as well as mean, SD and SE. .................................. 76 Table 3.6 Post-hoc analyses. ...................................................................................... 78 Table 3.7 Descriptive statistics for standardised tests and age (obtained from raw scores). ....................................................................................................................... 79 Table 3.8 Min, max and mean accuracy values are reported for perception, as well as Standard Deviation and Standard Error. . .................................................................. 80 Table 3.9 Min, max and mean accuracy values are reported for production, as well as Standard Deviation and Standard Error.. ................................................................... 80 Table 3.10 Descriptive Statistics for experiment 3, reporting max and min values for accuracy, mean, Standard Deviation and Standard Error. ......................................... 86 Table 3.11 2x2 ANOVA on English adults accuracy. ............................................... 87 Table 3.12 Examples of measure of distance between word beginning and uttering of a medial cluster. ......................................................................................................... 87 Table 3.13 Experiment 3 reaction times descriptive statistics. .................................. 88 Table 3.14 Post-hoc comparisons experiment 3. ....................................................... 89 Table 4.1 Hypotheses based on two theoretical positions are presented for each experiment. ............................................................................................................... 107 Table 4.2 Examples of lexical and morphosyntactic minimal pairs. ....................... 109 Table 4.3 Lexical and morphosyntactic minimal pairs used in the EEG experiment. .................................................................................................................................. 111 Table 4.4 Pairs with and without potential morphological information. ................. 115 Table 4.5 Descriptive statistics experiment 6. ......................................................... 127 Table 4.6 Correlations item based reaction times and PSF, BSF (obtained with British and American corpora) and Rhyme likelihood (obtained with British Corpora). .................................................................................................................. 130 Table 5.1 Distribution of non-initial clusters in the Children’s Test of Nonword Repetition, (CNRep), (Gathercole & Baddeley, 1996). ........................................... 161 Table 5.2 Distribution of word final clusters in the Test of Early Grammatical Impairment, (TEGI)(Rice & Wexler, 2001). .......................................................... 165 Table 5.3 Descriptive Statistics for children with a clinical condition. ................... 169 Table 5.4 Descriptive Statistics (based on raw number of errors) group 1. SLI + dyslexia. ................................................................................................................... 169 Table 5.5 Descriptive Statistics (based on raw number of errors) group 2. SLI. ..... 169 Table 5.6 Descriptive Statistics (based on raw number of errors) controls. ............ 169 Table 5.7 Five syllable nonwords used for the clinical data analysis. ..................... 170 Table 5.8 Post-hoc analysis between groups............................................................ 173 Table 5.9 Post-hoc analysis within groups. .............................................................. 173 Table 7.1 Orthographic productive rules used ......................................................... 231
Table 7.2 Examples of words used in DDE. ............................................................ 232 Table 7.3 Coloured Progressive Matrices standardised scores. ............................... 233 Table 7.4 Specifically designed stimuli: nonwords. ................................................ 234 Table 7.5 Real words used in experiment 4 ............................................................. 235 Table 7.6 Nonwords used in experiment 6. ............................................................. 236 Table 7.7 Block 1. .................................................................................................... 237 Table 7.8 Block 2. .................................................................................................... 238 Table 7.9 Block 3. .................................................................................................... 239 Table 7.10 Block 4. .................................................................................................. 240 Table 7.11 Block 5. .................................................................................................. 241 Table 7.12 Information on children with a clinical condition. ................................. 247
ACKNOWLEDGMENTS There are many people I would like to thank, and this comes as no surprise. Now that I am close to the conclusion of this PhD, I realise that these three years have been the most challenging period I have ever experienced. First of all, I would like to thank my supervisor, Professor Vesna Stojanovik. It is not easy to decide where to start in explaining how she contributed to the completion of this work. The trust she had in my ideas is probably what convinced me most to keep going, and the incredible amount of time she spent on my work is a rare perk in academic supervision. Her insights about theoretical and clinical linguistics guided my work, and contributed to shaping the linguist I am today. I am extremely grateful to my second supervisor, Professor Patricia Riddell. In the several meetings we had she contributed with exacting suggestions for making the research neat. Further, her insights revealed gaps and problems in my work that I did not spot at first. Her valuable experience contributed in shaping my work pragmatically, and gave it a psychological flavour that it would not have had otherwise. I am extremely grateful to my project monitors: Professor Theo Marinis and Professor Doug Saddy. Professor Theo Marinis was incredibly important during the first half of my PhD in helping me find the right structure for my project. At that time, my desire for exploration was poorly structured, and the experiments I had in mind were disjointed and lacked coherency. Theo spent more of the time he was required to in helping me narrow down the right questions and the right approach in facing them.
Professor Doug Saddy was incredibly important during the second half of my PhD. He was a great advisor in the design of the EEG/ERP portion of my PhD, and an important contributor in the design and interpretation of the final experiment of my thesis, which I personally find the most interesting of the set. Several other academics contributed greatly to the realisation of this work. Among them, special mention goes to Dr Vitor Zimmerer. Our discussions on the clash between generative and usage-based grammars were an invaluable resource of inspiration for my work. His lectures on language processing helped shape my experiments, and our discussions on the structure of my thesis helped me individuate the path I wanted to follow. Dr Matt Moreland was a great contributor when it came to creating experiments that were phonologically controlled. His excitement for my final experiment made me understand that it was the right thing to do. I am very grateful to Dr Simona Mancini and Dr Nicola Molinaro, who hosted me for a research visit at the Basque Centre on Cognition, Brain and Language, and helped me in the design of my EEG/EPR task. I am extremely thankful to Professor Luigi Rizzi and Professor Adriana Belletti, for their precious comments on the PhD and even more importantly for introducing me to linguistics during my BA in Siena. I am very grateful to Professor Alfonso Caramazza, who gave me the opportunity to teach in the Harvard Summer Program in Mind/Brain sciences, which proved to be a wonderful human and academic experience. I am extremely thankful to Professor Lisa Archibald and Professor Susan Gathercole, who provided the raw data on typically developing children I use as a baseline in the
chapter on the clinical implications of my findings. Without their data, the chapter would look extremely impoverished compared to what it is. I am also extremely thankful to Professor Harald Baayen. The time he spent running simulations with my data using his model is a wonderful gift for my intellectual exercise and for the interpretation of the data in this thesis. Dr Claudio Pacchierotti helped me on several occasions with mathematical issues, and helped me better understand (a bit of) the functioning of MATLAB. I am grateful to Dr Francesca Foppolo and Professor Teresa Guasti, who invited me to present at the Milan Bicocca Spring School on Language, and made my short stay in Milan extremely pleasurable. I am grateful to all the friends who sustained me during my stay in Reading. This includes the PhD crew, as well as the friends I made along the way who are now spread around the world. Likewise, I am grateful to my friends back in Italy, who are always there and, I hope, will always be there, no matter how far our experiences will bring us. I am deeply thankful to my girlfriend, who probably more than anyone else suffered the breakdowns of my final year, and managed to calm me down and sustain me until the end. Finally, I am very thankful to the members of my family. Although they may occasionally lose track of what I am doing, their long standing trust in my judgment always made me believe I could make my own choices. Looking back, I see that, with all the contradictions that are part of a family, it is because of them that I am who I am, and for that I am happy.
Linguistic and psycholinguistic studies show that different positions in the word are processed differently. Word beginnings are salient, i.e. the most important part of the word concerning processing, and word endings are salient only when carrying morphological information (otherwise they are non-salient and they are subject to disruptive processes such as deletion). Little is known about the presence of these effects in the sublexicon, particularly regarding input representations (i.e. perception). This project aims to fill this gap. With a set of 6 experiments the project investigated the nature of sublexical word position effects in perception. The first 3 experiments focused on nonword beginnings, the last 3 experiments focused on word and nonword endings. Using minimal pair discrimination tasks, the results of the experiements showed that: 1. Consonant clusters in nonword initial positions are detected more accurately when compared to consonant clusters in nonword medial positions. This effect was found to hold in Italian school aged children (aged between 8;03 and 10;01) and in Italian adults, but these effects were consistently observed only when the consonant clusters were in unstressed syllables. Similar results were found in English adults, who required more time to discriminate consonant clusters when the clusters were in the initial part of the nonword compared to clusters in the medial part of the nonword. The result obtained with English adults suggests that a larger amount of processing “resources” is spent for the initial part of the nonword compared to the medial part of the nonword. The fact that the results of these 3
experiments are obtained with nonwords suggests that the phenomena detected are sublexical. 2. Real word endings require more time to be processed when they carry morphological information. This was shown behaviourally with a reaction time experiment and was also reflected in neurophysiological activity by the fact that larger brain components (specifically, larger Mismatch Negativities – MMNs) were observed when participants had to discriminate items that did carry morphosyntactic information than when participants had to discriminate items that did not carry morphosyntactic information. Further, reaction time differences were also observed in a third experiment that used nonwords rather than real words, with participants taking longer to discriminate nonwords that contained potential morphosyntactic information than to discriminate nonwords without morphosyntactic information. The results of the final experiment suggests that inflection morphemes may be detected sublexically. Finally, we investigated the presence of word position effects in English children with dyslexia and specific language impairment (SLI) and in typically developing children by analysing previously collected data with the test Children’s test of nonword repetition, and we observed that word position effects may also be present in clinical groups. Specifically, the clinical data analysis showed that nonwords containing non-initial clusters were repeated less accurately than nonwords that do not contain non-initial clusters by children with SLI (however, nonwords with non-initial clusters contained more clusters overall, so given the data available it is not possible to assess whether the result observed is a consequence of a word position effect or a consequence of phonological complexity).
In conclusion, the experiments conducted in this PhD project suggest that the word position effects we investigated (saliency of word beginnings and optional saliency of word endings) are active sublexically in perception, but they also show that the effects can be influenced by other variables. For instance, results in experiments 1 and 2 were obtained only when the phonological clusters discriminated were unstressed and the results in the clinical data analysis may be influenced by the presence of clusters per se, rather than by the position of the clusters.
CHAPTER 1. INTRODUCTION 1. Introduction
There is a substantial amount of linguistic and psycholinguistic evidence suggesting that different positions in the word are processed differently: Word initial syllables are described as strong, since they license a large number of contrasts and resist reduction (Beckman, 1998, 2013, Smith, 2004) and non-initial syllables are described as weak, since they license a smaller number of contrasts and tend to reduction (Beckman, 1998, 2013, Harris, 2011). As Beckman (1998, 2013, p.49) explains: Word initial syllable onsets permit a wide range of marked segments (i.e. segments violating certain constraints), trigger directional phonological processes, and resist the application of otherwise regular alternations. Further, they play a prominent role in lexical access (Zwisterlood, 1989, Marslen-Wilson & Zwisterlood, 1989, Pitt & Samuel, 1995) and are more likely to be recalled in the Tip of the Tongue phenomenon (Browman, 1978). Word final material, on the other hand, is subject to deletion. For instance, when the end of the word is followed by a stressed syllable, as it occurs for /d/ in “hand-bag” (Harris, 2011), the /d/ in the word ‘hand’ is omitted. Word final material also has worse priming effects on neighbours than word initial material (Marslen-Wilson & Zwisterlood, 1989). However, the presence of morphology can modify these effects. When it comes to verb inflection, word endings resist phonological processes that are otherwise active (Pater, 2006), such as deletion (Harris, 2011), and entail merging processes (Padovan, 2013). Given the above, it could be summarised that word
beginnings are inherently strong (Beckman, 1998, 2013) and word endings are optionally strong (Pater, 2006). Most research on word position effects has focused on the lexicon and on position effects on lexical access (Brown & McNeill, 1966, Browman, 1978, Cole & Jakimik, 1980, Marslen-Wilson, 1984, Nooteboom, 1981). Only a few studies have focused on position effects in the sublexicon (Pitt & Samuel, 1995, Marshall & van der Lely, 2009). The distinction between the lexicon and the sublexicon is assumed by several different models (Ramus, 2001, Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001, Chomsky & Halle, 1968, Prince & Smolensky, 1997, Morton, 1969, Norris, McQueen & Cutler, 2000). The classic generative theory distinguishes between underlying and surface representations (Chomsky & Halle, 1968, 1990). Underlying representations are stored forms of words, in which some phonological traits are underspecified. Surface representations are the result of the application of the phonological rules of the language on the underlying representations. Optimality Theory develops this idea and distinguishes between lexicon and post-lexicon (Prince & Smolensky, 1997). With the term “post-lexicon,” supporters of optimality theory refer to the output form of a given word, after a set of constraints has been applied on the lexicon. This level of representation explicitly represents a development of the concept of surface representations (Kager, 1999). At the same time, the concept of lexicon is very similar to the concept of underlying representations (Prince & Smolensky, 1997). Both generative and optimalist models naturally describe a unidirectional process, production, and overlook perception. This problem was noticed by Ramus and colleagues (2010) who developed an Information Processing Model (IPM) based on generative theory, but that also takes into account
perception. The IPM (Ramus et al., 2010) assumes the existence of a lexicon and the existence of a sublexicon. The former contains prototypical word forms, the latter contains information on the phonological rules to be applied in perception and production to map the speech with these prototypical forms. Ramus et al. (2010) explicitly state that their distinction between lexicon and sublexicon corresponds to the generative distinction between underlying and surface representations (Chomsky & Halle, 1968)1 but that this distinction accounts only for the output pathway of their model (Ramus et al., 2010). The input pathway is described using a new level of representation that they call Input Sublexical Representations – ISR (see Figure 1.1below from Ramus et al., 2010). The ISR is tuned during language acquisition, and contains a mapping of the phonemes of a given language and information regarding relevant and irrelevant contrasts. Word position effects are rarely studied at the perceptual level (Pitt & Samuel, 1995, Marshall & van der Lely, 2009), and studies on word position effects in perception at the sublexical level are even rarer (Pitt & Samuel, 1995). This PhD aims to fill these gaps. The need for this type of research is theoretical and clinical. From a theoretical point of view, the lack of research in this domain corresponds to a substantial lack of understanding of the sublexical processes related to word position. Sublexical processing is, however, very important, because it provides a domain of research of linguistic phenomena relatively independent from semantics. From a clinical point of view, the lack of research in this domain corresponds to a poor understanding of some of the phenomena observed in clinical assessments with sublexical items. A good
Not very intuitively underlying representations correspond to the lexicon and surface representations correspond to the sublexicon (Ramus et al., 2010),
understanding of sublexical phenomena is however crucial in the clinical context, considering the volume of assessments and interventions that use nonwords (i.e. assessments and interventions that rely on items that tap into the sublexicon). In all our tasks we will use minimal pairs, which are widely used in clinical assessments and in speech and language-therapy protocols. Minimal pairs discrimination has been extensively studied both at the lexical (Dehaene-Lambertz, Dupoux & Gout, 2000, Dufour, Nguyen & Frauenfelder, 2007) and at the sublexical level (Hoonhorst et al., 2011, Raphael, 1972, Liberman, Harris, Hoffman & Griffith, 1957, Zlatin & Koeningsknecht, 1975). However, a systematic study of the effect of the position of the contrast in the word has never been attempted.
Figure 1.1 The model of phonology used as a reference in this PhD (Ramus et al., 2010).
Permission to use Figure 1.1 was generously granted by Franck Ramus. The following is the description from Ramus et al., 2010, page 313:
An information processing model of speech perception and production. Arrow 1a corresponds to output phonological processes, 2a to output phonetic implementation. 1b corresponds to phonological parsing – inverse phonology – and 2b to perceptual phonetic decoding. In the adult, these four processes are finely tuned to the phonological and phonetic properties to the maternal language-s. They may be mistuned during – first or second – language acquisition, or in cases of brain lesions or language disabilities. Our hypothesis is that two word parts (word beginnings and word endings) are perceptually salient, but in different ways and for different reasons: Word beginnings are inherently salient, and this is due to the psycholinguistic salience of material presented “first” (Beckman, 1998, 2013, Smith, 2004, Marshall &van der Lely, 2009). The rationale for this hypothesis is theoretical: the constraint proposed by Beckman applies in production, but, according to Ramus et al. (2010), we must assume similar principles to apply in perception too. Word endings, at the same time, are optionally salient, meaning that they become salient only if they are carrying potential morphosyntactic information (Pater, 2005, Padovan, 2013). The rationale for the second hypothesis is based on data from lexical studies (Kuriso, 2001, Pater, 2006) and on the idea, proposed recently by Grainger and Ziegler (2011), that inflection morphemes are detected sublexically. All experiments in this PhD are based on variations of the same methodology, which consists of the study of accuracy and reaction times in the discrimination of elements belonging to minimal pairs. Minimal pairs are pair of words which differ in only one phonological element and have a different meaning (Roach, 2000). (At the sublexical level, meanings are not activated (Ramus, 2001, Coltheart et al., 2001), so
we refer to minimal contrasts as pair of nonwords which differ in only one phonological element (Wedel, 2012)). The use of minimal pairs is motivated by methodological and clinical reasons. From a methodological point of view, minimal pairs allow for a coherent study of word position since the contrast can be moved in different positions in the word, complexity can be controlled changing the nature of the contrasts, and sublexical representations can be tested with the use of nonwords. From a clinical point of view, a better understanding of the processing of minimal pairs is desirable, considering the fact that minimal pairs are widely used in the assessment and remediation of phonological disorders (Pulvermüller et al., 2001, Crosbie, Holm & Dodd, 2005), particularly of dyslexia (Blache, Parsons & Humphreys , 1981, Barlow & Gierut, 2002). Consequences of our findings in the assessment and the remediation of dyslexia will be discussed in the last part of the dissertation. Results confirming these hypotheses would be relevant for two reasons: first, the psycholinguistic salience of word beginnings is described as a constraint that applies to the lexicon (Beckman, 1998, 2013, Prince & Smolensky, 1997). However we would show that the constraint (if we want to keep this terminology) applies sublexically, in perception, to any linguistic material, not just to the lexicon. Second, phonological and lexical models (Chomsky & Halle, 1968, Prince & Smolensky, 1997, Morton, 1969, Ramus et al., 2010, Coltheart et al., 2001) do not predict inflectional morphemes to be recognised at the sublexical level, but there is some evidence that this may actually happen (Grainger & Ziegler, 2011). We will explicitly address this question.
Directly comparing word beginnings and word endings is not a sound way to proceed, however, for two reasons. First, word beginnings refer to syllable onsets (Marshall & van der Lely, 2009), while morphosyntactic information is carried in the last syllable coda (Berko, 1958, Law & Strange, 2010). Second, word beginnings are supposed to be salient per se (Beckman, 1998, 2013), while word endings are supposed to be optionally salient, i.e. only if potential morphosyntactic information is present (Harris, 2011, Pater, 2005). As such, a direct comparison is not possible. 1) In order to test the two hypotheses, six experiments were carried out: three concentrating on word beginnings, three concentrating on word endings. Further, implications for clinical linguistics were discussed analysing data from children with language difficulties in light of the findings of our experiments. The list of experiments is summarised below, together with a short description of the findings. In the first experiment, we tested the discrimination of contrasts in the onset of the first syllable and contrasts in the onset of the second syllable (developing the study from Marshall & van der Lely, 2009). For this experiment, 22 adult native speakers of Italian were recruited. Subjects were asked to discriminate auditorily presented minimal pairs of nonwords differing in one phoneme, and specifically on the voice feature. Contrasts could be positioned in the onset of the first syllable or in the onset of the second syllable. Results show that Italian adults made a significantly larger number of errors on second-last syllables than word-initial syllables. 2) In order to understand if the finding is valid across development and to explore significant differences across development, the study was performed again on 34 Italian children. Results show that the effect of contrasts position is present in
school aged children (8 to 10), but only if the syllable is not stressed. Children made more errors in word medial syllables than in word initial syllables if the syllables were not stressed. 3) In order to find out if the finding is valid cross-linguistically, as Beckman’s (1998, 2013) proposal would suggest, the study was performed on 30 adult native speakers of English. Nonwords were manipulated so that wordlikeness was respected for English (only vowels were changed). The results show that word position effects are active in English as well, although some issues in the analysis emerged. 4) In order to test saliency-optionality of sublexical word endings, we compared the discrimination of lexical and morphosyntactic pairs. We conducted a RTs experiment on 20 English adults. Subjects were asked to discriminate elements in lexical and morphosyntactic minimal pairs. The result shows that participants were quicker in discriminating elements that do not carry morphosyntactic information. 5) The previous experiment was extended by a follow up experiment with Event Related Potentials (ERPs). Minimal pairs were presented with the Oddball paradigm. Elements carrying morphological information generated a larger Mismatch Negativity (MMN), a brain component that is enhanced when elements are perceived as “more” different. The experiment was conducted on 22 English adults. 6) With the last experiment we wanted to understand if the presence of potential morphosyntactic features is detected and processed differently by subjects when using sublexical items. In order to do that, we compared the discrimination of
nonwords differing in phonemes that do not carry any morphosyntactic information, and nonwords differing in phonemes that carry morphosyntactic information. This study was performed on 20 adult native speakers of English. Results show that it takes longer to discriminate elements that may carry morphological information. 7) With this set of experiments we showed that, at the sublexical level, word initial positions are perceptually salient, and word final positions are optionally salient. Nonwords (sublexical items) are widely used in the assessment and in the remediation of language disorders, particularly dyslexia. As such, in the last part of this PhD we discussed the consequences of our findings in the assessment and remediation of language impairment. In order to do that, we tested the prediction made by our findings in experiments 1-6 by including a sample of children with developmental language impairment (some children had comorbid reading difficulties). The analysis showed that some linguistic patters observed in children with language impairment may be a consequence of word position effects. A schematic summary of the six experiments and the clinical analysis is provided in Figure 1.2
Experiment 1: word initial vs word medial clusters. Tested on Italian adults
Experiment 4: lexical vs morphosyntactic minimal pairs. Tested on English adults
Experiment 2: word initial vs word medial clusters. Tested on Italian children
Experiment 5: lexical vs morphosyntactic minimal pairs. Tested on English adults (follow up with ERPs)
Experiment 3: word initial vs word medial clusters. Tested on English adults
Experiment 6: morphological vs non-morphological nonwords. Tested on English adults
Clinical data analysis: data from children with Language Impairment were analysed in the light of the findings in experiments 1-6
Figure 1.2 Schematic representation of the set of experiments conducted in this PhD.
CHAPTER 2. THEORETICAL FRAMEWORK AND HYPOTHESES 2. Theoretical framework
2. 1 The information processing model of Ramus et al. (2010) 2. 1. 1 Surface and underlying representations The phonological model used in this PhD is the model of Ramus et al. (2010). The model retains a distinction that was proposed in previous models, and introduced by Chomsky and Halle (1968): the distinction between underlying and surface representations. In order to present this distinction, a few historical facts need to be reported. The theoretical framework of reference for the model of Chomsky and Halle (1968) is known as generative linguistics. Generative linguistics is a branch of linguistic theory born during the 1950’s, after the seminal work of the American linguist Noam Chomsky (Carnie, 2013). For generative linguists, grammar is a set of specified rules, and the aim of linguistic research is understanding these rules. Early work by generative linguists (Chomsky & Halle, 1968) distinguished between at least two levels of representation of words (this is with reference to phonological representations). In other work, Chomsky (1957) discusses the existence of two levels of representation in syntax: surface and underlying. This distinction is crucial for the experiments presented in this PhD thesis, and is maintained, in a modified form, in the model we use as a reference (Ramus et al., 2010). The nature of this distinction is described below. In English, there are many words in which the only difference is the point of articulation of the plosive consonant (Kenstowicz, 1994, p.57). As in many other languages, the point of articulation of plosive consonants is said to be “distinctive” in
English (Kenstowicz, 1994, Roach, 2000). Not all features are distinctive, however. Some of them are redundant, meaning that they do not lead to change in meaning, and the contrasts they generate are complementary and can be predicted. This is the case, for instance, for the aspiration feature, formalised with the term in brackets: [spread gl]. The presence of aspiration respects distributional rules that do not lead to changes in meaning. The rules can be summarised as follows for English (Kenstowicz, 1994, p.58): a. All segments except for voiceless stops are [- spread gl] b. /ph/, /th/, /kh/ only appear syllable-initially c. /p/,/t/,/k/ do not appear syllable initially, but freely occur in other positions in the syllable. Some features are thus unpredictable while others are predictable. How is this system represented in our mind? Generative theory (Chomsky & Halle, 1968) proposes a distinction between two levels of representations, surface and underlying representations. Underlying representations (also called phonological representations) contain all and only the unpredictable distinctive information for each lexical item. A set of rules then derives a different form of the word based on the predictive distribution of the grammar, and generates surface representations (also called phonetic representations) that correspond to the actual pronunciation of the word (Kenstowicz 1994, p.60). The case of aspiration is quite elegant, and allows for strict formalisation. However, this is not the case for all processes with regard to surface and underlying representations (Kager, 1999). In certain cases, strict formalisation shows its limits
and different solutions have been proposed (Prince & Smolensky, 1997, Ramus et al., 2001). One of the recent developments of generative phonology is Optimality Theory (OT) (Prince &Smolensky, 1997). The OT model relies on the distinction between underlying and surface representations, as proposed by generative linguists, but describes in different terms the relation between the two levels. In generative theory, the two levels are linked via a set of explicit rules, which describe the changes from one level to the other. As Goldsmith and Laks (2012, p.20) explain: In OT, the explicit rules are substituted with an algorithm that selects surface representations from a wide set, choosing the representation that best satisfies an ordered set of constraints. Constraints are assumed to be universal, to make the model plausible in terms of language acquisition, but the ranking of the constraints varies according to the language. OT and generative theory also differ in that, while in generative theory an extremely detailed description of rules was desirable, in OT constraints are described in vaguer terms. Thus, even if the distinction between the two levels is accepted (Ramus et al., 2010) and grounded also by clinical data (Goldrick & Rapp, 2007), the exact nature of these representations is debated (Steriade, 1995). In a recent discussion of the issue, Ramus et al. (2010, p.316) simply assume that underlying representations are abstract in the sense that they do not include complete phonetic specifications of the word forms. This is sufficient for the needs of the current research project and is therefore the point of view adopted in this dissertation. OT distinguishes between markedness and faithfulness constraints (Kager, 1999, Prince & Smolensky, 1997, Jusczyk, Smolensky & Allocco, 2002). A structure that
violates a markedness constraint is said to be marked for that specific feature. For instance, all syllables ending with a coda violate the NO-CODA markedness constraint, and are said to be marked with the coda features (Jusczyk et al., 2002). The concept of markedness is not new and was originally developed by the Prague School of Linguistics (Jakobson, 1962, Trubetzkoy, 1969). Marked forms are suboptimal (an optimal form violates the smallest possible number of constraints) and tend to be avoided in speech (Jusczyk et al., 2002). How do we explain, then, the fact that they are often present in speech? For instance, if having a coda violates the NO-CODA constraint, and thus having a word ending with a coda is having a suboptimal form, why do so many words end with codas? According to Optimality Theory, faithfulness constraints are the answer to this question. Faithfulness constraints require underlying representations to resemble faithfully surface representations (Beckman, 1998). When a faithfulness constraint outranks a markedness constraint, the markedness constraint can be violated. If the constraints are ranked in the opposite order, the surface representation does not violate the markedness constraint (Jusczyk et al., 2002). Thus, in a word ending with codas we have 1) an underlying form that contains a coda, 2) a faithfulness constraint outranking a markedness constraint. In processing terms, the output of the lexical phonology is the input to the postlexical phonology. This means that the information processed in the post-lexical phonology does not play any role in any rule or generalization in the lexical phonology (Goldsmith & Laks, 2012). OT and generative theory, thus, account for a unidirectional process,
from stored lexical representations to post-lexical
representations. They precisely describe the nature of modifications taking place
between stored form of words and uttered versions of them. They formally acknowledge the mapping from abstract forms to forms that are pronounced by the speaker in everyday conversation, taking into account variations related to the context and the register, for instance. In doing this, they describe the process of production but ignore the process of perception (Figure 2.1). OT and classic generative theory are thus not appropriate for this PhD project, and a different model needs to be used. The model we chose is the Information Processing Model developed by Ramus, Peperkamp, Christophe, Jacquemot, Kouider, Dupoux (2010).
Figure 2.1 Generative and Optimality theories gap with perception.
The level of input sublexical phonological has mostly been overlooked (Ramus et al., 2010). However, some authors focused their research on that level: Darcy, Peperkamp and Dupoux (2007), Gaskell and Marslen-Wilson (1996) and Darcy, Ramus, Christophe, Kinzler & Dupoux (2009) showed that speakers of English and French are unable to perceive assimilation processes when presented
with on-line speech perception. However, when assimilation is presented out of context, subjects are able to perceive the contrast. This suggests that there is a set of phonological rules active at the level of input sublexical phonological representations that specifies which contrasts are relevant during speech perception, and facilitates the access to lexical forms. This level of representation has received less attention than the level of output sublexical phonological representations (Eisner 2002; Boersma 1998).
2. 1. 2 Information processing model The original model of Ramus (2001) (see Figure 2.2) presents an information processing model of lexical access that takes into account the
unidirectionality, and offers a bidirectional analysis of representations. The model of Ramus (2001) distinguishes between lexical and sublexical representations. Sublexical representations, containing information on relevant and irrelevant contrasts in a given language and a detailed description of phonemes, is directly linked to the level of orthographic sublexical representations. At this level, the phonological route of the classic dual-route model (Coltheart et al., 2001) is conceived, and graphemes are associated with phonemes. The lexicon is composed of three sub-modules: the phonological, the orthographic and the semantic lexicon. Strings of sounds, strings of graphemes, and sets of meanings are linked together, connected and at the same time separated. The directions of the arrows explain how each module influences the others. Sublexical representations are shaped by the phonological lexicon and by the sublexical orthographic representations. This idea is in line with studies coming from different approaches. Cheng, Schafer and Riddell (2014), for instance, showed that information on phonotactic probabilities is embedded sublexically. Comparing the processing of real words and nonwords (with and without neighbours) using EEG, the authors found that participants are able to discriminate nonwords with neighbours and nonwords without neighbours in an amount of time that is not sufficient to access the lexicon (suggesting that this distinction is taking place sublexically). Thus, information on phonotactic probabilities is extracted from the lexicon “off line”, but becomes available in the
sublexicon when processing speech2, together with the information on which contrasts are relevant (and which are not) in the language (the relevance of contrasts, similarly, is extracted making comparisons between lexical items, as it will be explained in section 2.3.1). Sublexical representations, conversely, shape sublexical orthographic representations and influence the acquisition of the phonological lexicon (in that they distinguish possible new words from unlikely new words). The orthographic lexicon shapes the phonological lexicon, and is shaped by it. At the same time, lexical and sublexical orthographies influence each other (Figure 2.2).
In other words, phonotactic probabilities are a direct consequence of the distribution of phonemes in the lexicon, but once information on phonotactic probabilities is extracted from the lexicon, it becomes available without the need of accessing the lexicon.
Figure 2.2 Information Processing Model according to Ramus and colleagues in its first version (Ramus et al., 2001).
In this model, the term “sub-lexical phonological representations” is explicitly used as a synonym of “surface representations”. However, this model does not account for differences between input and output sublexical representations, and a more detailed distinction revealed to be necessary in subsequent work (Ramus et al., 2010). In a more recent version of the Information Processing model, Sublexical Representations are distinguished in Input Sublexical Representations and Output Sublexical Representations (Ramus et al., 2010, Figure 2.3).
Figure 2.3 An up to date version of the Information Processing Model of Ramus et al. (2010).
Lexical Phonological Representations and Output Sublexical Representations explicitly represent the classic distinction between Underlying and Surface Representations, retained in Optimality Theory as the distinction between Lexicon and Post-lexicon. What is new here, however, is that the model accounts for inverse phonology (i.e. rules describing perception) and defines a partially autonomous set of rules and representations to account for it. The nature of the distinction between input and output sublexical representations is of great interest. As the authors point out, these two levels may be virtually indistinguishable during adulthood (Ramus et al., 2010). However, there is evidence to suggest that there may be a distinction between the two. The strongest evidence comes from neuropsychology, and specifically from a pathology known as conduction aphasia (Caramazza, Basili, Koller & Berndt, 1981). Patients with this condition have relatively intact speech perception and production, but are impaired in
the repetition of speech (i.e. the connection between input and output sublexical representations is impaired). As an emblematic example, Jacquemot, Dopoux and Bachard-Levi (2007) reported the case of an aphasic patient who could perceive nonwords but could not repeat them. As Ramus et al. (2010, p.315), this type of dissociation can only be accounted for if we assume dissociation between the two levels of representation. While the two levels may appear indistinguishable in monolingual adults, the two systems may be quite different during second language acquisition. For instance, Sheldon and Strange (1982) showed that late bilinguals are more proficient in the production of foreign contrasts than in their perception. The shape of input representations is thus strongly language dependent, and is shaped during language acquisition according to the contrasts present in the native language (Kuhl, Williams, Lacerda, Stevens & Lindblom, 1992). As the authors point out, the relation between input and output sublexical representations is poorly understood, and, similarly, it is not clear how these levels relate with lower (lexical) and higher (acoustic) levels of representation (Ramus et al., 2010). What is clear, however, is that levels are not completely independent from each other, but neither completely mirror each other (Ramus et al., 2010). The next section will consider an issue with regard to the perception of phonological contrasts. 2. 1. 3 On the phonology of perception: The perception of specified contrasts The definition adopted for describing underlying representations is the one from Ramus et al. (2010), page 316: Underlying (or “lexical”) representations are abstract in the sense that they do not include complete phonetic specifications of the word forms.
The under-specification of allophones and of assimilation features belong to the level of underlying representations. The perception of specified contrasts, instead, belongs to the input sublexical level (Ramus et al., 2010). The nature of the perception of specified contrasts has been widely studied. It appears that humans are provided with a biological instinct of phonological categorization (Munson, Edwards & Beckman, 2012). In the first few months of life, infants are able to perceive many consonant contrasts (Miller & Eimas, 1983). The contrasts they can perceive are not language-specific and, in fact, non-human animals can also perceive these same contrasts (Kuhl & Miller, 1975). However, perception quickly changes in humans. Kuhl et al.., (1992) showed that six months old infants prefer listening to vowels that are more similar to native-language vowels than ones that are not. This is considered one of the first language specific successes of children. By about nine months of age, children are not able anymore to perceive contrasts that are not in their native language (Werker & Tees, 1984). The structure of phonological information seems to be largely independent from meaning. Navarrete and Costa (2005) performed an experiment that shows the dissociation of phonology from semantics at one of the levels in which this dissociation may not seem intuitively present: the lexical level. The authors report presence of phonological effects in the context of no semantic effects in a lexical naming task. In their study, participants performed a picture-naming task. The picture was accompanied by a second picture, which could be either phonologically or semantically related. In the phonological condition subjects were significantly faster than in the semantic condition. At the same time, there was no difference between the semantically related and unrelated conditions. Authors showed that it is
possible to activate the phonological representations without activating semantic representations as well. Distractors could lead to phonological activation without leading to semantic activation. This finding underlines the dissociation between phonological and semantic representations even at the lexical level. One interesting aspect about the contribution of sublexical representations in the perception of specified contrasts is that they operate on lexical items as well. Lexical phonological representations contain abstract forms of words. But when contrasting
representations of those words. For instance, when comparing the words “cat” and “bat” we are not dealing with underspecified phonemes, or with allophones. We are contrasting fully-specified phonemes at the surface level (Chomsky & Halle, 1968), or, according to the terminology used in this dissertation, fully-specified phonemes at the input sublexical level (Ramus et al., 2010). 2. 1. 4 Bound morphemes and sublexical level The model used in this research is the model of Ramus (2010). The model has been used widely in the study of dyslexia and is designed to represent human’s representations (lexical and sublexical, as well as more peripheral systems such as acoustic and orthographic representations). The distinction between orthographic and phonological representations makes the model able to endorse some ideas proposed by Grainger and Ziegler (2011) for the processing of polymorphemic words. As Seidenberg (2005) points out, most models concentrate on single word processing. Even at a single word level, most models fail to account for word-structure processes, particularly morphological derivation. Many words present with
morphologically complex structures, and with phonologically attached functional elements, known as bound morphemes. There are two types of bound morphemes: derivational and inflectional. Derivational morphemes can change the word category of the word that they bind to, and are not grammatical, in the sense that they do not put the word in relation with other words in the sentence. Examples of this type are -ful, i.e. “painful”, or re-, i.e. “rewrite”. They can appear in word initial and in word final position in English. Inflectional morphemes, however, do not change the word category of the word they bind to, and express grammatical functions, expressing the relation between different parts of the sentence (for example, expressing the agreement between subject and verb). Examples of inflectional morphemes are -s, as in Mary speaks Bantu, or -ing, as in Mary is speaking Bantu. In English they always appear in word final positions. Speech perception models normally overlook the existence of morphologically complex words (Seidenberg, 2005). Exception to this is the model of Grainger and Ziegler (2011). The authors argue that certain strings of graphemes (we will later propose that the idea can be extended to certain strings of phonemes) are stored as morphosyntactic units, and processed separately. For instance, strings such as “-ed”, indicating a high likelihood of the presence of a bound morpheme, may be detected sublexically with a parser tuned to the detection of bound morphemes. This idea can be accounted by the model of Ramus (2010), which does not exclude strings of phonemes or graphemes to be processed as units at sublexical levels.
2. 2 Word beginnings and word endings 2. 2. 1 Word beginnings From a lexical point of view (Ramus et al., 2010), word beginnings have been shown to be salient by a number of authors. First, several studies suggest that word beginnings play an important role in lexical access. For instance, Browman (1978) showed that word beginnings are the part of the word most likely recalled in the tip of the tongue phenomenon. Marslen-Wilson and Zwisterlood (1989) have shown that word initial onsets play a bigger role in lexical access than word initial rhymes. Their hypothesis was tested using cross-modal priming in a lexical decision task. The prime could either be the rhyme of the word to be evaluated or the complete form of the item, including the onset. Results show that priming was effective only when the onset as well was used as a prime, and not effective when the rhyme was used in isolation as a prime. Theoretical work on word position effects goes in the same direction. Beckman (1998), bringing evidence from several languages (among them Shona, English, German, Hungarian, Polish and Lardil), showed that word onsets and non initial positions differ in the nature of the type and amount of segments they allow. While initial positions allow a huge variety of marked segments, non-initial positions undergo a set of phonological degenerative processes and tend to limit the number of phonemes allowed. For example, in Shona, a Bantu language of Zimbabwe, in initial position all the vowels of the language inventory are allowed, while in other positions the set of vowels allowed is restricted. Similarly, in Hungarian mid-vowels appear freely in word beginnings, but they are restricted in word medial and final positions.
Smith (2002) described the interaction between phonological and psycholinguistic phenomena on word position effects. According to the author, word beginnings are psycholinguistically salient, due to their importance in lexical access. As such, in those positions phonological processes such as deletion and assimilation (strongly active in final positions, Harris, 2011) are deactivated. Studies on the sublexicon (in the sense of Ramus et al., 2010) are however limited. Pitt and Samuel (1995) compared lexical and sublexical feedback in phoneme detection, and showed that similar detection processes are active at both levels.
Marshall and Van der Lely investigated word position effects in the
sublexicon in children with dyslexia and specific language impairment (SLI). The authors showed that children with dyslexia, children with SLI and children with comorbidity of the two disorders make more errors in the repetition of word medial clusters than in the repetition of word initial clusters. This result is in line with other studies suggesting that word initial positions are prominent, such as Beckman (1998), Smith (2002), Marslen-Wilson and Zwisterlood (1989). The idea that word beginnings are salient is well represented in models of speech recognition, particularly in the cohort model (McClelland & Rumelhart, 1981) and in developments of this model, such as the trace model (McClelland & Elman 1986) and more recent models such as the one proposed by Grainger and Ziegler (2011). Some authors, however, have questioned the saliency of word beginnings. Connine, Blasko and Titone (1993) failed to find a word beginning significant effect when nonwords were used to prime the access to real words. In their task, nonwords were generated modifying one or more phonemes of the real words to be accessed. The phoneme changed could either belong to word initial or to word medial
syllables. These nonwords were then used to prime lexical access in a lexical decision task. The ANOVA showed no significant difference when comparing performance in one or the other condition. The authors concluded that the mapping that takes place between spoken words and lexical representations does not attribute any special status to word initial positions. However, their interpretation may not be statistically sound: the evidence they report for the claim of the absence of word position is the absence of a significant interaction in their ANOVA. However, the absence of a significant effect does not entail a significant similarity between two conditions (Field, 2009). As such, it appears that the criticism of the word beginning saliency principle is not solid. 2. 2. 2 Word endings Word final positions are of great interest for phonologists and morphologists for different reasons. From a phonological point of view, word final positions are the locus of a series of processes that are not active, or at least less active, in other positions within the word. For instance, word final positions often undergo deletion processes, and feature assimilation with phonemes from following words (Harris, 2011, Roach, 2000). For instance, in the compound ‘handbrake,’ both processes are active: the final /d/ of hand is deleted, and the /n/ assimilates the bilabial feature of the following word and is pronounced /m/. From a morphological point of view, word final positions play a prominent role in a great number of languages (including English and Italian, the languages studied in these set of experiments) because they carry a large amount of morphological information (Pater, 2004). In English, all inflectional morphemes (-s, -ed, -ing) appear in word final position; the same can be
said about Italian (-o, -a, -iamo, -ate, -avo, -avi, -avate, -avamo, -avano, etc). The interaction between phonology and morphology was studied by a number of authors (Pater, 2004, Kuriso, 2001, Tyler, Randall & Marslen-Wilson , 2002). It appears evident, cross-linguistically, that the presence of morphological information undermines the application of phonological processes. See, for instance, this example from Pater (2006) from the Piro language. (1) /heta/+/ya/ → /hetya/ “see there” /heta/+/lu/→ /hetlu/ “see it” (2) /heta/+/nu/ → /hetanu/ “going to see” In (1) the verb “heta” (to see) is modified by an adverb, “there” and by a pronoun, “it”. No morphological modification is taking place. As a consequence, the word final position incurs in deletion. As one can see, after the application of the words “ya” and “lu” the final phoneme /a/ is deleted in both the examples in (1). In (2) the particle applied, “nu”, has a morphological value. It specifies that the action will take place in the future, and it operates as a bound inflectional morpheme (the translation in “going to see” is needed because English does not have any inflectional bound morpheme that expresses future). In this case the deletion process does not take place, and the word final phoneme of the stem, /a/, is pronounced in the inflected form of the word (Pater, 2006). Mahanta (2012) notices that the presence of morphemes can lead to exceptional alternations. Her analysis of Assamese shows that raising of vowels (a type of phonological change occurring in
certain phonological contexts) depends on the presence of morphemes, rather than on the pure phonological environment. Converging evidence of the special status of word final positions with morphological information comes also from clinical studies. For instance, there is evidence that English-speaking children with Specific Language Impairment (SLI) have difficulty with the phonological endings /s/, /z/ and /t/, /d/ only when these endings carry morphological information. For a child with SLI, it may be easier to produce a sentence such as “the bus is fast ” rather than a sentence such as “the coach goes fast ”, with the ending /s/ being critically morphological in the second example (Rice & Wexler, 1996, 2001). Further, if the presence of the morpheme leads to the generation of a phonological cluster, production is even more problematic (Marshall & van der Lely, 2007). The system by which inflected forms are generated is still not completely understood. According to one line of research, inflected forms are stored as units in the lexicon (Bertram, Schreuder & Baayen, 2000). According to another line of research, inflected forms are derived using rules (Pinker & Ullman, 2002a). The experiments conducted in the second part of this PhD will contribute new data to this debate.
2. 2. 3 The difference between word beginnings and word endings Word beginnings and word endings cannot be compared directly for two reasons. The first reason relies on the different nature of word beginnings and word endings. This difference can be understood by referring to the way we process the internal structure of words. First of all, one should notice that we do not process phonemes in isolation. Frisch, Large and Pisoni (2000) showed that wordlikeness judgments are influenced by suprasegmental features, such as the number of syllables, suggesting that speakers take into account syllabic structure when analysing and perceiving words. Treiman, Kessler, Knewasser, Tincoff & Bowman (2000) showed that, rather than processing single phonemes in isolation, we are sensitive to larger units (rime frequency, for example). Word endings, thus, have a different status compared to word beginnings. They belong to syllable codas, and are structurally close to the nucleus (main vowel of the syllable), while word beginnings belong to syllable onsets, as specifiers of the whole syllable. Comparing word beginnings and word endings, thus, would mean comparing elements that are structurally different in nature (for a graphic representation of the syllabic structure see Figure 2.4).
Figure 2.4 Prototypical syllabic structure.
The second reason for not directly comparing word beginnings and word endings relies on the different type of saliency we argue the two positions hold. While word beginnings are described as salient cross-linguistically (Beckman (1998), Smith (2002), Marslen-Wilson
& Zwisterlood (1989), McClelland &
Rumhelart (1981), McClelland & Elman (1986)) and also in clinical populations (Marshall & van der Lely, 2009), word endings are characterised by a more complex pattern: word endings are normally non salient (Harris, 2011) but optionally resist the activation of degenerative processes, when morphological information is present (Pater, 2006, Kuriso, 2001). In clinical populations, they may also show this optional saliency status, with word endings being difficult to process only when carrying morphological information, as, for example, in children with SLI (Rice & Wexler, 2001). Thus, directly comparing word initial position, which is inherently salient, and word endings, which are optionally salient, does not appear to be a sound procedure.
2. 3 The importance of using minimal pairs for the investigation of perception of phonological contrasts 2. 3. 1 Introduction All the experiments in this PhD use minimal pairs. While methodology is not the topic of this chapter, it is important to introduce the theoretical account of minimal pairs and the advantages that this type of structure can bring to the design of a study on perception. Minimal pairs are defined as pairs of words which differ in only one phonological element and have a different meaning (Roach, 2000). For instance, in English, the word “pear” and “bear” differ in only one phonological element, the initial sound, and as such they form a minimal pair. The contrast generated by the
minimal pair, i.e. the fact that the difference in one phonological element is able to generate two different meanings, is used as evidence in theoretical linguistics for the existence of phonemes, as well as for the existence of other phonological elements, such as tones in Chinese (Shen & Lin, 1991). A phoneme is the smallest linguistic unit which may generate a change in meaning. Thus, the sounds associated with the grapheme “p”, which phonologists write as /p/, and the sound associated with the grapheme “b”, which phonologists write as /b/, are phonemes. The minimal pair presented above, “pear” vs “bear”, thus, underlines the existence of the phonemes /p/ and /b/ in English; /p/ and /b/ are specified phonemes. They are differentiated based on the voicing feature, which is distinctive in English (/p/ being voiceless and /b/ voiced). The fact that this feature is distinctive in English means that the input sublexical representations are shaped according to it (Ramus et al., 2010). The contrast between /p/ and /b/ is then perceived sublexically. Before addressing the issue of minimal pairs, it is important to understand certain acoustic and cognitive aspects regarding the perception of speech sounds. When people speak their first language, the sounds that compose the language appear well defined, unconsciously, in our mind. We don’t need to understand the articulation instructions of our vocal system to understand that a certain sound is different from another one. We don’t need to know what our tongue and lips and vocal folds have to do in order to understand that pear and bear are two different words. However, from an acoustic point of view, the difference between pear and bear is very subtle. Specifically, there is a continuum of possible strings of sounds that goes from pear to bear and vice versa. This continuum is generated by the infinite number of sounds that can be uttered between the phonemes /p/ and /b/
(Liberman et al., 1957). From an articulatory phonology point of view, /p/ and /b/ are identical phonemes but for one trait. This trait, voicing, defines the physical vibration of the vocal folds. While uttering /b/ they vibrate, while uttering /p/ they don’t. Vibration, however can be manipulated on a continuum, and measured using computers (Liberman et al., 1957). We can then generate a set of intermediate sounds between /p/ and /b/. However, when a native speaker of English is faced with this intermediate sound, they either perceive it as /p/ or as a /b/, according to which is the closest prototypical phoneme. This property of our linguistic system is known as Categorical Perception, and is found very early in infancy, as early as 4 months of age (Eimas et al., 1971). Thus, the ability to discriminate words belonging to a minimal pair, such as “pear” and “bear”, relies on the ability to discriminate and categorise phonemes we develop during early infancy. This ability is non-acoustic, and properly cognitive, since what we operate is abstraction over the acoustic signal (Ramus et al., 2010, Chomsky & Halle, 1968). Even if articulation possibilities are infinite, we perceive specified contrasts in a categorical fashion (categorical perception will be discussed in detail in the section 2.3.2). When children acquire their first words, they are able to discriminate minimal pairs, showing ability to differentiate well- and oddly-formed words. For instance, Swingley and Aslin (2000), have shown that 18-month-old children are able to discriminate elements of a minimal pair formed of an existing common word and a modified version of the same word, such as baby and vaby. This ability is strongly linked to the first language of the speaker because input sublexical representations are shaped by the language of the environment (Ramus et al., 2010). A task that is very easy for an 18month-old native speaker may be very complex for a near-native L2 speaker. This is
what was demonstrated, for instance, by Pallier, Colomé and Sebastián-Gallés (2001), by studying native Spanish speakers’ perception of Catalan minimal pairs. At the same time, once defined during early infancy, this ability is very unlikely to be lost, even in extreme circumstances. For instance, Bouton, Serniclaes, Bertoncini and Cole
(2012) and Bouton, Cole and Serniclaes (2012), have demonstrated that
children with reduced hearing ability and cochlear implants remain quite proficient in the discrimination of minimal pairs, compared to their loss of other acoustic abilities. Saying that we are good at discriminating minimal pairs, however, does not imply that this ability is helpful in our language acquisition. Feldman, Griffiths and Morgan (2009) have shown that the presence of minimal pairs in human languages may actually be a confounding factor in the acquisition of phonetic categories. This finding may be explained by the theoretical analysis of Akamatsu (1995), who has shown that a minimal pair does not generate the sufficient number of contrasts needed to unequivocally identify a phoneme. What we need are minimal sets, that is to say a bigger number of words differing in one phoneme. In any case, there is evidence that minimal pairs are a cognitively valid concept, and data suggest that elements in a minimal pair are, at the cognitive level, coindexed (Astorkiza, 2007), i.e. stored in memory together with an abstract index that specifies their relation. Another theoretically relevant concept is the concept of morphosyntactic minimal pairs (Law & Strange, 2010). Terminologically, this refers to the condition in which words differ in one phoneme that is also an inflectional morpheme. In their study the authors have shown that words belonging to a morphosyntactic minimal pair are difficult for L2 speakers to distinguish if the inflectional morphemes leading to the difference are in a non-allowed position in L1. The miniparadigms acquired by
Italian children during early childhood (Guasti, 2004), such as mangio, mangi, mangia, pronounced /mʌ: ŋdᴣo/, /mʌ: ŋdᴣi/, /mʌ: ŋdᴣʌ/, may well be defined as morphosyntactic minimal pairs (minimal sets, because there are more than two elements) using the terminology of Law and Strange (2010). Law and Strange (2010) showed that perception of bound morphemes in morphosyntactic minimal pairs is influenced by the L1 in L2 speakers. Shtyrov and Pulvermüller (2002) elicited a specific brain component contrasting the morphosyntactic pairs “comes/come” and “come/comes”. Pulvermüller and Shtyrov (2003) showed that the ungrammatical fragment “we comes” generates a larger brain component than the grammatical fragment “we come” when elicited using an experimental design known as oddball paradigm. Apart from theoretical/experimental linguistics, minimal pairs are widely used in at least two domains of applied linguistics: second language (L2) teaching and speech and language therapy (SLT). In this chapter, after offering a more detailed account of the use of minimal pairs in research, we will describe the current practice in the use of minimal pairs in L2 teaching and in speech and language therapy (SLT). 2. 3. 2 Minimal pairs in research Minimal pairs tap into input representations. Classic categorical perception tasks are, in fact, minimal pair discrimination tasks (Hoonhorst et al., 2011, Liberman et al., 1957, Raphael, 1972). Thus, first of all, we can state that minimal pair discrimination tasks test the perception of speech contrasts. A selection of perceptual minimal pair tasks is presented below.
Liberman et al. (1957) carried out pioneering work on minimal pair discrimination. In order to explain their experiment and their findings, a few concepts on the acoustic nature of consonants need to be introduced. Each phoneme can be described with a specific limited set of features, corresponding to the articulatory features needed to utter that sound. A change in one feature leads to a different phoneme. For instance, the difference between /t/ and /p/ is in the position of the tongue/lips system (and nothing else): in one case, the tongue touches the alveolar ridge and the lips are slightly open; in the other, the tongue is lower and the lips touch one another. One of the features that can lead to different consonants is voicing. When vocal folds vibrate, a consonant is called voiced; when vocal folds do not vibrate the consonant is called unvoiced. This is the difference that occurs, for instance, between /p/ and /b/. While the articulation of /b/ requires the vibration of vocal folds, the articulation of /p/ does not. The distance, in time, between air release (plosion) and the vibration of vocal folds is called Voice Onset Time (VOT). In unvoiced consonants this value is large, since only the last moments of the consonant production correspond to fold vibration. In voiced consonants, this value is very small, if not negative, since fold vibration starts very early in comparison to plosion. VOT can be modified artificially using computers, and an infinite number of consonants in between prototypical /p/ and /b/ can be created. As already sketched in section 2.3.1, the crucial finding of Liberman et al.’s (1957) study is that human beings do not perceive these intermediate forms as intermediate forms, but categorically perceive either /p/ or /b/. Up to a certain boundary, all modified versions are perceived as /p/; after that boundary, all modified versions are perceived as /b/. Participants were asked to discriminate minimal pairs
of artificially-created syllables. If the consonants of the two syllables did not belong to different sides of the boundary, participants perceived the two syllables as identical. If the two consonants belonged to different sides of the boundary, they perceived different syllables. This categorical perception had nothing to do with the actual acoustic difference between the two sounds, i.e. it could not be predicted by the amount of variation in VOT between the two sounds, but only on the basis of boundary crossing. Following this pioneering work by Liberman et al. (1957), a number of studies on categorical perception assessed the precise nature of the boundary and the extension of categorical perception to other domains, such as facial expressions, but maintained the general features identified in its first definition (Hoonhorst et al., 2011 ) Several other lines of research take advantage of minimal pairs processing: Nazzi (2005) reports differences between consonants and vowels in minimal pair acquisition. In his experiment, 20-month-old children were trained to learn new words. First, children were presented with three physically different objects. The objects were given nonword names, and two of them received the same name. Word learning was evaluated with an object selection task. The words could differ in several ways. They could be completely different, or form a minimal pair. The minimal pairs used could differ in the first consonant, the medial consonant, the medial vowel or the final vowel. While children were able to acquire the consonant distinction, their performance on vowel contrasts was at chance level. These results raise interesting questions about minimal pair distinction, and suggest that consonants and vowels are treated differently, at least at a certain age, by the speakers. Dietrich, Swingley and Werker (2007) found that native language relevant
contrasts influence minimal pair discrimination at 18 months. Rost and McMurray (2009) report children’s failure to learn minimal pairs. Despite being excellent in perceptual task, discriminating, for instance, pairs of syllables (such as /pa/ vs /ba/), children around 12-14 months of age are surprisingly poor in discriminating pairs of words differing in the same phoneme (such as “pear” vs “bear”). Dehaene-Lambertz et al. (2000) investigated the perception of minimal pairs in L2 using electrophysiological measures such as brain Event Related Potentials (ERPs). The authors found that if the contrast is not relevant in L1 (even in nonwords tasks) participants do not perceive it and do not produce MMN. The perception of minimal pairs is also influenced by regional variations. Hayes-Harb (2007) assessed the contribution of minimal pairs in the acquisition of new contrasts in L2 compared to the acquisition based on statistical information only. The author shows that acquisition can be performed relying only on statistical information. However, the presence of minimal pairs corresponds to a significant improvement in the acquisition of new contrasts for adult L2 learners. Tyler et al. (2002) used minimal pairs to show that past tense formation is a morpho-phonological process: in this task subjects were asked to detect the difference between the past tense and the stem of regular (hugged/hug) and irregular (taught/teach) past tense verbs, as well as matched “pseudo” pairs (trade/tray and port/peach). Further, there was a control condition with nonwords (e.g. nugged/nug). Results show that subjects took consistently longer to discriminate inflected real regular verbs than irregular and pseudo verbs. Some of the subjects had mild to severe phonological deficits, but these neither correlate nor explain variance in the morphological minimal pairs task (Tyler at al., 2002, p.1154).
The data are interpreted as evidence for the presence of a dual mechanism model of tense inflection. Finally, it should be said that minimal pair discrimination can be seen as an instantiation of immediate priming (Diependaele, Grainger & Sandra, 2012). The first element present, being identical to the second but for one phonological feature, should strongly activate the other element in the pair, as shown by research on priming (James & Burke, 2000). 2. 3. 3 Minimal pairs in second language teaching Minimal pairs are also used in L2 language teaching, but their effect is debated (Levis & Cortes, 2008). In L2 acquisition one of the most complex processes learners have to face consists in perceiving and producing contrasts that are not present in their L1. For instance, it is well known that native speakers of Japanese struggle in the detection and production of the English contrast /r/ vs. /l/. Iverson, Hazan and Bannister (2005) show that using minimal pairs in teaching this specific contrast to Japanese adults learning English results in a significant improvement in their perception and production. In a similar study McCandliss, Fiez, Protopapas, Conway and McClelland, (2002) show that using minimal pairs in teaching unfamiliar contrasts can be successful, but only if precise and constant feedback is given. However, even if positive effects are reported, many authors and teachers identify several problems in the use of minimal pairs in L2 teaching. Foote, Holtby and Derwing. (2010) report that, when asked, two out of three Canadian English as a Second Language Teachers (ESLs) declare that minimal pairs should not be the main
tool to teach pronunciation of unfamiliar contrasts. Brown (1995) describes in detail why minimal pairs should not be used in L2 teaching. Contrasts can be generalised using features, rather than words. For instance, the contrasts /p/-/b/, /k/-/g/, /t/-/d/, /s//z/ and several others can be generalised as voicing contrasts, i.e. all pairs of phonemes are distinguished by whether the vocal folds vibrate or do not vibrate. Using minimal pairs, this generalization cannot be captured. 2. 3. 4 Minimal pairs in speech and language therapy Minimal pairs are widely used in language pathology treatment, and, specifically, in the treatment of people with a phonological disorder. According to Brancalioni, Bonini, Gubiani and Keske-Soares (2012), minimal pairs should be used from the very first stages of language remediation, i.e. during the assessment of the disorder. Many practitioners and researchers have tested the efficiency of minimal pairs (post assessment-treatment phase of phonological disorders). Earl (2011) reviewed the different ways minimal pairs are used in remediation. The first is minimal pairs treatment utilising five levels of training. In this type of treatment, children are asked to follow five different steps composed of: perception (discrimination) tasks, minimal pairs imitation tasks, independent naming tasks, explicit production of minimal pairs, and production of sentences. After the five steps, the children are assessed again. This approach is quite successful in terms of generalisation. Minimal pairs treatment is also used at the word and sentence level. After introducing minimal contrasts at the word level, the child is asked to find out and produce “silly sentences”, sentences that sound strange semantically with one
element of the minimal pair, but acceptable in the other case, such as “the pear is running in the snow” VS “the bear is running in the snow”. This treatment results in improved overall intelligibility (Earl, 2011) Minimal pairs treatment has also been used with a large number of contrasts. This form of treatment may well be applied to the other forms of treatment discussed before. It consists of a precise assessment of the phonological needs of the client, and in the identification of all the contrasts needed to expect a general improvement in them. These treatments proved to be successful in several different populations. Treatment success was demonstrated in people with Chronic Aphasia due to stroke (Pulvermüller et al., 2001), in children with Consistent Speech Disorder (Crosbie, et al. , 2005) and children with phonological impairments, such as children with dyslexia (Blache et al., 1981, Barlow &Gierut, 2002, Gierut, 2001). Even if successes are undeniable, some authors are critical about the way improvements are measured and how they actually generalise everyday use of language. For instance, Sabem and Ingham (1991) point out that even if treatments with minimal pairs lead to improvement in non-ecological, experimental conditions, the ability of patients to generalise this improvement to everyday speaking is contestable. Gierut (2001) points out that not all minimal pairs entail the same degree of complexity. Some contrasts are more complex in phonological terms, for instance, because the differentiating phonemes share many features, and, as such, there must be a correlation between the severity of the disorder and the nature of the minimal pairs used in treatments. Otherwise, improvements would just be apparent.
2.4 The concept of saliency The concept of saliency is difficult to define and it is not clear what type of linguistic behaviours one should expect when investigating participants’ processing of salient positions. The definition of the term offered by the Oxford Dictionary of English is the following: The quality of being particularly noticeable or important; prominence Beckman (1998) consistently refers to prominent positions when discussing the phonological variety of word-initial positions, or when discussing the “strong” nature of stressed segments. Marshall and van der Lely (2009) refer to word beginnings as “psycholinguistically prominent positions”, in line with previous work by Smith (2004). Using a nonword repetition task on children with dyslexia and/or SLI, Marshall and van der Lely (2009) report greater accuracy in the repetition of phonological clusters in salient position compared to the repetition of phonological clusters in non-salient positions. This finding suggests that clusters in salient positions are repeated more accurately by participants. Neuroscientific studies have also investigated the processing of salient positions. Warren, Wise and Warren (2005, p. 638), analysing and interpreting data from previous studies, offer a neuroscientific explanation of the perceptual processing of phonologically salient positions: Detection of acoustic differences between speech stimuli produces greater activation of the posteromedial Superior Temporal Plane (STP) when those differences are phonologically salient. The posteromedial STP is a sub-component of the auditory cortex which is responsible for the early processing of sound. In fact, the posteromedial STP is the
component that shows the earliest activation related to sound perception, with positive surface waves observed in non-human primates already 12 month after stimulus presentation (Liegeois-Chauvel, Musolino & Chauvel, 1991). Studies, such as LaBerge (1983), show that the use of an additional amount of resources (which in neuroscientific terms may correspond to greater activation in the brain) may be reflected in longer reaction times when performing tasks that require the use of these resources. For this reason, one may expect longer reaction times when completing tasks that tap into salient positions (compared to tasks that tap into non-salient positions). In summary, referring to the studies presented above, we use the term “salient” to refer to positions that are prominent. Prominent positions are positions of the word that have a particular importance in terms of processing, as suggested in Beckman (1998) and Smith (2004). These word positions are likely to be granted more resources by our cognitive system, as suggested by Warren, Wise and Warren (2005) and are also likely to be processed with greater accuracy, as suggested by the results of Marshall and van der Lely (2009). In the meantime, in line with studies such as LaBerge (1983), they may require a larger amount of time to be processed than non-salient positions.
2. 5 Hypotheses The two hypotheses tested in this PhD are presented in a concise way below.
Figure 2.5 Schematic representation of the Information Processing Model proposed by Ramus et al. (2010). Input sublexical representations, circled in red, are the focus of the research conducted in this PhD.
At the input sublexical level (Figure 2.5, circled in red): Hypothesis 1: Word beginnings are inherently salient Hypothesis 2: Word endings are optionally salient While there is evidence that word beginnings are salient and word endings are optionally salient at the lexical level, little is known about the existence of similar constraints at the sublexical level. The set of experiments in the following chapters deals with sublexical representations and investigates sublexical phenomena related to word position effects.
CHAPTER 3. WORD BEGINNINGS MATTER 3. Word beginnings matter
3. 1 Summary Note: a significant part of this chapter was published in the peer reviewed journal Dyslexia, in Cilibrasi, Stojanovik and Riddell (2015). From a lexical point of view there is evidence that word beginnings are psycholinguistically salient, both in perception and production (Zwisterlood, 1989, Harris, 2011). Data are explained in theoretical terms as the result of the effect of a highly ranked constraint applied to the lexical representations (Beckman, 1998, 2013, Smith, 2002). In light of recent models of representations (Ramus et al., 2010), we expect this principle to be active sublexically across age and languages. To test this hypothesis, three tasks were carried out (the first two on Italian speakers, the third on English speakers). Twenty-one adult native speakers of Italian were recruited and asked to complete a nonword minimal pairs discrimination task. Results show that Italian adults made a significantly larger number of errors in the second last syllable than in the first syllable, showing that word beginnings are salient even in a language in which stress does not normally fall on the first syllable [In Italian, words usually get stress on the penultimate syllable (Hayes, 2012)]. Similar tasks were then carried out with Italian children and English adults. Italian children were more accurate in word initial positions, while English adults were slower in discriminating word initial contrasts (accuracy was at ceiling), suggesting that they devoted a larger amount of resources to word beginnings. The pattern, thus, is reported at different ages and in two languages, and word beginnings are reported as salient in all tasks. Interestingly, all tasks were performed using sublexical items (i.e. nonwords), suggesting that the
phenomena observed depend on sublexical representations (i.e. the word beginnings saliency principle is part of sublexical representations).
3. 2 Introduction As discussed in Chapter 1, theoretical linguistics has often overlooked the problem of speech perception. Computational approaches to lexical access, on the other side, have often investigated speech perception, and as a consequence they may offer relevant insights in the discussion for this thesis. The most influential model of speech perception is the one from McClelland and Elman (1986), known as the TRACE model (Strauss, Harris & Magnuson, 2007). Trace is a connectionist model of speech perception, i.e. a computer-based software that simulates human-like speech perception using artificial representations. The interesting aspect introduced by speech recognition models is time-based processing. The activation of words is seen as an ongoing process, contrary to classic generative linguistics, in which analysis is performed offline. The trace model takes into account the psycholinguistic saliency of word beginnings in lexical access. When a word is presented to the model, all words that share the word beginning are activated. For instance, when presented with the word telephone several competitors are active,such as the words telegraph or television. However, while the processing spreads over time, the access to competitors is inhibited. In the end, thus, the competitor is rejected and the stimulus is recognised. The effect described by connectionist models, such as TRACE (McCLelland & Elman, 1986), finds confirmation in psycholinguistic studies of lexical access (Marslen-Wislon & Zwisterlood, 1989, Smith, 2002). For instance, Marslen-Wilson
and Zwisterlood (1989) have shown that word-initial onsets play a bigger role in lexical access than non-initial material. Their experiment consisted of a cross-modal priming task in which participants were asked to identify words (lexical decision). The prime could either be the rhyme of the word to be evaluated or the complete form of the item, including the onset. The results indicated that priming was effective only when the onset was used as a prime, and not effective when the rhyme was used without the onset. However, psycholinguistics research has not dealt with sublexical items, with few exceptions (Pitt & Samuel, 1995, Marshall & Van der Lely, 2009, see Chapter 1 for an in-depth discussion). The set of experiments in this PhD thesis builds on work by Marshall and van der Lely (2009). In their study, the authors showed that children with developmental dyslexia and/or Specific Language Impairment (SLI) have more difficulties with repeating nonwords containing consonant clusters found in word medial syllable onsets than if they are in word initial syllable onsets, and children with developmental dyslexia only (no co-morbidity) are less accurate in the repetition of consonant clusters in unstressed than in stressed syllables. Given the theoretical foundation of their study (particularly, given the work of Beckman, 1998, 2013 and Ramus et al., 2010), we expect similar word position effects to be found in Typically Developing (TD) children in perception and production. In fact, Marshall and van der Lely (2009) explain their result as an instantiation of a word beginning saliency principle, as proposed by Beckman (1998), and by a phonological effect of stress. Since Beckman’s principle is a general principle assumed to be active in human language, we expect similar word position effects to be found in TD children, but also in children and adults speaking a differently stressed language from English,
such as Italian. In fact, if the word beginnings constraint is a universal constraint (Beckman, 1998), these results should be replicated in other languages. Thus, the aim of the present study was to investigate whether these word position effects were found in Italian speaking adults and TD children, and in English speaking adults. Twenty-two Italian adults, 34 TD Italian children aged 8-10, and 30 English adults completed three specifically designed tasks: All tasks used nonwords containing consonant clusters consisting of a plosive plus liquid (eg. /pl/). The choice of these clusters was motivated by syllabification reasons: often it is not clear whether phonemes belonging to clusters are processed as a unit or separated in different syllables. For instance, most speakers would agree that the syllabification of the word cargo is car + go, with the cluster /rg/ shared by the first and second syllable. This is not the case for plosive + liquid clusters, which are always processed as units (Roach, 2000). This is due to the existence of a phonological principle known as the Maximal Onset Principle (Kahn, 2015). The principle states that intervocalic consonants are maximally assigned to the onsets of syllables in conformity with universal and language-specific conditions. In other words, syllabification tries to assign more material to onsets than codas, as long as this does not violate any other phonological principle. In our tasks, consonant clusters could be either positioned in a stressed or in an unstressed syllable, and could be either in initial position (first syllable) or in medial position (second syllable). In all tasks, participants were asked to discriminate between two nonwords differing in one phoneme belonging to a cluster by reporting whether two repetitions were the same or different. Results from this task showed that Italian children and adults performed more accurately when discriminating word
initial contrasts than when discriminating word medial contrasts, especially if the cluster was unstressed. English participants performed at ceiling for what concerns accuracy, but a post hoc analysis of reaction times shows that they took longer when having to discriminate word medial contrasts. These results suggest that Beckman’s principle is active not only in children with developmental dyslexia and/or SLI, but that it is a more general principle applying to speech perception: we detected it in typically developing children, in English adults and Italian adults. The principle proved to originate in perception, and to apply to unstressed syllables.
3. 3 Hypothesis We hypothesise that the word beginning saliency principle proposed by Beckman (1998, 2013) and detected in production by Marshall and Van der Lely (2009) in clinical populations is a general principle that applies to the perception of any spoken material in both typical and atypical populations across languages. This predicts that there will be better accuracy in the detection of consonant clusters in word beginnings compared to the detection and production of consonant clusters in word medial syllables in typical Italian-speaking children and adults, as well as in English adults. There are good reasons to start-testing on Italian. As already mentioned, consonant clusters in the beginning of the word were produced more accurately than all other clusters in Marshall and van der Lely (2009) by children with dyslexia and/or SLI. One explanation for such results may be linked to the nature of English. English stress is often defined as unpredictable (Roach, 2000). This definition is accurate, because stress can occur in all different positions in English words, such as initial, medial or final syllables. However, the statistical distribution of stress in
English reveals a relatively consistent pattern. It has been calculated that in English around 75% of words receive stress in the first syllable (Jusczyk, Cutler & Redanz, 1993). Considering the frequency of the stress in initial position, and considering the importance of stress as a saliency factor (Marshall & van der Lely, 2009), it may be argued that initial positions are salient from the stress point of view in English. This is not the case for all languages however. In Italian, for example, most of the words receive stress in the penultimate syllable (Guasti, 2004, Hayes, 2012). In French, most of the words receive stress in the final syllable. This contrast is considered to be a fundamental cue for lexical segmentation during early infancy. As soon as children realize the nature of the stress pattern in their language, they can determine where words start and finish in the phonetic wave (Guasti, 2004, Mehler, Dupoux, Nazzi & Dehaene-Lambertz, 1996, Saffran, Aslin & Newport, 1996, Echols, 1996, Ambridge & Lieven, 2011, Dehaene 2009, Guttorm, 2010, Benasich & Tallal, 2002, Benasich et al., 2005, Mannel & Friederici, 2008). Stress patterns, thus, are very important in speech perception, and may act against the word beginning saliency principle (Beckman, 1998) in a language such as Italian. Given the importance of the stress patterns and the nature of their stress patterms, Italian and English appear as ideal languages to investigate the relation between the word beginning saliency principle (Beckman, 1998) and stress. In Italian, the stress pattern saliency may operate in contradiction with the word beginning saliency principle, offering a peculiar context for the investigatigation of the relation between the two saliencies. In English, the stress pattern saliency may operate in concert with the word beginning saliency principle, offering a complementary domain of investigation to the phenonema addressed on Italian. On the other hand, the word beginning saliency principle may
operate independently from stress, offering consistent patterns in English and Italian across different stress conditions. Experimental data are crucial in order to make sense of this peculiar combination of conditions.
3. 4 Methods 3. 4. 1 Methods: Experiment I - Italian adults Ethics, recruitment and consent: The current study was approved by the University of Reading Research Ethics Committee and it was given favourable opinion. Adult native speakers of Italian were recruited by contacting the University of Reading Italian Society. The Society is a non-profit association of Reading University students who are interested in Italian language and culture. Members can be of any nationality. The society put the researcher in contact with a group of Italian students who were spending the academic year (2013/2014) in Reading as part of the Erasmus Programme. The students were informally approached during society meetings, where the researcher distributed information sheets about the present study with contact details. If interested in taking part, the students contacted the researcher and a date and a time for testing were agreed. Testing was conducted in a suitable laboratory at the School of Psychology and Clinical Language Sciences. Before testing, students were given time to reread the information sheet and if happy to proceed with the study, they were asked to sign a consent form. Participation was voluntary and students did not receive any reward for their participation in this study. After the testing, the students, if interested, were provided with a precise explanation of the task conducted and the study performed. All sensitive data about the participants were locked in a filing cabinet to which only the PhD supervisor and the
research student had access. Participants were allocated a numeric identifier which was used to anonymise the data. The information linking participants to this numeric identifier was stored in a separate, locked cabinet. Calculation of sample size: The sample size was calculated using an a-priori power analysis through the software G-power. Since data were going to be analysed using a Two-Way ANOVA, a medium effect size for ANOVA was chosen as the aimed effect size. The chosen value is 0.06 η2 (chosen according to http://imaging.mrc-cbu.cam.ac.uk/statswiki/FAQ/effectSize). The software G-Power was then used to calculate the sample size needed to find an effect of this size (0,06 η2). Since G-Power uses the f effect size (instead of η2) to compute this operation, a formula was used to obtain the f value starting from η2. The formula used is the following: f = √ η2 / (1 – η2). Using this formula, the resulting value of f is 0.26. The value of f was then added in G-Power, specifying also the α error probability (0,05) and the power aimed for (0,8). Conditions of the test were also defined: testing was going to take place in 1 group only and the conditions in the test were 4. The resulting sample size suggested by the software was 22 subjects. Participants: The study was conducted on 22 Italian adults, 13 female, 8 male (1 not specified) aged between 20 years and 2 months and 24 years and 8 months, with a mean age of 21years and 8 months, and a Standard Deviation of 1 year and 6 months, and reported clinical conditions. All participants were Erasmus students, grew up as monolinguals, and started learning English from the age of 10 or11, which is common in the Italian primary school system. All students had at least an intermediate command of English (B1 in the European framework). Before approaching the Erasmus students as potential participants, we considered testing
pure monolinguals living in Italy, to avoid confounds with the L2. However, in the end we opted for Italian Erasmus students because, considering the current education system in Italy, in order for us to find pure monolinguals we would have needed to test elderly Italians and this would have created other problems in the interpretation of the results. Stimuli: The task contained 40 pairs of nonwords. Half of the word pairs consisted of identical words and half of the pairs consisted of pairs of nonwords differing in one phoneme, generating a minimal pair. The nonwords used were trisyllabic and contained only the vowel /a/. Every nonword contained one phonological cluster of the type “plosive+liquid”, for instance /t/ followed by /r/ as in /tra:kata/. When nonwords in the pair differed in one phoneme, this phoneme was always part of the cluster, and the difference was always in one single trait, which was voicing. The reason for choosing this contrast could be defined as “historical”, in the sense that this contrast has been used in a large number of studies on perception, dating back to the first investigations of categorical perception: For a review about voicing studies see Hoonhorst et al.(2011) and for a detailed description of categorical perception, see the discussion in section 2.3.2 . Clusters were positioned in the first or second syllable, and, similarly, stress could be positioned on the first or second syllable. For instance, pairs of differing nonwords included /tra:kata/ VS /dra:kata/ or /praka:ta/ VS / braka:ta/. The full list of stimuli is available in Appendix 1. Recording of the stimuli: The stimuli were recorded in the sound booth room of the School of Psychology and Clinical Language Sciences by a trained linguist whose first language was Italian. The software used was Audacity, running on a
computer using Windows. The microphone used was an AKG D80, the mixer was a Behringer Mini Mon, the pre-amplifier was a B-tech phono-microphone preamplifier, and the speakers were Sony SS-CMD373. The length of the stimuli was controlled “analogically”, using the audio software Audacity. Once imported in Audacity, we ensured that the stimuli did not differ in length of more than 25ms (i.e. three percent of the total nonword length), using the Audacity timeline. The nonwords were recorded as “pairs”, in the sense that each nonword recorded was followed in the recording by the other nonword forming a minimal pair. This corresponds to reading the list of nonwords in Appendix 1 row by row. Testing session: the testing session lasted approximately 30 minutes. Participants were asked to read the information sheet and sign the consent form. After giving consent, participants were asked to sit comfortably in front of the computer and the researcher gave the instructions for the task orally. The researcher then left the room and let the participant read the instructions on the screen and put on the headphones. The instructions on the screen guided the participant through the testing session. Procedure: The task was programmed on the software E-prime. Participants were presented with pairs of nonwords. Each presentation of a pair consisted of a rapid succession of two nonwords. The nonwords in each pair could either be identical or form a minimal pair, differing in one phoneme. A white sticker was placed over the letter “w” on the keybord, and a black sticker was placed over the letter “b” on the keybord. Participants were asked to press white when they thought the two nonwords presented as part of a pair were identical and to press black when they thought that the two nonwords were different. The exact text of the instructions
was: In the experiment you will hear pairs of invented words presented in rapid succession, one next to the other. Sometimes the two words will be identical, sometimes they will be slightly different. Press white when you think the two words are identical, press black when you think that they are slightly different”. Considering that the position of the cluster was manipulated and considering that the position of stress was manipulated, there were four possible conditions in which the two nonwords differed, and four possible conditions in which the two nonwords were the same. The combination of possible stimuli is presented in Table 3.1. In the table, stress is represented in its main manifestation in Italian: the lengthening of vowels. Stressed syllables, thus, are syllables that in this table end with the symbol “:”, representing lengthening. Stimuli were presented randomly, using the random function of E-prime. The order was different for every subject. Each minimal pair was never repeated twice in one experimental session.
Table 3.1 Example of stimuli for experiment 1. The four conditions are presented together with their contrasting items.
Description of stimuli Cluster in the first syllable, stress in the first syllable Cluster in the first syllable, stress in the second syllable Cluster in the second syllable, stress in the second syllable Cluster in the second syllable, stress in the first syllable
Each trial (i.e. each presentation of a pair) was composed of four different slides: the first slide, named “fixation”, introduced the subject to the new trial. It was
one second long, it did not contain any sound and it only consisted of a visualization of the symbol # in the centre of the screen. The second slide, named “soundout1”, consisted of the presentation of the first sound in the pair, which in Table 3.1 is referred to as “prime”. During the presentation of this slide, the screen was completely black (the stimulus was presented aurally). The length of this slide was 1000ms, although the length of the stimuli was, on average, 730ms (SD 39ms). The length was also calculated post-hoc with the automatic measures of the Voice Activity Detector (VAD) software and the automatic text-grid of Praat3 (with the help of Professor Radek Skarnitzl of Charles University in Prague). The final slide was named “TextDysplay”, and it contained an arrow ( --> ), that indicated the end of the trial and the approach of the new trial. Also the final slide had a length of 1s (1000ms). Scoring: The measures collected by E-prime were the following: the answer the subject gave (whether white, black, a non-valid key or no answer) and the time they took to provide an answer. Subjects had a limit of 1000ms to give an answer (from the beginning of the target stimulus). If no answer was given within 1000ms, the software coded “no answer”. Outcome measures: Accuracy was calculated by dividing the number of correct answers by the number of given answers. Since the task was measuring the ability to discriminate phonological contrasts, we only used the trials in which the two elements contrasted in the pair were different. The pairs in which the two
The automatic and analogical measures of length substantially overlap. One should note that the analogical measure may actually be more accurate, since the software is unable to interpret final vowel lengthening and is relatively inaccurate in the detection of the beginnings of words starting with voiced plosives.
elements were identical were used as fillers. Given the type of coding used, accuracy could be any number between 0 (none of the given answers is correct) and 1 (all of the given answers are correct). For instance, if, in a given condition, the subject answered 5 times and none of the answers was correct, the accuracy for that participant in that condition was coded as 0, since this is the result of the ratio 0/5. If the subject answered 5 times and 4 of them were correct, accuracy was coded for that participant in that condition as 0.8, since this is the result of the ratio 4/5. If, for instance, the subject answered only 2 times, but both answers were correct, accuracy was coded as 1, since this is the result of the ratio 2/2. A similar choice was made in previous research (i.e. Mora, 2005). Reaction times were measured from the beginning of the target stimulus (slide Soundout1). No further coding was operated since reaction times were not used in the data analysis. 3. 4. 2 Method: Experiment II - Italian children Participants: Thirty-four children from a state primary school in Siena (Tuscany, Italy), aged 8;03 to 10;01, were recruited (Mean age 8;09, SD, 6 months, 19 M). None of the children had a diagnosis of developmental disorders, and the children were seen individually. Children’s non verbal abilities were assessed using the Coloured Progressive Matrices (CPM) (Raven, 1995). The mean standard score for the CPM was 98, sd. 15. No child scored lower than 2SDs below the mean. Individual scores are available in Appendix 1. Reading performance was also assessed using a standardised measure of reading performance for Italian called Batteria per la Valutazione della Dislessia e della Disortografia Evolutiva - DDE-2 (Sartori, Job & Tressoldi, 2007). The children
completed subtests 2 and 3. Subtest 2 is a real word reading task, consisting of 4 types of words: highly concrete and frequent words, highly concrete and infrequent words, highly abstract and frequent words, and highly abstract and infrequent words. Subtest 3 is a nonword reading task, consisting of three types of words: short shallow words, long shallow words, and opaque words generated with regular orthographic rules (for more details, see appendix 1). Considering that Italian has a shallow orthography, TD children between the ages of 8 and 10 are quite accurate in reading; hence the time needed to perform the reading task is usually preferred as a measure of variability. The results showed that children’s mean reading time was 183 seconds (sd. 54). Reading accuracy was at ceiling and as a consequence we are confident in excluding the presence of phonological/reading deficits. Production Task: The production task required the child to repeat 40 trisyllabic nonwords containing clusters. Accuracy was measured. Nonwords were presented in a child-friendly context. Children watched a video of a dancing parrot that seemed to pronounce the 40 nonwords. They were asked to repeat what the parrot was saying as accurately as possible. They were only allowed one repetition. The video could be stopped at any point by the child pressing the space bar and was re-started by pressing the same key. Nonwords were generated so that each contained a phonological cluster. The cluster was always formed of a plosive consonant, followed by a liquid consonant, followed by the vowel /a/. The cluster could be either in the first or in the second syllable, and stress was either in the first or in the second syllable. This gives 4 conditions (see table Table 3.2 for details). Scoring: Initially we measured whether the child repeated the nonword or decided not to repeat a certain word. Using this distinction, we obtained a measure of
given and missed answers. We then transcribed all of the answers given by the children. In this context, an inter-reliability check was run by another trained linguist on 20% of the subjects. The check revealed a reliability of 96% with the main transcription. The accuracy in the repetition was then measured. The measure was performed using 16 different variables and concentrated on the phonological clusters. Eight variables measured substitutions of phonemes in the clusters. Specifically, 4 variables measured the substitutions of the plosive consonant in the cluster (one for each possible combination of cluster and stress position) and 4 variables measured the substitutions of the liquid consonant in the cluster (again, one for each possible combination of cluster and stress position). The remaining 8 variables measured the amount of deletion. Four of the variables measured the number of deletions of the plosive consonant in the cluster (one for each combination of cluster and stress position), the other four measured the number of deletions of the liquid consonant in the cluster (again, one for each combination of cluster and stress position).
Table 3.2 Example of stimuli for the production task in experiment 2. The four conditions are presented.
Description of Stimulus 1. cluster first syllable, stress first syllable 2. cluster first syllable, stress second syllable 3. cluster second syllable , stress first syllable 4. cluster second syllable , stress second syllable
Example i.e. /pla:kata/ i.e. /plaka:ta/ i.e. /ka:plata/ i.e. /kapla:ta/
Clusters were formed as a combination of plosives and liquids so that, in word medial position, the two consonants were always processed as belonging to the
same syllable. Ten words for each condition were created. For a complete list, see appendix 1. Perception Task: As in experiment 1, the perception task contained 40 pairs of words. Half of the word pairs were identical words and half were pairs of words differing in one phoneme generating a minimal pair. Children were asked to press white when they thought the two words were identical and black when they thought the two words were different. The words used were trisyllabic and contained only the vowel /a/. When words in the pair differed in one phoneme, this phoneme was always part of the cluster, and the difference was always of one single trait: voicing. This contrast has been used in several previous studies (for a review, see Hoonhorst et al., 2011). For instance, pairs of differing words included /tra:kata/ VS /dra:kata/ or /praka:ta/ VS /braka:ta/. Clusters were positioned in the first or in the second syllable, which was either stressed or unstressed. Thus there were four possible conditions in which the two words differed, and four possible conditions in which the two words were the same (Table 3.3). For this task children had no time limit, while for adults in experiment 1 the time limit was set to 1 second. Scoring in the perception task mirrored the procedure used in experiment 1, but since children in experiment 2 had no limit in time, a “lack of answer” was never coded.
Table 3.3 Example of stimuli for the perceptual task.
Description of stimulus Target
Cluster in the first syllable, stress in the first syllable /tra:kata/ /dra:kata/ /tra:kata/ Cluster in the first syllable, /traka:ta/ /draka:ta/ /traka:ta/ stress in the second syllable Cluster in the second syllable, /katra:ta/ /kadra:ta/ /katra:ta/ stress in the second syllable Cluster in the second syllable, /ka:trata/ /ka:drata/ /ka:trata/ stress in the first syllable Note: The stimuli for this task are identical to the stimuli used in experiment 1, but the time window allowed for answering was longer for children compared to adults. 3. 4. 3 Method: Experiment III – English adults A similar task was performed on 30 English adults, aged between 18 and 22, mean age 19 years and 11 months, Standard Deviation 1 year and 4 months. Participants were undergraduate students, recruited with the SONA system of the University of Reading. Participants did not present with any clinical condition. Regarding the stimuli, word-likeness was respected for English: this allowed the same clusters but required different vowels (schwa in unstressed positions), i.e. /tra:kətə/ vs /dra:kətə/. The reason for this change is related to the nature of the syllabic structure in English (Roach, 2000). In English, only one vowel is usually fully specified in every word, and this vowel receives stress. In monosyllabic words, which are the majority in English (Hayes, 2000), vowels are fully specified. In polysyllabic words, surface representations (i.e. sublexical representations) normally contain the underspecified vowel schwa in unstressed syllables. From a phonological point of view, schwa is defined as mid, central and lax (Roach, 1998). Mid refers to the opening of the mouth (not close nor open), central refers to the position of the tongue (not front not back) and lax refers to the amount of energy employed to
produce the sound, which is limited. In English, thus, stress is not only defined as a function of vowel length or pitch but also through the nature of the vowel employed. Examples of nonwords for this task are presented below (Table 3.4).
Table 3.4 Example of stimuli for the English perceptual task. Description of stimulus
Cluster in the first syllable, stress in the first syllable: Cluster in the first syllable, stress in the second syllable Cluster in the second syllable, stress in the second syllable Cluster in the second syllable, stress in the first syllable
3. 5 Results The results of Experiment 1 are presented first, followed by the results of Experiment 2 and then Experiment 3. 3. 5. 1 Results: Experiment I Table 3.5 Descriptive Statistics for experiment 1, including min and max values for accuracy in the four conditions, as well as mean, Standard Deviation and Standard Error
Key: cl1 str1 = the cluster is in the first syllable and the stress is in the first syllable, cl1 str2 = the cluster is in the first syllable and the stress is in the second syllable, cl2 str2 = the
cluster is in the second syllable and the stress is in the second syllable, cl2 str1 = the cluster is in the second syllable and the stress is in the first syllable; Accuracy is presented as the
proportion of correct responses..
In order to ensure that participants were engaging in the task and not answering randomly, D-prime values were calculated (using hit rate and false alarm rate). Using one sample t-tests, D-prime values were compared to 0 and to 1. Dprime values showed to be significantly bigger than 0, t (20) = 6.51, p < .0001, and significantly bigger than 1, t (20) = 3.92, p = .0008 two tailed. These results suggest that participants were not answering the discrimination task randomly, and it also indicates that performance was better than chance and that it was accurate in more than 70% of cases (MacMillan & Creelman, 2005). Since the study compares means in 4 different conditions (generated by two variables) in one group of people, a 2x2 repeated measures ANOVA was conducted. Data were analysed using a 2x2 repeated measures ANOVA. Factor one was word position (the cluster could be either in the first or in the second last syllable), factor two was stress (the cluster could be either stressed or not). Descriptive statistics are presented in Table 3.5.The 2x2 ANOVA shows a significant effect of stress, F (20,1) = 8.42, p = .009, no significant effect of word position, F (20,1) = .17, p > .05, and a significant interaction, F (20,1) = 30.08, p < .001. Figure 3.1 shows visually the comparison of means and the factors used in the 2x2 ANOVA.
Figure 3.1 Mean comparison for Italian adults’ accuracy.
Post-hoc t-tests were carried out to investigate which contrasts led to the significant stress effect and to the interaction. Six contrasts were analysed. For this reason, the Bonferooni adjusted alpha used is p = .05/6 = 0.008. The results obtained are summarised in Table 3.6.
Table 3.6 Post-hoc analyses.
Contrast t value p value Cl1Str1 – Cl1Str2 t (20) = .275 p = .78 Cl2Str2 – Cl2Str1 t (20) = -5.58 p < .001 Cl1Str1 – Cl2Str2 t (20), 2.87 p = .009 Cl1Str2 – Cl2Str1 t (20), -2.88 p = .009 Cl1Str1 – Cl2Str1 t (20) = -2.48 p = .02 Cl1Str2 – Cl2Str2 t (20) = 1.72 p = .09 Key: cl1 str1 = the cluster is in the first syllable and the stress is in the first syllable, cl1 str2 = the cluster is in the first syllable and the stress is in the second syllable, cl2 str2 = the cluster is in the second syllable and the stress is in the second syllable, cl2 str1 = the cluster is in the second syllable and the stress is in the first syllable.
According to the post-hoc analysis, the contrast that is driving the ANOVA is the contrast between stressed and unstressed syllables when the cluster is in medial position. This suggests that the significance of the stress effect is uniquely driven by the difference between stressed and unstressed syllables in medial position, and it suggests that the interaction is driven by the fact that the contrast is significant in medial position but absent in nonword initial position. 3. 5. 2 Results: experiment II Table 3.7 (below) shows descriptive statistics for experiment 2. The table reports values for the age in months, performance in CPM measuring non-verbal abilities, performance on DDE2 (Test per la valutazione della Dislessia e della Disortografia Evolutiva), which is a test of reading performance in Italian (high values represent worse performance) and TROG (Test for the Reception of Grammar), which assessed understanding of different grammatical structures.
Table 3.7 Descriptive Statistics for standardised tests and age (obtained from raw scores).
CPM (correct answers)
TROG (correct answers)
In the tables below descriptive statistics is presented for the experiment on Italian children. Table 3.8 presents values for perception, Table 3.9 presents values for production.
Table 3.8 Min, max and mean accuracy values are reported for perception, as well as SD and SE Accuracy is presented as proportion of correct responses.
Key: cl1 str1 = the cluster is in the first syllable and the stress is in the first syllable, cl1 str2 = the cluster is in the first syllable and the stress is in the second syllable, cl2 str2 = the cluster is in the second syllable and the stress is in the second syllable, cl2 str1 = the cluster is in the second syllable and the stress is in the first syllable.
Table 3.9 Min, max and mean accuracy values are reported for production, as well as Standard Deviation and Standard Error.
Key: cl1 str1 = the cluster is in the first syllable and the stress is in the first syllable, cl1 str2 = the cluster is in the first syllable and the stress is in the second syllable, cl2 str2 = the cluster is in the second syllable and the stress is in the second syllable, cl2 str1 = the cluster is in the second syllable and the stress is in the first syllable; Accuracy is presented as
proportion of correct responses.
Correlations: Initially, an analysis of correlation between age and accuracy in all the tasks was performed, in order to understand if age accounts for significant variance in accuracy. None of these correlations were significant. Age and Perception Accuracy, r = .24, p > .05, Age and Production Accuracy, r = -.25, p >
.05. Thus, age was not related to task accuracy and so was not considered in further analyses. Accuracy in the perception task was found to correlate significantly with reading performance, r = .38, p < .05. Accuracy in the production task (calculated dividing the number of errors by the number of given answers) did not correlate with reading performance, r = -.19, p > .05, but a partial correlation between number of missed answers and reading time (with accuracy as control) was significantly correlated with reading time using a one-tailed hypothesis (justified, for instance, by Torgesen & Burgess, 1998), r = .28, p = .05. Figure 3.2 and Figure 3.3 below show a scatterplot representing the correlation between reading and nonword discrimination, and reading and nonword repetition respectively. A bigger value in the reading task indicates lower performance, since it indicates that the child needed more time to complete the task.
Figure 3.2 Scatterplot of the correlation between accuracy in the perception task and performance in the reading task in seconds.
Figure 3.3 Scatterplot of the correlation between accuracy in the production task and performance in the reading task in seconds.
Production task (nonword repetition task): Production was analysed measuring accuracy in children’s performance and comparison performance in the four conditions. Figure 3.4 summarises the results in the analysis. Children made more errors in unstressed compared to stressed syllables. Word position effects were absent. The interaction between position and stress was also marginally significant.
Figure 3.4 Production: comparison of the means in the four conditions.
Following the correlation analysis, an error analysis was performed (for the production task). Deletions were quite rare in this task, occurring less than once in every hundred words (26 errors in 2720 nonwords presented), and were therefore not analysed separately. Instead, deletions and substitution errors were combined in one analysis. The relation between errors, stress and word position was analysed using two-way ANOVA: the first factor was the position of the cluster (word initial and word medial), the second factor was whether clusters were stressed or not (cluster stressed, and cluster unstressed). The analysis of errors shows a significant effect of stress, F (33, 1) = 23.096, p < .001, with children making more errors in unstressed compared to stressed syllables. There was no effect of word position, F (33, 1) = 1.84, p > .05, but we detected a marginally significant interaction, F (33, 1) = 3.82, p = .059. Post-hoc analysis shows that the contrast between stressed and unstressed syllables in word medial position is highly significant, t (33) = -4.08, p < .001, and
that the same contrast in word initial position is only marginally significant, t (33) = 2.45, p = .02 (Bonferroni adjusted alpha = .025). Perception: Perception was analysed with a 2x2 ANOVA that had position of the cluster and presence of stress in the syllable as factors. The result is presented graphically (see Figure 3.5) and explained below.
Figure 3.5. Perception: comparison of the means across stressed and unstressed, and initial and medial clusters.
Perception: comparison of the means across stressed and unstressed, and initial and medial clusters was conducted. There was a significant effect of stress, with children making more errors in the unstressed than in the stressed condition. There was also a main effect of position with children making more errors in the medial than in the initial position. However, there was an interaction between stress and position: children showed no difference in rate of errors between stressed and unstressed syllables in the initial position so differences between stressed and
unstressed syllables were limited to the medial position. Being a same-different task, we checked for the presence of biases using d-prime analysis, as it is suggested in the study of children and non-human primates (Blöte, Resing, Mazer & Van Noort, 1999, Katz, Wright & Bachevalier, 2002). We calculated hit rate, false alarm rate and the d-prime value for each participant. We then compared the d-prime values to 0 and 1 using one sample t-tests. The t-tests showed that the d-prime values are significantly different from 0, indicating that performance is not random (MacMillan & Creelman, 2005), t (33) = 8.27, p < .0001, and they are also significantly bigger than 1, indicating an overall accuracy for both different and same trials of more than 70% (ibid), t (33) = 3.42, p = .002, two tailed. In order to investigate word position and stress effects, we then conducted a two-way ANOVA having position of the cluster and stress as factors. The two-way ANOVA shows a significant word position effect, F (33, 1) = 12.76, p = .001. Children made more errors in the detection of contrasts when the clusters were in the medial than in the initial position (initial vs medial means, .421, .503, SE, .026, .036). There was also a significant effect of stress, F (33, 1) = 14.75, p = .001, with children making more errors when the clusters were unstressed than stressed (stressed vs unstressed, .400, .524, SE, .032, .032). Finally, there was a significant interaction, F (33, 1) = 8.18, p = .007. Post-hoc comparisons showed that children made significantly more errors in the medial position when the syllable was unstressed rather than stressed (t = 4.38, p < 0.0001) and made more errors in unstressed syllables when the cluster was in medial position compared to when the cluster was in initial position (t = 5.67, p < 0.0001). Other comparisons did not reach significance (see figure 6).
3. 5. 3 Results: experiment III Results from the experiment on English adults were analysed in a similar fashion. Descriptive statistics is reported in Table 3.10:
Table 3.10 Descriptive Statistics for experiment 3, reporting max and min values for accuracy, mean, Standard Deviation and Standard Error.
Key: cl1 str1 = the cluster is in the first syllable and the stress is in the first syllable, cl1 str2 = the cluster is in the first syllable and the stress is in the second syllable, cl2 str2 = the cluster is in the second syllable and the stress is in the second syllable, cl2 str1 = the cluster is in the second syllable and the stress is in the first syllable; Accuracy is presented as
proportion of correct responses.
Initially, we ran a 2x2 repeated measures ANOVA having word position as first factor (the cluster could be either in first or second position syllables) and stress as a second factor (the cluster could be either stressed or not). Table 3.11 presents the result of the 2x2 ANOVA measuring accuracy and having position and stress as factors.
Table 3.11 2x2 ANOVA on English adults accuracy.
Hypothesis Degress of Freedom
Error Degrees of Freedom
Position * stress
The analysis shows a word position effect (F (29,1) = 6.41, p = .01), which goes against our prediction. However, the result is based on a very low number of errors, since performance is at ceiling (English adults were accurate 9 times out of 10). We then performed an analysis of RTs. When the clusters were in word medial positions the distance in time between the uttering of clusters and word beginnings was measured in a set of randomly chosen words, in order to obtain an indicative value of the distance. Words with stressed clusters differed from words with unstressed clusters. Three random words were chosen for each condition. Length values are summarised in Table 3.12:
Table 3.12 Examples of measure of distance between word beginning and uttering of a medial cluster. Word A Word B Word C Median Stressed
This values show that there is certain variability among words in the moment of utterance of the cluster within the same condition of stress. Aware of the limits of this type of analysis, we used the median value for each condition to compute new
variables that would (partially) take into account the distance between word beginning and the uttering of word medial clusters. The new variables are obtained by subtracting the values identified above from the reaction time values for each condition4. Descriptive statistics are presented in Table 3.13. One subject was excluded because their response times were unrealistically approximating 0.
Table 3.13 Experiment 3 reaction times descriptive statistics.
Minimum Maximum Mean
Stress and word position effects were then analysed using a 2 x 2 ANOVA. The analysis shows a significant effect of stress (F (28,1) = 4.15, p = .05), a significant effect of position, (F(28,1) = 146.84, p = < .001 and a non significant interaction, F (28,1) = .43, p > .05). Figure 3.6 graphically reports the effects. Crucially for our discussion, contrasts in word initial positions required a larger amount of time to be recognised, suggesting that a larger amount of time (and possibly, of resources) was needed to process contrasts in word initial position. Caution is, however, advised in the interpretation of this analysis, since the analysis of RTs is a Post-hoc analysis, and the way variables were computed is relatively arbitrary (i.e. the procedure used to obtain computed variables that would take into account the distance in time between the uttering of clusters in medial positions and 4
The formulas used are the following: AdjCl2Str2=Cl2Str2–164; AdjCl2Str1= Cl2Str1 –248
word beginnings is arbitrary). We also reported a significant stress effect, with stressed contrasts taking shorter to be detected than unstressed contrasts.
Figure 3.6 Analysis of reaction times.
Post-hoc comparisons were run to assess the contribution of each condition to the main effects. The results of the Post-hoc comparisons are summarised in Table 3.14. Table 3.14 Post-hoc comparisons experiment 3.
List of comparisons Pair 1 cl1str1 - cl1str2 Pair 2 cl1str1 - cl2str2ADJ Pair 3 cl1str1 - cl2str1ADJ Pair 4 cl1str2 - cl2str2ADJ Pair 5 cl1str2 - cl2str1ADJ Pair 6 cl2str2ADJ - cl2str1ADJ
t values 1.055 8.705 11.997 7.932 11.417 1.847
DF 28 28 28 28 28 28
p values .301 .000 .000 .000 .000 .075
The post hoc analysis shows that any contrast in which the position of the cluster differed revealed a significant difference, while the two contrasts in which the position of the cluster did not differ (i.e. the comparison was between clusters in the
same position, and one of them was stressed and the other unstressed) did not show a significant difference.
3. 6 Discussion The aim of the three experiments was to investigate whether word initial positions are salient and how stress interacts with word position effects. The results showed that word beginnings are perceptually salient across age and in two different languages. This word position effect was found to interact with stress: it was consistently found only when clusters were in unstressed syllables. Interestingly, however, the word position effect was not predicted by the stress pattern of the language: even if words in Italian and English typically receive stress in different positions, the saliency principle applied to initial syllables in both cases. This finding is also relevant for the discussion of the phenomenon known as speech segmentation. One of the first important tasks children have to deal with during the early stages of language acquisition is segmenting words from the stream of speech, since speech is acoustically continuous (moments of silence correspond to plosive articulation and not to word boundaries) (Guasti, 2004). The process of speech segmentation consists of different connected stages. Vowel frequency plays an important role in the detection of language differences in very early stages of acquisition (from the first days). In the first instance the child may recognise the rhythm pattern of their language using the frequency distribution of vowels (Mehler et al., 1996). For instance, a child may recognise that French is different from Japanese because vowels are more frequent (or less distant) in Japanese than in French. For the same reason, at this stage a child may not recognise the difference
between Spanish and Italian (Nazzi, Bertoncini & Mehler, 1998). The frequency of vowels could tell the infant the average duration in syllables of the words in their language, using a basic assumption: languages with complex syllables tend to have shorter words, while languages with simple syllables tend to have longer words, in terms of syllables. This implicit knowledge helps in performing boundary detection (Nazzi et al, 1998): Languages such as English, with complex syllable structure, i.e. languages that allow complex phonological clusters between vowels, tend to have a higher number of monosyllabic and disyllabic words. Languages such as Japanese, which admit only CV syllables, tend to have a higher number of polysyllabic words. The reason can be understood by referring to combinatorial possibilities. In a system that uses monosyllabic words and simple syllables, such as CV syllables, it is not possible to generate a big vocabulary, because combination possibilities are very limited. This is why Japanese words tend to be polysyllabic. Thus, children can make an estimate of the syllabic length of words in their language using syllabic structure. When children have a reliable estimate of the number of syllables in the words in their language, the frequency of adjacency of syllables could tell them where word boundaries fall (Saffran et al, 1996). Japanese children may look at the probability of strings of four syllables, while English children may focus on two syllable strings. Syllables that often occur in adjacency are recognised as words (Saffran et al, 1996). When children have a sufficient number of words in their vocabulary, they can extract the stress pattern of these words and try to use this pattern as a cue for further speech segmentation (Echols, 1996). For instance, English children may assume that words start with a stressed syllable, while French children may assume that they finish with a stressed syllable, since it is very likely that the first words they learn
will follow this pattern. Thus, stress is a fundamental cue in early speech segmentation, and this is an accepted position both in the constructivist and the generativist approaches to language acquisition (Ambridge &Lieven, 2011). The stress pattern of the language reveals which positions are salient in continuous speech. However, our results suggest that performance in tasks using words in isolation (such as the task used in this set of experiments) may be driven by a more general principle, which is independent from the stress pattern of the language but, at the same time, de-activated in stressed syllables. The reason behind the deactivation is not clear. A possible interpretation may be offered by referring to Optimality Theory. The word beginning saliency principle (Beckman, 1998) is described as a constraint within optimality theory. As discussed in chapter 1, constraints are different from rules in that they can be violated (Prince & Smolensky, 1993). Thus, when we state that there is a constraint establishing that word beginnings are salient, we are not stating that every speaker in every context will concentrate their resources on word beginnings, but rather that this will happen most of the time. The constraint is, according to Beckman, highly ranked (Beckman, 1998). This means that most constraints will not be able to lead to its violation, and only constraints positioned in a higher position will lead to its violation. Our data suggest that a phenomenon of this type is taking place with stressed syllables. If we assume the existence of a STRESS constraint, the interaction between word beginning effects and stress becomes predictable. For instance, if we assume the existence of STRESS constraint that establishes that stressed positions are equally salient, and we assume this constraint to be ranked in a higher position than Beckman’s principle, then we would predict the word beginning saliency effect to be
active only in unstressed syllables. This explanation may seem strongly speculative, but it could be accounted for in neat psycholinguistic terms. From a phonetic and psycholinguistic point of view, it is well acknowledged that stressed positions are salient (Roach, 2000). On the other hand, there is evidence from lexical access studies showing that word beginnings are salient (Zwisterlood, 1989). The present study investigated, by psycholinguistic means, the interaction between these two psycholinguistic phenomena, showing that stress operates somehow at an earlier stage compared to the word beginning principle. The interaction was reported across two languages, and this suggests that it depends on non-language specific systems. It was also reported across age, suggesting that word position effects emerge at a relatively early stage. Finally, and most importantly, the interaction was obtained using non-lexical, though language-specific stimuli. This suggests that the phenomena observed take place sublexically, and that stress and word position saliency may largely depend on sublexical representations. Another relevant finding of the present set of the experiments is the positive correlation which was observed between reading performance and performance in the minimal pairs discrimination task. This suggests that sublexical input performance is related to reading performance, and the fact that these relations occurs may offer relevant insights for the study of reading disorders. Reading disorders are diagnosed at an arbitrary cut-off point on the continuum of reading performance. The National Institute of Neurological Disorders and Strokes defines dyslexia with the following statement (Williams & Lind, 2013, p23):
Dyslexia is a brain-based type of learning disability that specifically impairs a person’s ability to read. These individuals typically read at levels significantly lower than expected despite having normal intelligence. The notion of “lower than expected” is, however, problematic. As said, performance can be mapped onto a continuum. The boundary is arbitrary, even if researchers and educational psychologists make a decision using normative data. Given this consideration, it is possible to study the range of reading abilities and possible reading deficits in populations with no formal diagnosis. With a sufficiently large sample, the performance of people in reading tasks exhibits a normal distribution. The lower limits of the distribution represent people with reading difficulties (and probably some of the people who fall in this part of the distribution would get a diagnosis if tested), while the central part of the distribution (most people) represents controls and the upper limits of the distribution represent exceptionally good readers. Ramus et al. (2003) showed that a phonological deficit is the only one which alone is able to explain the difficulties of children with dyslexia. In their paper, the authors assessed a large sample of individuals with dyslexia with tasks tapping on the cerebellar domain, on the phonological domain and on the magnocellular one, and showed that phonological problems are the only ones that can occur in isolation. In short, given the relation between dyslexia and phonology, and considering that performance is normally distributed in large samples, it can be stated that the performance of subjects with no clinical diagnosis may be informative and/or relevant in the understanding of dyslexia. Along these lines, the present study also investigated whether phonological discrimination can predict reading performance in TD children. Performance on the
reading task correlated with performance on the cluster discrimination task in TD children: the better the children performed in the reading task, the better they performed in the discrimination task. Performance in the discrimination of clusters which differ in only one phoneme highly correlates with reading performance (DDE2). Therefore, children’s reading abilities are related to their ability to discriminate minimal pairs generated by clusters. This result is in line with previous research on the topic (Hoonhorst et al., 2011) The ability to discriminate minimal pairs relies on the amount of information associated with single phonemes. In this task, words differed on one phonological trait: voicing. Discrimination in this task therefore requires exhaustive mapping in traits of every single phoneme. Each phoneme can be described exhaustively as a list of traits: by specifying a list of parameters, such as the position of the tongue, or the vibration of the vocal folds, it is possible to identify unambiguously any single phoneme. Our input sublexical phonological representations contain this information (Ramus et al., 2010). In order to perform a minimal pairs discrimination task in which one trait is leading to the contrast, as the one used in this set of experiments, participants need to have A) a precise list of traits associated with each phoneme, in order to identify unambiguously each of them; and B) a system that is able to contrast traits between phonemes, and that is able to identify minimal differences such as one trait contrasts. This system is set in the input sublexical phonological representations (Ramus et al., 2010). Children with dyslexia are unable to create an exhaustive phonological mapping of phonemes, and as a consequence they struggle in phonological tasks that require a high level of phonological awareness (Dehaene, 2009). The reason why this deficit can lead to poor reading performance is relatively
straightforward to present. Reading is a very complex process that involves several sub-processes. Among them, and of central importance during reading acquisition, is the ability to associate graphemes to phonemes. In order to associate graphic symbols to specific sounds, a very precise mapping of sounds is required, and children with reading difficulties may lack such a precise system. Our set of experiments suggests that the phonological phenomena observed (with large effects) in children with a clinical condition may actually be found in non-clinical populations, even if less pronounced. This may be related to the fact that the definition itself of clinical populations depends on a cut-off point in the continuum of what is considered clinical and what is considered non-clinical. Consequences of our findings for clinical assessments are discussed in detail in chapter 5, where clinical data are also analysed.
3. 7 Conclusion We hypothesised that the saliency principle proposed by Beckman (1998, 2013) and detected in production by Marshall and Van der Lely (2009) in clinical populations is a general principle that applies to the perception of any spoken material and across age. The main hypothesis is confirmed: word position effects are present in TD Italian speaking children, in Italian speaking adults and in English speaking adults. Italian adults were less accurate in the perception of word medial contrasts, when the clusters were unstressed. Italian children were more accurate in the discrimination of word initial contrasts than in the discrimination of word medial contrasts, if the clusters were unstressed, in the perception task. With regard to the production task, no word position effect was detected, but the data showed a
significant stress effect, with stressed clusters being repeated more accurately. Concerning English adults, we reported an unexpected result in accuracy measures, but this may be due to the fact that performance was at ceiling. Reaction Times, on the other hand, presented an interesting result, with subjects being slower in discriminating word initial contrasts than in discriminating word medial contrasts. Longer reaction times may be interpreted as a manifestation of a form of saliency, in the sense that subjects devoted a larger amount of time to word initial positions. A similar interpretation is offered in previous work, in which longer reaction times are interpreted as the use of a larger amount of resources (LaBerge, 1983). We also reported a significant stress effect, with stressed contrasts taking less time to be detected than unstressed contrasts. If we assume stressed conditions to be salient (based on previous research, such as Marshall & van der Lely, 2009), the only possible interpretation is that the two forms of saliency are substantially different, leading to completely different patterns in the reaction time measures. The results we obtained in the three experiments extend Marshall and van der Lely’s work (2009). They also confirm Beckman’s principle (1998, 2013) in perception, and are in line with the predictions of Ramus et al. (2010)’s model of phonological representations. Further, accuracy in the perception task was found to correlate significantly with reading performance, extending to Italian a cross-linguistically well- established correlation between voicing contrast perception and reading (Hoonhorst et al., 2011).
CHAPTER 4. WORD ENDINGS ARE OPTIONALLY SALIENT 4. Word word endingsendings are optionally salient
4. 1 Introduction In English, word endings are the locus of a series of important morphological processes. Among them, crucial relevance is given to the process of verb inflection. The inflection of verbs is a complex phenomenon in which phonological and morphological processing co-occur. For this reason, verb inflection is an excellent ground of research for the investigation of the interaction between word position effects and morphology. The following chapter investigates word position effects in inflected verbs. The processing of inflected verbs is a strongly debated issue in psycholinguistics. Some researchers suggest that inflected verbs are stored as units (i.e. root + bound morpheme) in the lexicon as evidenced by the presence of strong frequency effects in lexical decision tasks with inflected verbs. Others suggest that inflected verbs are decomposed into stems and affixes in perception (morpheme stripping) and generated by the application of a rule in production (for instance stem +ed, stem +s in English), as evidenced by phenomena such as over-regularizations. There have also been proposals which suggest that a morpheme stripping process takes place sublexically, at least in reading. The current study investigated processes underlying morpheme stripping using lexical and sublexical items in a set of three perceptual experiments. The aim was to find evidence either for the separation of stems and affixes, or for the unification of stems and affixes in the processing of lexical and sublexical items containing bound morphemes.
To test this, three experiments were designed: In Experiment 4, participants were asked to discriminate elements belonging to lexical and morphosyntactic minimal pairs and reaction times were measured. In Experiment 5, event related potentials were measured to determine the size of the mismatch negativity (MMN) using the same stimuli as in Experiment 4. In Experiment 6, participants were asked to discriminate between pairs of nonwords, with both morphological and lexical contrasts 5. In Experiment 4, we found that elements belonging to morphosyntactic minimal pairs took longer to be discriminated than lexical minimal pairs. In Experiment 5, the MMN was larger when morphemes were present, and in Experiment 6, non-words with potential morphological information took longer to be discriminated than those without. The results of the study suggest that inflected verbs are decomposed into stems and affixes perceptually, both at the lexical and sublexical levels. Our findings are in line with dual models of verb processing that assume rule-based processing of inflected forms, such as the one proposed by Pinker and Ullman (2002a), although a more flexible explanation is proposed, in order not to ignore the evidence for whole-word processing reported in other studies.
4. 2 Background and hypothesis As explained in detail in the section 2.2.2. word final positions are of great interest. On the one hand, they are the locus of disruptive phonological processes such as deletion and assimilation. On the other hand, these processes can be blocked when
For an alternative approach to the problem, see Appendix 3, where the stimuli developed for our experiments are used as data for a connectionist model developed by Harald Baayen at the University of Tuebingen.
morphological information is present. This is what happens, for instance, with inflected verbs in English, Italian, French, and Czech. The way in which inflected words are processed and stored is a matter of intense debate. Most psycholinguistic research has so far focused on English. Since this is a poorly inflected language, much of the research has focussed on a limited number of inflectional processes, with most work addressing the derivation of the past tense (Pinker & Ullman, 2002a). From this work, some authors have suggested that inflected forms are stored as units in the lexicon (Stemberger & MacWhinney, 1986, Bertram et al., 2000, Maslen, Theakston, Lieven, Tomasello, 2004). “Played” and “cared”, for instance, would be stored in our declarative memory together with their bound morphemes, ed. The evidence for this comes from strong frequency effects in lexical access tasks. The time needed to access inflected forms is strongly related to the frequency of those inflected forms in the corpora available for analysis (Stemberger & MacWhinney, 1986). Further evidence comes from the study of richly inflected languages. For instance, studies in Finnish reveal similar frequency effects (Kirjavainen, Nikolaev & Kidd, 2012). The idea of the storage of inflected forms is also supported by research on language acquisition (Tomasello, 2006; Diessel, 2012). According to Tomasello (2006), during the early stages of word acquisition, children store all the available inflected forms in the lexicon. Then, by analogy, they extend the principle observed (for instance, +ed for regular past tense in English) to other occurrences of verbs in the past. The nature of this analogic effect, however, is poorly formalised in principle, since research suggests that many uses are idiosyncratic to the single item, and as such, creating new past forms from analogy does not mean extracting a general rule and applying that rule to all occurrences
(Tomasello, 2000). Pinker and Ullman (2002b, page 474), describe the circularity of attempts to describe tense inflection in terms of analogy, in a comment on the connectionist model of past tense learning by McClelland and Patterson (2002): We agree that connectionist networks are not always analogy mechanisms. Our point (based on explications by McClelland and other connectionists) is that pattern associators (the most common connectionist model of the past tense) tend towards analogy when learning competing patterns under standard training regimes. This is what gives such models their predictive power with irregular forms. The claim that some connectionist models can, given a specific architecture, training schedule and input features, approximate any linguistic phenomenon might be true, but it is in danger of reducing connectionism to a universal statistical approximation technique rather than a source of empirical predictions. Language cannot be treated as just a collection of ‘regularities in the input’ that can be approximated by some mechanism; those regularities are themselves the products of human minds and need to be explained. As Pinker and Ullman (2002b) explain, the evidence that the past tense is generated through associations and analogy is not conclusive. First, they notice, the frequency effects we observe are a consequence of complexity, not a cause of it. The regularities and irregularities that we find in the corpora are a result of the processing that takes place in our mind. Second, the fact that past tense is formed through analogy may well be true, but analogy must be the object of research itself. The questions to be asked are: What exactly are the analogical or associative processes that lead to tense generation? Are these rule-like processes? Pinker and Ullman (2002a) suggest that these are rule-like processes. Regular past tense is derived
through the application of a rule, which in English is “+ed”. The bound inflectional morpheme is added to the stem, and the lexicon contains only non-inflected forms. This proposal has a solid theoretical foundation in the concept of economy (Blevins, 2004). For highly inflected languages in particular, this type of derivation would ensure that a significantly smaller number of words are required to be stored in memory. In Italian, for instance, there are at least 75 different inflectional bound morphemes (Palermo, 2000). Storing only stems for verbs, then, would be substantially more economical in terms of memory load. In English, the advantage is less evident, but if we consider the three regular inflections –s, –ed, and –ing, storing stems and affixes separately would be three times more economical than storing all forms. Further evidence that inflected verb forms may not be stored as units comes from classic work by Berko (1958). This shows that children at a young age derive inflected forms using rules. When presented with a new item, for instance, a picture of an animal that the researcher names “wug”, children are able to derive the plural form “wugs” on their own, without any previous exposure to that specific form, suggesting that they are creating the form ‘wugs’ on the basis of a rule contained in their computational system. Similar processes are shown to be active in the derivation of invented verbs (Berko, 1958, Marcus et al.., 1992). Further crosslinguistic evidence suggests that children across the world experience a period in which they over-regularise non-regular verbs (Guasti, 2004, Marcus et al., 1992). Forms such as “blowed” or “comed” are typical in English, for instance (Gleason & Ratner, 2009). These forms appear to be derived through the (wrong) application of the rule “+ed” on stems stored in memory.
Thus, according to Berko (1958) and Pinker and Ullman (2002a) the derivation of verbs takes place through the application of a rule. In production, this consists of the adding of bound morphemes by a computational system. In perception, this consists of the identification and isolation of bound morphemes from stems, a process called “morpheme stripping”. There is some evidence that the process of morpheme stripping is not blind (i.e. the parser does not strip the morpheme without analysing the stem as well). Evidence for this comes from studies of the sublexicon, a level of word processing in which meaning is not activated. In a visual word recognition task, Caramazza, Laudanna and Romani (1988) showed that the same morpheme takes more time to be stripped if applied to a real stem than if applied to a non existing stem. 1) cantevi 2) canzevi 3) cantovi 4) canzovi Examples 1-4 are taken from an Italian study. In (1) and (2) the same morpheme, “evi”, a real morpheme in Italian applied to verbs belonging to a category called Conjugation II, is applied to two different stems, creating two different non words. The first nonword, “cantevi”, is composed of the existent stem “cant” and the existent morpheme, “evi”. However, the two elements belong to different conjugations and cannot combine. The second nonword in (2), “canzevi”, is composed of a nonexistent stem, “canz” and of the existent morpheme, “evi”. Results showed that it took longer to judge (1) as nonword than (2). This suggests that the conflict between stem and affix affects the decision, i.e. that the analysis of the
bound morpheme is synergistic to the analysis of the stem. In (3) and (4) the same stems are used, but the morpheme applied is non-existent. Results showed that nonword recognition for (3) is less accurate than (4). Again, this suggests that the analysis of the affix is not blind. If it had been, the judgment of the two impossible affixes would have led to the same performance, i.e. participants would have been able to assess whether these were words or nonwords in the same way by checking the affix only. The result, instead, suggests that the analysis of the affix is always synergistic to the analysis of the stem. The nature of this synergy is, however, poorly understood: we do not know whether stripping takes place in response to purely morphological rules; whether the meaning of the stem is activated; or whether morphology interacts with phonology in word final positions. An interesting aspect emerging from pioneering work by Caramazza et al. (1988) pertinent to this discussion is that some form of morpheme stripping takes place even when we do not access the lexicon. This idea is integrated in recent connectionist models of reading, such as that proposed by Grainger and Ziegler (2011). According to this model, while reading, two different forms of sublexical decoding take place through two different systems: one is fine-grained orthography, defined as a system that decodes single graphemes into phonemes and small groups of letters into bound morphemes (in other words, a “morpheme probe system”). The other one is coarse-grained orthography, that performs whole-word processing, similarly to what was already suggested in Coltheart et al. (2001) in the dual route model of reading (which will be discussed in detail in section 5.2). The morpheme probe system in the fine-grained orthography makes it easier to spot affixes, such as the sequence “–ed” in English words. The model proposed by Grainger and Ziegler
(2011) opens an interesting experimental question. It is possible that a morpheme probe system of this nature is active in human’s perception, but no study has investigated this issue directly so far. Morpho-phonological
neurophysiological methods (Scharinger, Lahiri & Eulitz., 2010; Shtyrov & Pulvermüller, 2007; Krott & Lebib, 2013). Shtyrov and Pulvermüller (2002) elicited the mismatch negativity (MMN) component using event related potential recording in response to auditory presentation of the forms comes (frequent)/come (infrequent) and come (frequent) /comes (infrequent). Both conditions generated an MMN. However, the condition in which “comes” was infrequent showed an MMN with longer latency, which suggests that more time was needed to process the inflection morpheme, and therefore that it must take longer to access the area of the brain in the left perisilvyan which is known to store inflection morphemes. In a similar experiment, Pulvermüller and Shtyrov (2003) showed that the auditorily presented ungrammatical fragment “we comes” generates a larger MMN than the grammatical fragment “we come” when presented as the deviant (infrequent) stimuli in an oddball paradigm. Shtyrov, Pulvermüller, Näätänen & Ilmoniemi (2003) showed that in morphosyntactic minimal pairs, MMN is enhanced when the second word contains a morpheme that makes the inflection ungrammatical. Bakker,
Pulvermüller & Shtyrov (2013) showed that auditorily presented nonwords generated by over-regularization, such as “flied”, generate bigger MMN than phonologically comparable nonwords, such as “smide”, suggesting that a form of morpheme stripping is: active, contemporary to the analysis of the stem and it also suggests that the process is pre-attentive. Leminen, Leminen, Kujala, & Shtyrov (2013) showed a
contrast in the MMN generated by auditorily presented derivational and inflectional morphemes, a contrast that suggests that derived forms are stored and inflected forms are not. 6 In summary, there is evidence that inflected forms may be stored as units (Stemberger & MacWhinney, 1988; Bertram et al., 2000; Tomasello, 2006; Diessel, 2012). But there is also contrasting evidence to suggest that stems and inflection morphemes are stored separately (Pinker & Ullman, 2002a; Guasti, 2004; Berko, 1958). There is also evidence that some form of morpheme stripping takes place at the sublexical level (Grainger & Ziegler, 2011; Caramazza et al., 1988), and that this stripping is not blind but synergistic to the analysis of the stem (Caramazza et al., 1988). There is, however, poor understanding of the nature of these sublexical processes. The aim of the current study is to investigate the contribution of morphology and the contribution of frequency effects in the perception of lexical and sublexical items containing bound morphemes. Three different experiments are performed with three different methods: 1) behavioural effects of morpheme stripping phenomena on lexical items are investigated using reaction times (RT); 2) neurophysiological effects of morpheme stripping are investigated using ERPs; and 3) morpheme stripping at the sublexical
There is also relevant evidence from other components: Newman et al. (2007) showed that regular verb violations generate Left Anterior Negativity (LAN) while irregular verb violations do not, suggesting the presence of a dual system of inflection (Pinker & Ullman, 2002a). A recent study by Krott and Lebib (2013) on past participles shows the ERP component known as Left Anterior Negativity (LAN) normally elicited by morphological violations can be generated also by regularly inflected verbs, and also that lexical frequency does not play a role in the elicitation, suggesting that rule based mechanisms are active when the processing of inflected verbs take place.
level are investigated using nonwords and reaction times. The hypothesis for the three experiments are summarised in Table 4.1. The three sub-hypotheses marked A are based on whole word processing explanations, such as Stemberger and MacWhinney (1988) or Bertram et al. (2000): In contrast, the three sub-hypotheses marked B are based on morpheme stripping explanations, such as Pinker and Ullman (2002a).
Table 4.1 Hypotheses based on two theoretical positions are presented for each experiment.
Hypothesis 1 time needed to discriminate lexical and morphosyntactic minimal pairs will be predicted by lexical frequency with shorter reaction times for items with high frequency time needed to discriminate lexical and morphosyntactic minimal pairs will be predicted by morphological complexity, with longer reaction times for pairs containing bound
Hypothesis 2 the neurophysiological response to lexical and morphosyntactic minimal pairs will be predicted by lexical semantics (lexical minimal pairs will generate bigger Mismatch Negativity MMN) the neurophysiological response to lexical and morphosyntactic minimal pairs will be predicted by morphological complexity (morphological minimal pairs will generate bigger MMN)
Hypothesis 3 time needed to discriminate non-words with and without potential morphological information will be predicted by phonotactic probabilities with shorter reaction times for items with high phonotactic probabilites time needed to discriminate nonwords with and without potential morphological information will be predicted by the presence of potential morphemes (longer reaction times if potential morphemes are present)
4.2 Methods 4. 2. 1.Method: Experiment IV - Lexical and morphosyntactic minimal pairs Participants: Twenty adult participants (17 Female, 3 Male) were recruited through the SONA system of the University of Reading, with students receiving one credit for participating in the experiment. Students were undergraduate students in psychology, mean age 19 years and 9 months, SD 1 year and 6 months. Procedure: The experiment consisted of a same/different task with lexical and morphosyntactic minimal pairs presented aurally. If the two words were judged to be identical, participants were asked to press a white button; if the two words were judged to be different, participants were asked to press a black button. The task consisted of 80 trials. Half of the paired words were identical and half were different. Stimuli: There were two types of minimal pairs: lexical and morphosyntactic. By definition, minimal pairs are pairs of words that differ in only one phonological element so the two words have a different meaning (Roach, 2000). In many languages, bound morphemes which are used to mark inflection generate minimal pairs, or miniparadigms (Guasti, 2004). These types of pairs are called morphosyntactic minimal pairs (Law & Strange, 2011). In the morphosyntactic condition of our experiment, the contrast was generated by bound morphemes, for instance “asked” vs “asks”, pronounced in standard BBC English: /ɑːskt/ - /ɑːsks/. In the lexical condition, minimal pairs were carefully chosen so that: 1) the difference was always in final position, in order to be comparable to the morphosyntactic condition; and 2) the contrast was always generated by morphemes that are potentially morphological in English, i.e. /z/ & /d/ or /t/ & /s/, for instance “slight” vs
“slice”, “tact” vs “tax”, “right” vs “rice”. All words in this task were monosyllabic. Examples are provided in Table 4.2 below. For a complete list see the appendix.
Table 4.2 Examples of lexical and morphosyntactic minimal pairs.
Lexical minimal pairs
Phonetic /ɑːskt/ - /ɑːsks/ (i.e. transcription asked vs asks) (and meaning in /keəd/ - /keəz/ (i.e. parenthesis) cared vs cares) /laikt/ - /laiks/ (i.e. liked vs likes)
Morphosyntactic minimal pairs /slaɪt/ - /slaɪs/ (i.e. slight vs slice) /nəʊd/ - /nəʊz/ (i.e. node vs nose) /tækt/ - /tæks/ (i.e. tact vs tax)
Stimuli were recorded in a sound booth by a trained linguist whose first language is English. The length of strings (words) were measured analogically using Audacity to ensure that, in each condition, the length of the overlapping string of phonemes was similar. 4. 2. 2 Method: Experiment V - MMN and lexical/morphosyntactic pairs Participants: Twenty-two right handed, native speakers of English were recruited through the SONA system of the University of Reading, and received one credit for attending the experiment. Participants were undergraduate students in Psychology, 21 Female and 1 Male, mean age 21 years and 8 months, Standard Deviation 1 year and 6 months. The experiment was conducted in the EEG lab of CINN – Centre for Integrative Neuroscience and Neurodynamics, University of Reading.
Procedure: This was a Mismatch Negativity (MMN) task with minimal pairs. As in experiment 1, we compared lexical and morphosyntactic minimal pairs. In this type of task, the same stimulus needs to be repeated a large number of times, and, for this reason, we used a reduced set of eight stimuli evaluated as prototypical of the condition tested. In MMN tasks, participants are presented repeatedly with a frequent stimulus (standard condition) 80% of the time (Näätänen, Paavilainen, Rinne & Alho , 2007), while an infrequent stimulus (the deviant stimulus) is presented in the remaining presentations with randomisation of presentation of frequent and deviant stimuli. This typically leads to a difference in the averaged waveforms of the standard and deviant stimuli at 100-150 ms after stimulus presentation, with the deviant waveform typically producing a larger response than the standard stimulus (Näätänen et al., 2007, Steinberg, Truckenbrodt, & Jacobsen, 2011). Stimuli ending in /d/ were the “standard” in our task, while stimuli ending in /z/ were the “deviant”. For instance, in condition A (lexical pairs), participants were presented with the word “side” in 80% of presentations and with the word “size” in the remaining 20% of presentations. In order to keep participants attentive, they were asked to press the space bar as fast as they could whenever the noticed the deviant stimulus (please note that in our task it is not the MMN itself that is relevant, but the comparison between MMNs. Thus, the motor response potentially recorded in the deviant condition does not present a risk of artefacts). Reaction times to deviant stimuli were also measured. All participants were presented with both conditions with the order counterbalanced across participants. EEG was recorded using 32-electrodes conventional caps (Easycap, Brain Products Gmbh). Impendence was maintained below 5KΩ. The recordings were
amplified by two BrainVision amplifiers with a bandpass filter of 0.1 to 200 Hz, and digitized continuously with a sampling rate of 0.5 kHz. Reference was recorded from the fronto-parietal position. Data were recorded using the BrainVision recorder (Brain Products Gmbh). Artefact rejection was performed using the software BrainVision Analyser (Brain Products Gmbh). Details about artefact rejection are provided in the analysis section. Stimuli: Stimuli were chosen from Experiment 1 according to measures of syllabic structure: We maintained the same structure across the two conditions; and we avoided consonant clusters because it would have been impossible to have the same clusters in all final positions, and there is evidence that different types of clusters require different forms of processing (Kirk & Demuth, 2005). We opted for the structure CVC, where the final C represented a bound morpheme in the morphosyntactic condition. For the morphological condition we selected the pairs “cared/cares” and “chewed/chews”. For the lexical condition we chose the pairs “side/size” and “bud/buzz”. All words in this task, thus, contain the required CVC structure. The length of vowels (or diphthongs) is different in the four conditions, but this was taken into account in the data analysis (see Results section). The stimuli are summarised in Table 4.3. Table 4.3 Lexical and morphosyntactic minimal pairs used in the EEG experiment.
Lexical minimal pairs
Morphosyntactic minimal pairs
side – size bud – buzz
cared – cares chewed – chews
4. 2. 3 Method: Experiment VI - nonwords and morpheme stripping Ethics, recruitment and consent: the current study was approved by the University of Reading Research Ethics Committee and it was given favourable opinion to proceed. The study was advertised on the School of Psycholgy and Clinical Langauge Sciences SONA system. SONA is a cloud-based participants’ recruitment service that allows the connection between researchers and students within an institution. Students received one credit for their participation in this study. The online page advertising the study contained information about the task and the research project. Students were able to decide voluntarily whether to participate in this study by enrolling in one of the slots available for testing. Once they arrived at the testing room (which was in the School of Psychology and Clinical Language Sciences), participants were asked to sit comfortably and reread the information sheet. After that, they were asked to sign a consent form if they agreed to take part in the study and to write their date of birth. All data were stored in a locked filing cabinet. Participants were allocated a numeric identifier which was used to anonymise the data. The information linking participants to this numeric identifier was stored in a separate and secure location. Participants: Twenty adult native speakers of English were recruited (the same people participated in experiments 4 and 6). They were undergraduate students in psychology, mean age 19 years and 9 months, Standard Deviation 1 year and 6 months. Seventeen were female, three were male; The order of presentation was balanced, such that half of the participants completed experiment 4 and then experiment 6, and the other half completed experiment 6 and then experiment 4).
Stimuli: in this experiment, we investigated morpheme stripping effects using sublexical items. In English, regular verbs ending in /l/ “take” the /d/ ending when inflected in the past (e.g. kill - killed) and the ending /z/ when inflected in the third person present (e.g. kill - kills). Although morphological in other contexts, /t/ and /s/ do not bring grammatical information when following /l/. We compared discrimination of nonwords ending in /ld/ vs /lz/ and nonwords ending in /lt/ vs /ls/. All nonwords were deemed phonotactically legal using the Vitevich and Luce (2004) calculator (see appendix 2 for detail). Stimuli were created using the following procedure: first of all, 4 starting consonants were chosen. These were: /v/, /n/, /θ/, and /dʒ/. The choice was motivated by two factors: all these consonants are allowed in word initial position in English, as it is shown by the fact that the positional segment frequency value for these consonants in initial position is never zero. Positional segment frequency is a statistical measure obtained through the analysis of corpora. The measure indicates how often a specific phoneme appears in a specific position in words. For instance, /v/ has a positional segment frequency of .02 if used in word initial position. This means that /v/ appears in word initial position in 2 words out of 100. If /v/ was not allowed in word initial position in English, its positional segment frequency as a word beginning would have been zero. Positional segment frequencies for the word beginnings chosen for this study is shown in Appendix 2. As seen in Appendix 2, this value is never zero in the entire test. In the meantime, the consonants chosen have a relatively low frequency in word initial position. In fact, the values of positional segment frequencies vary between .02 and .006 for word initial phonemes in our test. The choice of having word beginnings with a relatively low frequency was an advantage in terms of nonwords generation.
In fact, having infrequent word beginnings substantially reduced the risk of “creating” existing words. All words in this task are monosyllabic and as such there is only one vowel per word. The vowels used in this experiment are the following: /ɪ/, /aɪ/, /æ/, /ɔ/, /ʌ/. The choice of these vowels was motivated by biphone segment frequencies and positional segment frequencies. All these vowels are allowed in the second position of a word and all of them are allowed as the second phoneme of a biphone having any of the consonants presented above as a first phoneme. The fact that they are allowed in second position is demonstrated by the fact that the positional segment frequency of these vowels is never zero. The values can be checked in Appendix 2 (the positional segment frequency of the vowel is the second value from the left reported in each box). The fact that these vowels are allowed as members of a biphone having one of the consonants presented above in initial position is demonstrated by the fact that the biphone segment frequency of these biphones is never zero (values below the first box in the appendix 2). All these values are given in Appendix 2. The onset and nucleus of the nonwords were then combined with the potentially morphological codas presented at the beginning of this section:. /lz/ and /ld/, and non-morphological codas, /ls/ and /lt/. The nonwords were created using rules that allowed us to combine onsets, nuclei and codas. The productive rules for creating the nonwords was that each onset was combined with each nucleus. This enabled us to obtain 20 base forms. The four different codas were codas added to each base form, thus generating 80 nonwords. Forty of these contained potential morphological information, and 40 did not contain morphological information.
A third control condition was added to control for voicing effects. Without the third condition, the contrast between potentially morphosyntactic and non morphosyntactic minimal pairs could be explained by the fact that the two final phonemes in the first condition are both voiced, while the two final phonemes in the second condition are not. With the third condition we exclude this possibility. In the third condition, the two final phonemes are both voiced but they do not carry morphological information. The codas used in the control condition were the following: /lb/ and /lm/. The base forms were also applied to these codas to create the control condition, leading to a further 40 nonwords. Thus the final test contained 120 nonwords. A summary of the type of stimuli used in the test is presented in Table 4.4. The full list of stimuli is available in Appendix 2 in two versions: one in phonetic transcription, the other one in software readable transcription (Vitevich & Luce, 2004), with specification of positional segment frequencies and biphone frequencies. Condition
Table 4.4 Pairs with and without potential morphological information. Potentially Non Voicing control condition morphosyntactic morphosyntactic minimal minimal pair pairs
/vɪld/ - /vɪlz/
/ vɪlt/ - /vɪls/
/vɪlb/ - /vɪlm/
Features of the nonwords
plosive/fricative voicing coherent morpho ending
plosive/fricative voicing incoherent non morpho ending
plosive/nasal voicing coherent non morpho ending
Although the productive rules used ensured the generation of a large number of nonwords, a few nonwords were created with different vowels because the
productive rules led to the generation of real words. Specifically, in the block starting with /dʒ/, /ɑ/ was chosen instead of /ɪ/ because the use of /ɪ/ would have lead to /dʒɪlz/, which is an existing word (the plural of Jill, or the possessive form for Jill). In the block starting with /v/, /ɛ/ was used instead of /aɪ/ because using /ɛ/ would have led to one of the nonwords having one of its values of biphone segment frequency equalling zero. Recording of the stimuli: The stimuli were recorded in the sound booth of the School of Psychology and Clinical Language Sciences by a trained female linguist whose first language was English. The linguist was instructed to record them in pairs. In reference to the phonetic list in Appendix b, this corresponds to reading the words row by row. The linguist was informed about the nature of the task, thus recording pair by pair to ensure the recording of a subtle vowel lengthening in the morphological condition, as it is typical in British speakers of English when producing inflected verbs. The software used was Audacity, running on a computer using Windows. The microphone was an AKG D80, the mixer was a Behringer Mini Mon, the pre-amplifier was a B-tech phono-microphone pre-amplifier, and the speakers were Sony SS-CMD373. Procedure: After signing the consent form, participants were given spoken instructions about the task by the researcher. The information they received was the following: “You are now going to hear pairs of words. The two words presented in each pair may be identical or slightly different. Press white when you think the two words are identical, black when you think they are slightly different. Try to be as quick as you can”- After that, E-prime was launched and the participant was left alone in the room. Instructions on the screen guided the participant through the
testing session. Similarly to Experiment 4, participants were presented with a same/different task. Experiment 6 was conducted using nonwords and contained 180 trials, 3 conditions and 60 items per condition. Each trial consisted of the presentation of two nonwords that could be either identical or that could differ in the final phoneme. The first slide was a fixation slide, lasting 1000ms, containing only the symbol “+”, presented in white in the centre of a black screen. The second slide was labelled in E-prime as Soundout1. It lasted 1000 seconds and corresponded to the presentation of the first nonword. While presenting the first nonword, the screen appeared completely black. The third slide was labelled in E-prime as Soundstimulus. It contained the second nonword and it lasted 1000ms. While presenting the second nonword the screen was completely black. During the presentation of this slide participants pressed “black” or “white” to express their judgment on the similarity of the nonwords. The fourth slide informed the participants that they were moving to the next trial, and was composed of an arrow -> presented in the centre of an otherwise black screen. The order of presentation of the trials was randomised using the random function of E-prime. As a consequence, the order of presentation of the trials was different for each subject. Each pair was never presented to the same participant more than once. Scoring: E-prime was set to record the answer given (either “black”, “white”, a non valid key or no answer). A no answer was coded when participants did not press any key for the entire duration of the third slide (1000ms). For any type of given answer, E-prime measured the time (in ms) that the subjects took to make their choice and press the button.
Outcome measures: The task was administered to assess the ability to discriminate nonwords forming minimal pairs. Since previous evidence showed that the processing of morphological information requires more time, we coded the responses in order to have a time measure of successful discrimination. In order to do that, we calculated for each participant the average reaction times (RTs) needed to successfully discriminate elements in the three different conditions. The average RTs were calculated by dividing the sum of the RTs of trials in which the participant successfully discriminated different nonwords by the number of successful discriminations. We also coded item-based reaction times. We did this by dividing the sum of RTs obtained across participants for a successful discrimination of a nonword by the number of times that nonword was successfully discriminated from its prime.
4. 3 Results 4. 3. 1 Results: Experiment IV The time needed to discriminate elements in lexical and morphosyntactic minimal pairs was compared using a paired samples t-test. An average time per stimulus pair was calculated for each participant by dividing the sum of RTs for correct answers by the number of correct answers, in the condition in which the elements in the pair were different. The paired samples t-test shows that elements in morphosyntactic minimal pairs took longer to be processed than elements in lexical minimal pairs, Paired samples t-test: t (19) = -11.985, p < .001. The result is presented visually in Figure 4.1. We also conducted item-based correlations between average RT for a specific word and frequency of that word in the British National Corpus. The correlation was not significant r = 0.004, p > .05.
Figure 4.1 Reaction times for the discrimination of elements in lexical and morphosyntactic minimal pairs.
4. 3. 2 Results: Experiment V One participant was excluded because of a recording error, and so analysis was conducted on the remaining 21 participants. Data were segmented according to stimulus onset times. The time limit was set from -200ms to 800ms in relation to the onset of the stimulus. Artefacts were then removed using a standard semiautomatic procedure. Two conditions were set: A) check gradient: max voltage step = 100 μV. Time window: Before -200ms – after 800ms. With this function the software gets rid of all the parts of the recording in which the step between two adjacent datapoints was larger than 100 μV. In fact, a step of this type cannot be generated by brain activity and indicates the presence of artefacts. B) check maximal and minimal: MIN = -100 μV, MAX = 100 μV. With this function the software gets rid of all the parts of the recording in which the amplitude values are lower than -100 μV or higher than 100 μV. These levels of amplitude cannot be generated by brain activity and indicate the presence of artefacts. Before -200ms – after 800ms (0 being the beginning of the word, see figure 2 for a graphic exemplification) was the time window used. These values indicate the amount of recoding used to perform the artefact rejection. Locking the epoch timeline to the beginning of words, one second of recording was used. The second started 200ms before the beginning of the word and ended 800ms after the beginning of the word. Baseline correction was performed, with the limits -200ms – 0ms. These values indicate the time window in which baseline correction was performed. Baseline correction is a standard procedure performed to align the amplitudes of different waves at the time 0, in order to make
amplitudes comparable across conditions. Two separate data sets were then averaged for each condition for each participant and each electrode: one analysis was conducted for the average of amplitudes (μV) for the deviant stimuli, the other for the average of amplitudes (μV) for the standard stimuli.
Figure 4.2 Example of sound wave. The flat line in the end of the first word indicates the beginning of the plosive consonant and thus the disambiguating point.
In all stimuli, the length of the vowels forming the nucleus of the syllable (and, to a lesser extent, the length of the onset consonants) is different, and, as a consequence, the disambiguating point varies across conditions (see, as an example, the disambiguating point in the pair “side/size”, Figure 4.2). The position of the disambiguating point was calculated for each stimulus. In order to do that, the software Audacity was used. As shown in Figure 4.2, the production of plosive consonants (for instance /d/ in “side”) is characterised by a short moment of silence. In fact, the very nature of plosive consonants consists in a moment in which air is completely constrained in the mouth before plosion is allowed and a sound is uttered.
That moment of silence is represented in the wave by a complete flattening (see Figure 4.2, in proximity of the black vertical line). When measuring the distance between the flattening and the beginning of the word we obtain a measure of the disambiguating point (in fact, that moment of silence is the first difference in the sequence of phonemes of the two pairs). Then, 150 data points (measuring 300ms) following the disambiguating point were used for the data analysis. The measure of 300ms was chosen because mismatch negativity is obtained on average at a latency of 150ms, thus it seemed reasonable to keep the 150ms between the expected ERP and the disambiguating point, as well as an equal amount of time (150ms) following the disambiguating point. An analysis of negativity peaks was carried out. Since several authors reported the results of MMN in lexical tasks with respect to signals recorded from frontal electrodes (Näätänen et al., 2007, Cornell, Lahiri & Eulitz, 2011, Gunawardena, Krikeb, Moelijker, Puppi & Witteveen, 2011, Hagoort & Brown, 2000, Pulvermüller, Shtyrov, Kujala, & Näätänen, 2004, Angrilli, Dobel, Rockstroh, Stegagno & Elbert, 2000, Shtyrov, Kujala & Pulvermüller, 2010,), we will also report measures only from frontal electrodes. The analysis was performed on the average of the 3 frontal electrodes, Fp1, Fp2 and Fz (i.e. electrodes cluster, Rugg & Curran, 2007, Rinker et al., 2007). Peaks were calculated individuating the smallest value (MIN) in terms of amplitude in the amplitude list of each participant in the 300ms following the disambiguating point (more about the disambiguating point above), for each condition. In fact, the MMN represents a negativity in the amplitude, i.e. a very small
value in the amplitude occurring after the disambiguating point (normally, between 100ms and 200ms after the disambiguating point). Latency was obtained by extrapolating coordinates (position in the list) of the MIN value. This is because the position of MIN in the list indicates the latency. The following lines report an example that exemplifies the rationale for this type of analysis. A MIN occurring in position 80 indicates a latency of the peak of 160ms. In fact, each datapoint represents 2ms, because amplitude was measured every 2ms. If the MIN value, i.e. the smallest amplitude value, occurred after 79 measures of amplitude, 160ms had to pass before the amplitude could reach its smallest value. The time needed to get to the MIN value indicates the latency of the MIN value, or, in other words, the latency of the peak of the negativity. Peak amplitude was compared across conditions using a 2X2 ANOVA, with MMN (standard/deviant) and TYPE (lexical/morphological) as factors. There was a significant main effect of MMN, F (17, 1) = 62.61, p < .001, and a significant main effect of TYPE, F (17, 1) = 4.72, p = .044. There was also a significant interaction between MMN x TYPE, F (17, 1) = 8.11, p = .01. Post-hoc analyses show that the difference in peaks in the deviant conditions is driving the effect: planned comparisons showed no significant difference between maximum amplitude for frequent lexical and morphological stimuli, t (17) = .034, p > .05, but a significant difference for maximum amplitude to deviant lexical peaks and deviant morphological peaks, t (17) = 2.73, p = .01. The comparison is shown in Figure 4.3. The difference in amplitude between points referring to the same type (lexical or morphological) indicate that both conditions produced a negativity (the difference between standard and deviant was always significant). The graph also shows the
nature of the interaction. While there is no significant difference in amplitude between lexical and morphological items in the standard condition, morphological items manifest a stronger negativity in the deviant condition.
Figure 4.3 Average of peaks.
As shown in Figure 4.3, MMN was obtained in both conditions, but the effect was stronger when the minimal pairs were morphosyntactic (the difference in amplitude between standard and deviant is larger in the morphological condition than in the lexical condition). We then analysed the contrast between mean amplitude in the lexical condition and mean amplitude in the morphological one (Handy, 2005, p. 8; Luck, 2005, p. 229). Specifically, we conducted an analysis on grand-averages, comparing averaged amplitudes in the two conditions (i.e. calculating the mean value on the 150 data points following the disambiguating points). We found a main effect of MMN, F
(149,1) = 285, p < .001, and a main effect of TYPE, F (149, 1) = 123, p < .001. The interaction MMN, TYPE was also significant, F (149, 1) = 1440, p < .001. Post-hoc analyses showed that the change in mean amplitude between standard and deviant stimuli takes place only for morphological pairs: comparison of mean values of standard and deviant lexical amplitudes, t (149) = -1.173, p > .05; comparison of mean values of standard and deviant morphological values, t (149) = 26.367, p < .001. The comparison is shown in
Figure 4.4. The lexical condition shows
no change in amplitude between standard and deviant conditions, indicating a lack of mismatch negativity. In contrast, there was a significant difference in amplitude of the MMN in the morphological condition, with a larger MMN response in the deviant condition.
Figure 4.4 Averages of mean amplitudes.
Below, a graphic representation of the brain activity in frontal position is presented. Figure 4.5 shows waves from 200ms preceding stimulus presentation to
800ms following stimulus presentation. Data come from the morphosyntactic condition. A widespread negativity is observed in the deviant condition. All electrodes show a negativity for the deviant condition, peaking around 660 ms after the beginning of the word, i.e. after 160 ms following the presentation of the bound morpheme. The difference between the two lines in each plot indicates the difference in amplitude between standard and deviant condition. The graph shows a widespread negativity for deviant stimuli, reported in 5 frontal electrodes: Fp1, Fp2, F3, FZ. F4. Figure 4.6 presents grand-averages in the 4 conditions in the 300ms following the disambiguating point. Only the measure of the activity in the morphological deviant condition is consistently below the other measures. While the difference between the two lexical waves is not significant (red and blue lines, Figure 4.6) the difference between the standard and the deviant waves is significant, with deviant stimuli producing a widespread negativity, peaking around datapoint 80 (i.e. 160 ms after the disambiguating point, since each datapoint represents two ms)
Figure 4.5 Example of MMN over time in the frontal region with morphosyntactic minimal pairs.
Figure 4.6 ERP components elicited by the four conditions in the 300ms following the disambiguating point, averaged in the cluster of frontal electrodes.
4. 3. 3 Results: Experiment VI Experiment 6 used a same/different paradigm. Descriptive statistics are presented in Table 4.5.
RT Morphological RT non 20 morphological RT control 19
Table 4.5 Descriptive statistics experiment 6. Minimum Maximum Mean SE 698.09 972.25 892.37 14.86
There were three conditions defined by the type of information carried by the items: Condition 1 in which the elements carry potential morphological information; Condition 2 in which the elements do not carry morphological information, but the word endings could potentially be morphological in isolation; and Condition 3 in
which the elements do not carry morphological information, and voicing effects are controlled (Table 4.5). A D-prime analysis was carried out to check whether participants engaged significantly in the task, using hit and false alarm rates, as suggested by MacMillan and Creelman (2005). These measures refer to the number of times particiapants correctly detected a contrast and the number of times participants perceived that there was a contrast while the two stimuli were actually identical. Using these two measures, it is possible to assess whether participants engaged in the task and whether they answered randomly or not. D-prime values are significantly different from 0, t (19),= 7.55, p < 0001, two tailed, and significantly different from 1, t (19), = 4.47, p = 0003, two tailed. According to MacMillan and Creelman (2005) this result indicates that participants did not answer randomly to the stimuli and that they were accurate in at least 70% of cases. Since there were 3 conditions and 1 group of participants in the study, reaction time data were compared using a 3 factor ANOVA (the three factor refers to the three conditions: presence of potential morphological information, absence of morphological information and control condition). Since the task was meant to assess the ability to discriminate minimal contrasts, we focused the second part of the analysis on pairs in which the two nonwords were different. There was a significant difference between the three conditions, F (15,2) = 17.7, p< .001. Pairwise comparisons show that elements with potential morphological information took more time than non-morphological elements to be discriminated, t (18) = 6.31, p < .001, and a significantly larger amount of time than the phonological control condition, t (17) = 2.56, p = .02. Further, they show that elements in the non-morphological condition were discriminated more quickly than
elements in the phonological control condition, t (16) = -4.65, p < .001. The results are presented visually in Figure 4.7.
Reaction times (ms)
880 860 840 820 800 780 760 740 720 1
1. Morphological 2. Non-morphological 3. Phonological control
Figure 4.7 Reaction times for the discrimination of elements with and without morphological information (+ control).
Item-based correlations were run between RTs and rhyme likelihood, biphone segment frequency and positional segment frequency. First of all, item-based RTs were calculated. In order to do that, the number of correct answers for each specific nonword was calculated. Then, the sum of RTs across participants for each specific nonword was calculated, and divided by the number of answers. The values obtained for each nonword were then correlated with the rhyme likelihood, biphone segment frequency and positional segment frequency relative to that same specific item. This was done in order to test the hypothesis 3A (see Table 4.1). None of the correlations were significant (with Bonferroni correction). See Table 4.6 below for details:
Table 4.6 Correlations item based reaction times and PSF, BSF (obtained with British and American corpora) and Rhyme likelihood (obtained with British Corpora). ItemRT PSF Brit BSF Brit Rhyme PSF Ame BSF Ame Pearson’s 1 -2.43 0.09 0.009 -2.61 0.092 correlation ItemRT Sig .062 .492 .944 .044 .485 ItemRT
Key: PSF = Positional segment frequency, BSF = biphone segment frequency
Finally, a post-hoc power analysis was performed to obtain the achieved power. First, the η2 was calculated by dividing the sum of squares relative to the conditions in the test of within-subject effects by the total sum of squares. The total sum of squares was calculated by summing the sum of squares relative to the conditions in the test of within-subject effects, the sum of squares associated with the error and the sum of squares of the error in the test of between-subjects effect (methodology from the website PsychoHawks). The values used are reported below: η2= 96080 / (96080 + 65921 + 108055) = 0.35 Since G-power uses r instead of η2 as a measure of effect size in ANOVA, the r value was calculated using the following formula: f = √ η2 / (1 – η2). The calculation is presented below: f = √ 0.35 / (1 – 0.35) = √ 0.5 = 0.7 We then used the obtained value of the effect size (0.7), the α value (0.05) and the sample size of the experiment (20) to calculate the achieved power (1-β), which was 0.9.
4. 4 Discussion The aim of this block of experiments (4, 5 and 6) was to investigate morpheme stripping phenomena with lexical and sublexical items, using both reaction time measures and event related potentials. The main findings are that: 1) discriminating elements belonging to lexical minimal pairs requires less time than discriminating elements belonging to morphosyntactic minimal pairs; 2) MMN is bigger when there is a morphological contrast compared with a phonological contrast and 3) nonwords with potential morphological information take longer to be discriminated than those without morphological information. Each of these results will be addressed in turn. The first research question focused on the process of morpheme stripping with lexical items, i.e. whether we need more time to process pairs with morphemes than pairs without morphemes, and whether varying reaction times are a consequence of whole-unit frequencies. The experimental hypothesis B was confirmed in that reaction times were longer when participants had to discriminate between elements belonging to morphosyntactic minimal pairs compared to elements belonging to lexical minimal pairs. The item-based correlations showed that itembased reaction times could not be predicted by item-type frequencies, which suggests that hypothesis A should be rejected, at least when performing this type of task. The contrast between lexical minimal pairs can be interpreted as an instantiation of morpheme stripping effects in perception. The fact that elements in the morphosyntactic condition take more time suggests that the comparison of elements of this type does not result in a pure phonological comparison of stored strings of phonemes, but that, to a certain extent, the morphological information present is
processed separately. This is consistent with Caramazza et al.(1988)’s study with Italian nonwords, which found that the processing of nonwords with potential morphology takes longer than the processing of nonwords without morphology. In summary, results from Experiment 4 suggest that some form of morpheme stripping procedure takes place at the lexical level. Further discussion of Experiment 4 is provided after the discussion of Experiment 5, since the two experiments are partially interdependent. The second hypothesis (that the MMN response would be predicted by morphological complexity) was tested using an EEG study. Experiment 5 showed that the MMN generated by elements differing in the bound morpheme was larger than the MMN generated by elements differing from a semantic point of view (elements in lexical minimal pairs). This finding can be interpreted by appealing to the nature of MMN. There is evidence that MMN is larger when discrepancies between standard and deviant stimuli are bigger (Näätänen et al., 2007). When comparing two phonemes, for instance, the greater the number of features by which the two sounds differ, the larger the MMN these produce (Näätänen et al., 2007; Dehaene-Lambertz & Baillet, 1998). In this task, the phonemes contrasted were always the same in all four conditions, /d/ and /z/, thus a phonological features explanation would not suffice. If MMN was generated by semantics, the larger MMN should be obtained in the lexical condition, since elements in lexical minimal pairs are semantically more different than elements in a morphosyntactic minimal pair. However, if we assume bound morphemes to be processed separately from the stem, then the condition in which elements in a morphosyntactic minimal pair are compared corresponds, from the MMN negativity point of view, to an ideal condition
in which elements from a limited set are compared. Inflectional morphemes belong to the group of closed class words (Carnie, 2013): there is a limited number of inflectional morphemes (very small in English) and all morphological information possibly conveyable is included in this limited set of items. Comparing two items from this set, thus, corresponds to comparing elements with substantially different information, in terms of amount and nature. For example,s /d/ conveys information with regard to past tense marking; /z/ on the other hand, conveys information about present tense, marking for 3rd person singular and agreement. However, when the two phonemes, /d/ and /z/, do not carry morphological information, the nature of the contrast is substantially different. When we contrast /d/ and /z/ in a context in which they do not carry morphological information, the comparison has to be interpreted as occurring between elements of a much larger set, the set of all phonemes allowed in word final position. From this point of view, the phonemes /d/ and /z/ not bringing bound morphology are more similar to each other than the morphemes /d/ and /z/. The difference between the two is more nuanced when being part of a larger set (all phonemes allowed in word final position), than when the two elements are compared within a morpho-syntactic set. Interestingly, a viewpoint based on semantics makes a completely different prediction. Another possible explanation is that elements in morphosyntactic minimal pairs are semantically strictly related, which might lead to the prediction that MMN negativity should be bigger in the lexical conditions (Näätänen et al., 2007). However, the results of our study do not support this view. Therefore, we conclude that it is most likely that the morphological contrast is driving the MMN effect. The finding in Experiment 5 is crucial also in the interpretation of Experiment 4. The pattern observed in Experiment 4 may be
explained in two ways: by referring to semantics and by referring to morphology. When referring to semantics, the results may be explained by proposing that participants were quicker in discriminating elements belonging to lexical minimal pairs because those elements were semantically more different than elements in morphological minimal pairs. A morphological explanation, on the other hand, would be that participants took longer to discriminate elements in morphological pairs because the presence of the morphemes required a further step in speech perception.
If the semantic explanation of Experiment 4 was correct, then in
Experiment 5 we would have observed a larger MMN for lexical pairs. Elements in lexical pairs should have been perceived as different. This, however, is not the case, and a morphological explanation thus seems plausible in both Experiments 4 and 5. In Experiment 6, participants were presented with a same/different task using nonwords. The results show that elements carrying potential morphological information took longer to be discriminated than elements that do not carry potential morphological
characteristics suggests that a purely phonological explanation cannot be used. What is most intriguing, however, is the evidence that morpheme stripping is not blind, i.e. that the parser does not operate on the word ending in isolation, and there is evidence from the results that there is an interaction between phonology and morphology during stripping. In order to explain this interaction, the following examples are reported: 1) /vIld/ /vIlz/ 2) /vIlt/ - /vIls/
Word endings in (1) are potentially morphological. These are the endings that any English regular verb ending in /l/ would take. For instance, the verb to call when inflected in the past becomes called and when inflected in the present third person becomes calls, pronounced /kɔːlz/ . The endings in (2) can carry morphological information in English, for instance when following a regular verb ending in /k/, as ask. However, these endings cannot bring morphological information when following a regular verb ending in /l/. For instance, the application of /t/ on the regular verb kill does not lead to a regular inflection, but to the generation of a different noun, kilt. If morpheme stripping was blind, i.e. if the morpheme detector was simply looking for potential morphosyntactic information in the end of the word, then (1) and (2) would not be different. All four words end in a phoneme, which, in isolation, potentially carries morphological information. However, the fact that there is a significant difference in reaction times in the discrimination of elements in (1) and the discrimination of elements in (2) suggests that a different process takes place. Morpheme stripping does not take place on the word ending in isolation, but is integral to the analysis of the stem. The procedure by which inflectional morphemes are analysed most likely consists of a combination of the analysis of the bound morpheme/word ending and of the analysis of the stem/rest of the word. In particular, the parser must take into account morpho-phonological rules, i.e. it has to consider what type of bound morpheme is allowed considering the phonological characteristics of the stem. In sum, this experiment suggests that morpheme stripping takes place sublexically, and that the process operates by taking into account the analysis of the stem, and specifically, that stripping takes into account the analysis of the phonological characteristics of the stem.
The result of the item based correlation was rather surprising. Using software developed by Moreland (2011), we calculated, for each nonword, positional segment frequency, positional biphone frequency and rhyme likelihood. Positional segment frequency is the most theoretically ungrounded measure among the three. It measures how often a specific phoneme appears in a specific position in the corpus7. It treats positions in a purely ordinal way, with no relation to syllable structure. The issue can be explained using the following example: 1) stop 2) atrium According to this measure, the phoneme /t/ appears in position 2 in both (1) and (2). However, from a syllabic point of view, the two sounds are quite different. In (1) /t/ is the second consonant of a syllable initial onset cluster, in (2) /t/ is the first consonant of a cluster in a non-initial syllable. However, there is evidence that the processing of phonemes depends on syllabic rules rather than on ordinal rules (Moreland, 2011), i.e. evidence showing that we process phonemes as part of syllables and not autonomously. Considering the fact that this measure is theoretically ungrounded, we did not expect any correlation between positional segment frequency and item-based reaction times. Positional biphone frequency measures how often a chunk of two phonemes appears in a certain position. For similar reasons, it is not a theoretically grounded
The software used (Moreland, 2011) gives measures calculated on the Kucera and Francis (1967) American corpus, also known as the Brown corpus, as in Vitevich and Luce (2004), and also measures calculated on the British National Corpus, separately. Surprisingly, correlations with the values obtained through the American corpus revealed to be stronger, even if all participants tested in this experiment are of British nationality.
measure, even if it might be preferred, because it captures the likelihood of two phonemes being next to each other. Rhyme likelihood measures how likely a certain rhyme is to appear in a stressed syllable. This measure has a stronger theoretical foundation because it includes information on rhyme structure and stress, two phonological properties that are shown to influence phonological processing (Fudge, 1987; McCarthy, 1979; Scheer & Szigetvári, 2005). The limitation of offering only a measure on stressed syllables is also accounted for in this experiment, since all words in this task are monosyllabic and, as a consequence, all rhymes are stressed. Considering the fact that this measure is theoretically well-grounded, we expected to find a correlation between rhyme likelihood and item-based reaction times. Even if the rhyme likelihood measure was the most promising in the item based correlation analysis, results show that the only significantly predictive measure is positional segment frequency, r = -.26, p = .044. The correlation, however, is not significant after Bonferroni correction is applied. In short, it appears that there might be some likelihood and frequency effects (hypothesis A) but that they are not as effective as the morpheme stripping hypothesis (hypothesis B) in predicting reaction times in this task. In conclusion, with the first experiment, we reported evidence that the discrimination of elements in morphosyntactic minimal pairs takes longer than the discrimination of elements in lexical minimal pairs; in Experiment 2 we showed that morphological contrasts elicit a larger MMN component than lexical contrasts; and in Experiment 3 we showed that participants are quicker in discriminating nonwords without morphological information than in discriminating nonwords with potential
morphological information. Further, with Experiment 3, we also showed that stripping is not blind but synergic to the analysis of the stem. Overall, our data suggest that a form of morpheme stripping is taking place, and thus it suggests that inflected forms are decomposed in stems and affixes (Pinker & Ullman, 2002a). The presence of these effects with nonwords suggests that the proposal of Grainger and Ziegler (2011) for reading (that we can detect morphemes sublexically) may be extended to speech. The presence of some frequency effects with sublexical items suggests that a form of whole-form processing may be active as well, in line with claims such as Stemberger and MacWhinney (1988) and Bertram et al. (2000), although we report that these effects are smaller in speech perception than what was reported in lexical access tasks by the authors. It is possible that a redundant system, in which stripping coexist with whole-form processing, may accommodate our data and previous data available.
CHAPTER 5. IMPLICATIONS FOR THE ASSESSMENT OF SPECIFIC LANGUAGE IMPAIRMENT AND DYSLEXIA 5. Implications for the assessment of Specific Language Impairment and dyslexia
5. 1 Summary Sublexical representations are crucial during reading and our findings about sublexical processing may have an impact on the assessment of reading disorders. In this chapter, the clinical implications of our findings are discussed. The chapter is organised as follows: In the first part of the chapter, the relation between reading and sublexical representations is described. We then describe the most common reading disorder, dyslexia, together with a commonly comorbid disorder, Specific Language Impairment (SLI). Following this, two commonly used assessments for dyslexia and SLI, the Children’s test of nonword repetition (CNRep, Gathercole & Baddeley, 1996) and the Test of Early Grammatical Impairment (TEGI, Rice & Wexler, 2001), are analysed, in light of the findings reported in the previous chapters. Finally, the performance of 34 children on the CNRep is analysed, using the previous discussion as a guide for the analysis. The sample is composed of 9 children with dyslexia+SLI, 9 children with SLI only, 12 age-matched controls and 4 children with impairments of different origin. Our analysis shows that word position effects are present in all groups, although children with dyslexia+SLI are significantly poorer in their performance than children with SLI only and TDs.
5. 2 Reading and models of reading Sublexical representations are crucial in reading. In order to present the relation between reading and sublexical representation, a short introduction is needed. Reading is the process that allows the recovery and the comprehension of
information that is stored in written form. According to an influential model of reading, known as the Simple View of Reading (Hoover & Gaugh, 1990), the process is composed of two subcomponents: decoding, which refers to the process by which graphic symbols are transformed into speech sounds by the reader, and which relies substantially on phonological representations (Ramus et al., 2003), and linguistic comprehension. Both components are needed to read successfully. In the seminal work of Hoover and Gaugh (1990), this idea was investigated on a longitudinal sample of English-Spanish bilinguals from first to fourth grade. In this study, the authors show that both listening comprehension and the ability to decode can predict reading comprehension, but the combination of listening comprehension and the ability to decode is a significantly better predictor of overall reading comprehension. In less skilled readers, decoding and listening comprehension can be inversely correlated (Hoover & Gaugh, 1990, p.127, partially adapted). The authors conclude that reading can be defined as the product of listening comprehension and the ability to decode. The term decoding refers to the process by which graphic symbols are transformed into speech sounds by the reader, and it relies substantially on phonological representations (Ramus et al., 2003). The relation between decoding and phonological representations is the focus of a large amount of reading research. It is well established that the quality of phonological representations influences reading performance (for a review see Dehaene, 2009). In parallel, there is evidence that learning to read enhances the quality of phonological representations. In a study conducted by Reis and Castro-Caldas (1997), data are reported showing that learning to read and write generates new rules within the language system. These new rules
change the way in which some operations are performed. Their study was conducted on a group of elderly Portuguese citizens, who had not received formal training in literacy throughout their life. Their performance was compared to the performance of a control group of literate elderly people on several tasks. The authors found that illiterate people were worse than controls in typical tasks of phonological assessment, such as nonword repetition, the memorisation of phonologically related words, and the ability to create new words using a trained phonological rule. As the authors argue, illiterate people use strategies that are good for semantic processing, but their phonological processing is poor when compared to literate people. This study shows that familiarity with literacy significantly shapes our phonological processing system. Thus, phonological representations are intimately related to reading performance. According to connectionist dual-route models, lexical and sublexical phonological representations contribute differently to reading. The contribution of sublexical representations is crucial for this work, since this is the level at which the word beginning saliency principle and morpheme stripping operate, according to the results reported in Chapters 3 and 4 in this thesis. The most influential dual-route model of reading (Elliott, Braun, Kuhlmann & Jacobs, 2012) is the Dual-route cascaded model of reading aloud, by Coltheart et al. (2001). This is a connectionist model. Connectionism is an approach to psychological problems in which human behaviour is simulated using artificial networks implemented in computers. Networks are combinations of nodes, and each node receives information from other nodes and sends it to other nodes.
Nodes are simple and highly interconnected units. The basic functioning of a connectionist network is well summarised in this passage from Fodor and Pylyshyn (1988, p5): Connectionist systems are networks consisting of very large numbers of simple but highly interconnected “units”. Certain assumptions are generally made both about the units and the connections: each unit is assumed to receive real-valued activity (either excitatory or inhibitory or both) along its input lines. Typically, the units do little more than sum this activity and change their state as a function (usually, a threshold function) of this sum. Each connection is allowed to modulate the activity it transmits as a function of an intrinsic (but modifiable) property called its “weight”. Hence the activity on an input line is typically some non-linear function of the state of activity of its sources. The behaviour of the network as a whole is a function of the initial state of activation of the units and of the weights on its connections, which serve as its only form of memory). The types of information each node can send vary from model to model and also from node to node. The way each node processes information and the way nodes are connected defines the nature of each model, and defines what each model is able to do. Researchers developed connectionist models that are able to perform a relatively large number of tasks. To name a few relative to language, connectionist models can recognise letters, create new past tense forms of verbs and process sounds as sequences occurring over time (Thomas & McClelland, 2008). The model of Coltheart et al. (2001) accounts for the transformation of single written words into phonological representations. The system is able to read written words aloud and perform a lexical decision task. The model of Coltheart et al. (2001) is defined as a
cascaded model. The term cascade refers to the fact that each node sends information to subsequent nodes even when the amount of information available is small. Many other connectionist models, instead, use thresholds: information is sent to subsequent nodes only if a certain threshold is achieved. Interestingly, and quite unusually for a connectionist model, the structure is hardwired rather than learned with an algorithm. This is done in order to make the model in line with psycholinguistic findings on phonological representations. In fact, back-propagation algorithms may generate structures that do not reflect human structures. A backpropagation algorithm is an algorithm that changes according to the outcome of the processing. When the outcome of the processing does not comply with the expected outcome of the algorithm, the weights of the connections are modified, and the algorithm changes slightly. For instance, if the model is presented with the stimulus “pint” and the outcome is a different word (i.e. the system “reads” a different word), a negative feedback signal is sent back to earlier stages of the algorithm, and the weights of the connections are changed so that the model is more likely to perform correctly when reading “pint” the next time it is presented with that stimulus. One problem with backpropagation algorithms is that they are likely to be non-human in nature. When a child learns to read, they do not receive positive or negative feedback for every word they try to decode. For this reason, a hardwired algorithm such a Coltheart et al. (2001) presents an interesting alternative to connectionist models based on backpropagation algorithms. Crucially, the learning curve of the model of Coltheart et al. (2001) is not based on what is known as a “teacher signal” (McClelland & Cleeremans, 2009). In other words, the structure of the model does not depend on the feedback that the model receives at the end of each cycle, and it is not modified by a
signal that can teach the algorithm how to perform the task. The model, instead, is built on the basis of psycholinguistic evidence about reading. This approach is, at least, peculiar. In fact, historically, connectionist models appeared as an attempt to reply to cognitive models (Fodor & Pylyshyn, 1988). Crucially, cognitive models normally rely on the use of symbols, and levels of representation correspond to computations operated on these symbols (ibid.). Connectionism models, instead, do not need symbols, and they actually operate “pragmatically”, in function of the task they aim to solve without defining abstract units of representation such as symbols. In Fodor and Pylyshyn’s words (1988, p5): In contrast [to classical models of the mind], connectionists propose to design systems that can exhibit intelligent behaviour without storing, retrieving or otherwise operating on structured symbolic expressions. The model of Coltheart et al. (2001), instead, presents a hybrid nature: it is a connectionist model in that it uses nodes, connections between the nodes and weights, but it also contains elements from classical cognitive science. In fact, levels of representation are maintained, they are created according to classical models of reading and they are based on psycholinguistic evidence. Some architectural characteristics of the model are presented below, followed by a parallel between those characteristics and psycholinguistics.
Figure 5.1 The dual route cascaded model of reading aloud by Coltheart et al. (2001)
The following are some of the architectural characteristics of the model (see Figure 5.1 for a graphic representation): Levels are linked without a threshold. This means that activation spreads always through levels, and that it is graded (see above for furher discussion of the term cascaded). Further, information flows bidirectionally. The representation of words at the lexical level is local rather than distributed. Sublexical processing takes place at the phoneme level. This means that a single letter, or a short string of graphemes, such as ph, is associated with a single phoneme. The model works as follows: the first two units represent the reading visual process. The visual level activates or inhibits individual letters in the letter unit level. For instance, the presentation of “p” would activate all letters that share certain
traits, for instance, a straight vertical line, and a quasi-circular appendix, such as “p”, “b”, and “d”. However, after a sufficient number of cycles, the system recognizes more relevant features, such as the position of the quasi-circular appendix, and is able to give a larger amount of activation to “p”. At this point, the letter units send activation to the lower routes. The aim of the model is to combine letters and understand if the combinations correspond to existing words in the lexicon. This process takes place in parallel routes. What is crucial for our discussion is the existence of the two routes: the lexical route (also called a direct route) and the sublexical route (also called indirect). When the model decides whether a string of letters is a word or not, the final destination is the phoneme level, where words are pronounced. The lexical route is composed of an orthographic input lexicon and a phonological output lexicon. The sub-lexical route, on the other hand, consists of a level that converts graphemes into phonemes. If the system is presented with the visual input “meat”, it will convert the stimulus into a string of letter units which activate or inhibit a number of entries in the orthographic and phonological output lexicon, and then activate phonemes in the phoneme system. At the same time, it will also activate a rule system that assigns a grapheme to a phoneme through specific rules. As a consequence of this parallel process, the string of phonemes /mi:t/ in the phoneme system receives a strong amount of activation and the system is then able to pronounce and recognise the word. Additionally, the model takes into account frequencies: highly frequent words are activated more promptly by the system. The distinction between lexical and sublexical routes is maintained in psycholinguistic and clinical linguistics models of reading. For instance, Friedmann and Lukov (2008) describe different types of dyslexia using a dual-route model that
distinguishes between lexical conversion and grapheme-phoneme conversion (more on this in section 4.3.3). Ramus et al. (2010) propose a model for the acquisition of phonology in which orthographic and lexical representations are separated and influence each other (more on this in chapters 1 and 2). The ability to read both regular and irregular words is considered one of the reflections of the existence of a dual-mechanism: on the one hand, the phonological mechanism associates graphemes to phonemes and is useful for regular words, and on the other hand, the whole-word mechanism associates strings of sounds to strings of graphemes and is useful for irregular words. Some authors, however, argue “against” this claim and against the existence of a dual-route system (Seidenberg, 2005). Seidenberg (2005) notices that the distinction between regulars and irregulars is trivial and that dual route theories fail to account for a widely present property of English words, which he calls quasi-regularity. According to the author, assigning two different processes for regulars and irregulars creates a paradox. Non-regular verbs are most of the time obtained with highly predictable algorithms. The spelling of irregular words is not completely arbitrary, but it shares partial regularities with regulars. The author points out that when learning to read, regular and irregular words play a role on each other. When learning to read [pint] or [pants], children are likely to generalise, for instance, that the symbol [p] is pronounced /p/. His idea is then that the system is not rule governed at all. The system is rather quasi-regular (Seidenberg, 2005, page 238). There are different degrees of consistency in the mapping from spelling to sound. These range from rule-like (e.g., initial b is always pronounced /b/) to more complex contingencies.
morphologically complex words: A baker bakes and a thinker thinks, but there’s no corn in corner and a slipper is a kind of footwear, not a person who slips. Similarly, the past tense seems to be rule governed (step–stepped) but there are many partially overlapping exceptions (e.g., sleep–slept, creep–crept, keep–kept). More research is needed on quasi-regular processes, although the linguistic evidence on these suggests that they are treated as rule-like processes (i.e. certain irregular forms are not really irregular for our linguistic system (Prasada & Pinker, 1993)). In fact, neuroimaging studies seem to provide evidence for the existence of two routes as well (Jobard, Crivello & Tzourio-Mazoyer, 2003). In order to assess the existence of two routes in the brain, the authors compared the brain activation generated by the processing of words (tapping on both routes) and nonwords (tapping only in the grapheme-phoneme one), analysing data from 35 studies conducted with functional Magnetic Resonance Imaging (fMRI). The analyses led to the conclusion that different brain circuits are recruited by the two routes. Crucially for our study, the authors show that the processing of nonwords relies consistently on circuits that are supposed to be the locus of sublexical processes. From a linguistic point of view, this is the same level in which the word beginning saliency principle and morpheme stripping operate, according to the results reported in Chapters 3 and 4 in this thesis. Thus, models of reading and neurolinguistic evidence about these, combined with our findings reported in chapters 3 and 4, suggest that: 1) the word beginning saliency principle and morpheme stripping may have effects on reading, 2) these effects may be important when trying to assess reading and language difficulties.
In sum, reading is a complex process: Phonological representations are of central importance in reading, and both lexical and sublexical levels are crucial. Text decoding takes place through two parallel routes: on the one hand, single graphemes are converted into single phonemes; on the other hand, chunks of graphemes, representing entire words, are directly decoded into the phonemic representation of the word. The two routes are competing and, as a consequence, ensure quick and accurate reading performance. In reading regular and pseudo- words, the contribution of grapheme-phoneme conversion is larger; in reading irregular words, the contribution of lexical conversion is larger.
5. 3 When reading is impaired: Dyslexia 5. 3. 1 Theories for dyslexia Dyslexia is defined as a brain based type of learning disability that specifically impairs a person’s ability to read (Williams & Lind, 2013); it is identified if a child has poor literacy skills despite adequate intelligence and opportunity to learn (Bishop & Snowling, 2004, p.858). Dyslexia shows patterns of inheritance among families, suggesting a genetic base of the disorder. Although the disorder varies from person to person, common characteristics among people with dyslexia include difficulties with spelling, phonological processing and/or rapid visual-verbal responding (National Institute of Neurological Disorders and Stroke, 2012, no page). Several authors also report visual deficits in a high proportion of people with dyslexia (Stein, Riddell & Flower, 1988). Ramus et al. (2003) reviewed the three most common explanations for the disorder. The three theories are known as the phonological theory, the magnocellular theory and the cerebellar theory.
The phonological theory assumes that reading difficulties are generated by a cognitive deficit in phonological awareness (Szenkovits & Ramus, 2005), which leads to poor grapheme-phoneme correspondence. From a neurological point of view, it is usually assumed that the origin of the disorder is a congenital dysfunction of the left hemisphere. Anatomical (Galaburda, Menard & Rosen, 1994) and functional (Paulesu et al, 2001) studies confirm this assumption. The cerebellar theory (Nicolson, Fawcett & Dean, 2001) focuses on the deficit people with dyslexia have in automatic sound processing and in certain motor tasks. The cerebellum plays a role in speech articulation and is involved in automatic tasks, such as driving and reading. The authors report a brain imaging study (Nicolson, Fawcett & Dean, 2001) that shows that dyslexic participants can manifest metabolic anomalies in the cerebellum. The magnocellular theory provides a sophisticated account of dyslexia that does not single out either phonological, visual, or motor deficits (Stein & Walsh, 1997, p.147). The theory postulates that the magnocellular pathway is disrupted, and that the disruption leads to deficiencies in visual processing. Anatomical studies show abnormalities in the magnocellular layers of lateral geniculate nucleus (Galaburda, Menard & Rosen, 1994). Since the cerebellum receives inputs from many magnocellular systems, the magnocellular theory could easily account for cerebellar anomalies. As a few authors point out (Ramus et al., 2003, Vidyasagar & Pammer, 2003), the theory does not exclude a phonological deficit, but it stresses the importance of a visual account in at least a certain proportion of people with dyslexia. Further, even if the auditory system does not have a magnocellular pathway, there is a subsystem responsible for analysing acoustic transients that might
mirror this, and that could be disrupted in people with dyslexia (Stein and Walsh, 1997). Ramus et al. (2003, p. 844) identify the limits of the three accounts: The phonological theory suffers from its inability to explain the sensory and motor disorders that occur in a significant proportion of dyslexics, while the magnocellular theory suffers mainly from its inability to explain the absence of sensory and motor disorders in a significant proportion of dyslexics. The cerebellar theory presents both types of problems. There is a rich body of evidence showing that dyslexia is a very heterogeneous disorder (Stein & Walsh, 1997, Ramus et al., 2003, Pennington, 2006), and any attempt to explain all cases appealing only to one of the three theories is likely to fail. Thus, when facing a study of dyslexia from a specific perspective, one should expect to be able to explain only a certain percentage of cases. Ramus et al. (2003) tried to assess the three theories administering psychometric, phonological, auditory, visual and cerebellar tests to a sample of participants with dyslexia. Results suggest that the phonological deficit is the only one which can appear in the absence of any other sensory or motor disorder, and is sufficient to cause literacy impairment. 5. 3. 2 Dyslexia and phonology Children and adults with dyslexia typically have difficulties that primarily affect the phonological domain (Snowling, 2005, Ramus et al., 2003). Furthermore, children with dyslexia often present with phonological limitations that seem to be related to deficits in short term memory (Snowling, 2005) (see also the verbal working memory hypothesis of Vender, 2009).
Most individuals with dyslexia
present disruption in tasks that involve phonological processing, for instance: rapid word naming, nonword repetition and memorization of lists (Moura, Simoes & Pereira, 2014).
Reading difficulties may depend on a deficit in phonological
representations (Ramus et al., 2003). This means that reading difficulties may not be a dissociated problem in children with dyslexia, but they could be related to deficits in linguistic computation. These problems affect, to a certain degree, spoken language too (Dehaene, 2009, Marshall & van der Lely, 2009). Basing her analysis on the dual route model of reading presented in the previous section, Friedmann and her colleagues describe different types of dyslexia (Friedmann & Lukov, 2008, Friedmann & Gvion, 2001, Zoccolotti & Friedmann, 2010): when the lexical route is impaired, children present with surface dyslexia: they can properly read nonwords and regular words, but have problems in reading irregular words. Conversely, when the grapheme-phoneme route is impaired, children present with phonological dyslexia. In this case, children have problems reading nonwords and new words (thus, learning to read is particularly challenging). Several other forms are identified, depending on the exact deficit reported in the module, and also depending on other modules involved (such as the visual analyzer or the semantic system). Language production tasks tend to be easier than receptive language tasks for children with reading difficulties (Stojanovik & Riddell, 2008). In their study, Stojanovik and Riddell (2008) presented a sample of 17 children with specific reading difficulty with standardized behavioural tests of receptive and expressive skills, and they found that children with dissociated reading problems perform significantly better in expressive tasks than receptive tasks, with results comparable to controls in production. This phenomenon has a neuropsychological explanation.
People with reading difficulties associated with other disorders, such as Attentional Deficit Disorder (ADD) and motor deficits have been shown to manifest a certain disruption in the prefrontal cortex and in the cerebellum. On the contrary, children with dissociated reading problems have been shown to manifest disrupted activation in the angular gyrus, an area that is supposed to be related to receptive language skills only (Naeser & Hayward, 1978). Productive problems found in children with dyslexia, then, may be a consequence of perceptual problems, considering that input and output sublexical representations shape one another (Ramus et al., 2010). In sum, dyslexia is a heterogeneous disorder, which in most cases is associated with a phonological impairment that affects reading performance and learning. Running in families, the disorder is assumed to be of genetic origin, although the precise genetic base of dyslexia is far from understood. Even if dyslexia manifests differently as a consequence of different orthographies, the disorder is observed cross-linguistically.
5. 4 A frequently co-morbid disorder: Specific Language Impairment (SLI) 5. 4. 1 Introduction SLI is a disorder diagnosed when oral language lags behind other areas of development for no apparent reason (Bishop & Snowling, 2004, p.858). Many studies report that SLI affects a relatively large percentage of the child population, around 5%-10%, (McArthur et al., 2000). Furthermore, McArthur et al. (2000) showed that more than 50% of children with dyslexia also meet the criteria for SLI (see also Marshall, Harcourt-Brown, Ramus & Van der Lely, 2009 and McArthur & Hogben, 2001). Bishop and Snowling (2004) claim that diagnostic criteria for SLI
are vague. This is because children with SLI display significant difficulties with one or more linguistic domains, such as phonology, syntax, semantics and pragmatics. In fact, the defining criterion states that a child has SLI if their difficulties cannot be explained by deficits in other aspects of cognition that are linked to language acquisition, such as intelligence, hearing, oral-motor skills and language exposure (Dollaghan, 2008), but does not specifically define any subfield of linguistics as crucial for the discrimination. 5. 4. 2 SLI and morphology Friedmann & Novogrodsky (2006) present a detailed discussion of variability in SLI. The authors distinguish between phonological-SLI, syntactic-SLI, semanticSLI and pragmatic-SLI. In their study, the authors tested children with SLI using tapping tasks in dissociated form on syntax, semantics, phonology or pragmatics. They observed that children’s difficulties can be dissociated to one domain, leaving the other domains intact. However, this is not frequent, and usually children with SLI present impairment in at least two domains. Morphological and syntactic problems received particular attention (for more information on syntactic processing in SLI see Fonteneau & van der Lely, 2008, Novogrodsky & Friedmann, 2010, Contemori & Garraffa, 2010, Adani et al., 2013, Friedmann, Rizzi & Belletti, 2009 and for SLI in different modality see Mason et al, 2010). In English, there is a large body of research focusing on the process of inflection, for instance, Joanisse (2004), Joanisse, Manis, Keating & Seidenberg (2000), Marshall and van der Lely (2007), Marshall and van der Lely (2012), Miller, Leonard and Finneran (2008), Oetting and Rice (1993), Poll, Betz and Miller
(2010), Rice and Wexler (1996), Robertson, Joanisse, Desroches and Terry (2012), van der Lely and Ullman (2001), Van der Lely and Ullman (1996), van der Lely (2005). In fact, tense inflection was identified as the main marker for (grammatical) SLI in English (Rice & Wexler, 2001). English children with SLI have problems with inflectional morphemes, particularly with past-tense morphemes and presenttense third person (Matchmann, Wulfeck &Weismer, 1999).
Oetting and Rice
(1993) analysed the spontaneous production of number markings, and showed that this domain is relatively intact in children with SLI. Miller, Leonard and Finneran (2008) showed that children with SLI have difficulties in detecting bound morpheme omissions even in later stages of their development, and the gap with TDs is still large in adolescents. Van der Lely and Ullman (2001) investigated the formation of past-tense in children with SLI using real and novel verbs. The authors showed that TDs manifest an advantage for regular real and novel verbs inflection, while children with SLI do not. Lexical frequency affected children with SLI, but not controls, with regard to regular verbs (Van der Lely and Ullman, 2001, p.178, partially adapted). For irregular verbs frequency was relevant for both groups. Overall, children with SLI produced fewer regular verbs. These results are consistent with a dual-route model of tense generation (as in Pinker & Ullman, 2002a) and suggest that rule generation is impaired in children with SLI. A similar result was obtained in a previous study (Van der Lely & Ullman, 1996). 5. 4. 3 SLI as a morpho-phonological disorder Joanisse (2004) argues that a phonological explanation of SLI is better at describing these patterns than one based on morphosyntax. Particularly, the author
points to the fact that an explanation based on (morpho)syntax fails to explain certain behavioural patterns reported in children with SLI, for instance difficulties with irregular verbs and difficulties with phonological tasks. If the difficulties of children with SLI were limited to syntactic rules application, they should perform poorly only on regular verbs inflection, since inflected forms of irregular verbs are supposed to be stored in memory (in rule based accounts, such as Pinker & Ullman, 2002a). Data show that Joanisse (2004) is right: Marshall and van der Lely (2007) investigated the interaction between phonological complexity and the derivation of inflected forms. Testing children with grammatical-SLI with a purposely created elicitation task, they showed that children with SLI find it harder to generate inflected forms when these forms generate a phonological cluster in word final position. The control group did not show such an effect. In a similar study (Marshall & van der Lely, 2012) they failed to find the contrast between SLI children and controls. Instead, they observed the phonological complexity effect on morphology in both groups. Robertson et al. (2012) investigated the elicitation of the past tense and phonological performance in children with SLI, children with dyslexia, and controls. They showed that while both clinical groups performed poorly in the phonological task, only children with SLI performed poorly in the morphological one. In a different paper, van der Lely (2005) analysed the disorder from a more heterogeneous perspective, showing that phonological, morphological and syntactic problems coexist in children with SLI. At the same time, the author showed that non-linguistic factors cannot, in isolation, generate the behavioural patterns observed in children with SLI.
5. 4. 4 Phonology as a bootstrap for the morphological deficit Many children with SLI present with phonological and morphological problems simultaneously. The relation between phonology and syntax in SLI has been discussed by a number of authors (Leonard et al., 1993, Chiat, 2001, Marshall & van der Lely, 2007, Nithart et al, 2009). From this line of research, much attention has been given to the proposal according to which SLI children face problems in interpreting phonologically weak elements, for instance elements without stress (Leonard et al., 1993). Marshall and van der Lely (2007) show that phonological complexity plays a role in the severity of the disorder. Particularly, they showed that children with SLI find it harder to produce inflected forms when these forms generate phonological clusters (in the word final position). Chiat (2001) proposes a model in which the phonological impairment bootstraps the syntactic one. The proposal of Chiat (2001) is based on evidence that phonological processing is crucial in the acquisition of syntax and morphosyntax. The acquisition of English inflection morphemes is a good example of this interaction: firstly, English inflectional morphemes are phonologically weak elements, and are prosodically non salient. The second thing to consider is that, as with all function words, they do not carry meaning “per se”. Their meaning corresponds to their grammatical function. For instance, the morpheme -s indicates present tense and third person in that it modifies the verb and is used when the verb refers to a third person (in syntactic terms) doing or experiencing something in the present. The morpheme cannot bring this meaning if used in isolation. The questions to ask are what cues are available for the child in the speech they hear for the acquisition of this morpheme to occur. A related question is what information can the child rely on to map this function onto its
phonological form. Differently to what happens for the acquisition of content words, gaze interpretation and semantic bootstrapping are of no help in this situation, because the function word is empty in isolation. What the child needs to do is notice the phonological contrast that occurs between different forms of the verb. For instance, the child may notice the phonological difference between play - plays played. In most languages, verb paradigms produce what are defined as “morphosyntactic minimal pairs” (Strange & Law, 2008), as it was thoroughly discussed in chapter 4 in this dissertation. Once this phonological contrast is noticed, the child will be able to map the different phonological forms with different syntactic scenarios (i.e. different patterns of who is doing what). The acquisition and the meaning mapping of these morphosyntactic particles are thus dependent on phonology. It is not surprising, then, that most children with SLI have problems with phonological tasks such as nonword repetition and morphosyntactic tasks involving elicitation of inflectional morphemes. During language acquisition, the two processes are intimately related. Notice that this proposal can also account for contrasts, such as the one observed in Jakubowicz, Nash, Rigaut & Gerard (1998): given two forms with the same phonological complexity, it will still be harder for children with SLI to process the form that required more complex phonological mapping during language acquisition8. 5. 4. 5 Clinical markers for SLI Poll, Betz and Miller (2010) assessed the usefulness of three different tasks in the assessment of SLI: nonword repetition, sentence repetition and grammaticality 8
Jakubowicz et al. (1998) showed that for French children with SLI is easier to produce the determiner “le” than the object clitic with the same phonological form “le”.
judgments on sentences missing inflectional morphemes. Results showed that each of these tests was a good predictor of the presence of SLI, and that the most reliable way of making a prediction was using a combination of the three. In a largely resonant study, Rice and Wexler (1996) investigated the effectiveness of four different potential grammatical markers for the identification of SLI: the auxiliaries DO and BE, and the bound morphemes -s and -ed. The authors showed that each of these elements can be useful in assessing SLI, and suggest that the grammatical function of tense could be used as a clinical marker. Conti-Ramsden and Hesketh (2003) assessed four different markers using sensitivity, specificity and accuracy analyses. They compared the performance of SLI children and younger controls using four tests. The four tests used were: a digit recall task, the Children’s test of Nonword Repetition (CNRep), a past tense elicitation task (similar to TEGI, Rice & Wexler, 1996) and a plural elicitation task. The results suggest that CNRep is the most reliable task in identifying risk factors, i.e. in identifying whether the performance is low because of normal variation observed in the population, or if the performance is low because a disorder is present. In other words, children who scored low on CNRep were the most likely to have SLI. More recently, Chiat and Roy (2007) showed that nonword repetition can also be used to identify SLI children in pre-school years. With a purposely developed test, the Preschool Repetition test (PS Rep Test), they showed that children as young as 2;6 can be differentiated from TDs using repetition performance.
5.5 Analyses of CNRep and TEGI 5. 5. 1 CNRep and word position effects Gallon and Marshall (2009) discussed the relation between nonword repetition and dyslexia, and showed that reading performance can be predicted by performance in nonword repetition. In previous research, Gathercole (1995) investigated the relation between lexical knowledge, phonology and nonword repetition. The authors claim that nonword repetition tasks are extremely important in the assessment of reading disorders, particularly in the assessment of dyslexia. Several authors also reported the effectiveness of nonword repetition tasks in the assessment of SLI (Norbury, Bishop & Briscoe, 2001 with the CNRep), and poor performance in nonword repetition tasks is regarded as a clinical marker of SLI (Conti-Ramsden & Hesketh, 2003). Given the relevance of CNRep (Gathercole & Baddeley, 1996) on the assessment of dyslexia and SLI, we investigated whether our findings on word position effects, reported in Chapter 3, would have consequences on the interpretation of the test. The CNRep assesses short-term memory (which often correlates with both language and reading abilities) and phonological processing, and is often used as part of a battery in the assessment of developmental disorders (Gallon & Marshall, 2009, Gathercole & Baddeley, 1996.). In this test, there are 4 types of nonwords, divided according to number of syllables: 10 two syllable nonwords, 10 three syllable nonwords, 10 four syllable nonwords and 10 five syllable nonwords. Normative data suggest that longer nonwords are repeated less accurately by all age groups (ibid.). However, this claim does not take into account word position effects generated by clusters. We showed in chapter 2 that non initial
clusters are processed less accurately than word initial clusters, across age and across languages. Inspection of the distribution of non-initial clusters in the CNRep task shows that they are not balanced across syllable length. Non-initial clusters are all positioned in the four and five syllable words, and never in the two and three syllable words. A Fischer’s exact-test (chi-square could not be used because the count of cells is smaller than 5 in 4 cases) shows that the distribution of clusters in non-initial position is significantly unbalanced: χ (3) = 11.3, p = .004 two tailed (see Table 5.1 below for detail). This suggests that the normative data obtained for the CNRep assessment may be influenced by the unbalanced distribution of non-initial clusters, not only by the length of the word. Note: as in chapter 3, only clusters composed of plosive + liquid are considered, since they cannot be shared by two syllables and as such they are always processed as units (Roach, 2000).
Table 5.1 Distribution of non-initial clusters in the Children’s Test of Nonword Repetition, (CNRep), (Gathercole & Baddeley, 1996).
Non initial cluster yes no yes no yes no yes no
Number of syllables
0 10 0 10 4 6 5 5
2 2 3 3 4 4 5 5
The clinical implication of this imbalance could primarily manifest through false positives in the assessment. It is not just the number of errors that will indicate the presence of dyslexia and/or SLI, but also the distribution of these errors. For instance, if a child made a relatively high number of errors, but their distribution resembles what our findings predict, it may be argued that the pattern is due to general word position effects, and not to the presence of a disorder. Given the unbalanced distribution of non-initial clusters, there is no way to assess whether an error with a cluster in the CNRep is due to the cluster itself, to word length, or to the fact that the cluster is non-initial (i.e. to word position effects). This is a potential issue because variations in performance related to word position effects may have to be considered non-clinical, as shown in Chapter 3. 5. 5. 2 TEGI and word position effects In our fourth experiment (see Chapter 3), lexical and morphosyntactic minimal pairs were used to investigate word final position effects in perception, and their interaction with morphological information. The stimuli we used were carefully chosen so that the phonological context was identical for final phonemes that carried and those that did not carry morphological information. Difficulties with production and perception of inflectional morphemes are a well established marker for SLI in English (Rice & Wexler, 1996) and, therefore, inflectional morphemes have a central importance in assessments. One test focusing on the elicitation of inflection morphemes is the Test of Early Grammatical Impairment (TEGI) (Rice & Wexler, 2001). In this test children are asked to complete sentences using either words containing morphological information or words without morphological information.
For instance, children may be primed to produce the word “bus”, a word ending with the phoneme /s/ in a context in which the phoneme does not carry any morphological information, or to produce “asks”, a word ending with the phoneme /s/ in a context in which the phoneme carries grammatical-morphological information. The test is composed of three sections: one eliciting a phonological probe, such as “bus”, one eliciting a third person singular probe, such as “asks “, one eliciting a past tense form, such as “asked”. The directions the test administrator provides in the three conditions are different and examples are provided below. For phonological probes, the elicitation resembles a naming task. In this condition the therapist says (directions reported from Rice & Wexler, 2001): I am going to show you some pictures and ask you to name some things. Then a picture as below (Figure 5.2) is presented:
Figure 5.2 Pictures of this type are used in the Rice & Wexler (2001) assessment to elicit words ending in the phoneme /s/.
For the third person singular probe, the elicitation is slightly more complex. In this condition the therapist introduces the task and guides the child, using pictures as in Figure 5.3: I am going to show you some pictures and ask you to tell me what each person does. Let’s try one. Then, a picture as seen below is presented: here is a wall painter. Tell me what a wall painter does. The target is PAINTS
Figure 5.3 Pictures of this type are used in the Rice & Wexler (2001) assessment to elicitate third person verbs ending in the phoneme /s/.
For the past tense probe, two pictures are used: one in which the process is ongoing and one in which the process has finished. Directions from the therapist include non linguistic interaction such as pointing: I have two pictures. I will describe the first one and you tell me about the second one. Let’s try one (pointing to raking picture), here the boy is raking (point to raked picture), now he is done. Tell me what he did. The results from Experiments 4, 5 and 6 (see chapter 4) suggest that the processing of affixes is of central importance in influencing word position effects, with the morphophonological context of word final positions playing a crucial role. The TEGI test, however, does not take into account the morphophonological context in which the final phoneme appears. In the phonological probe condition, all elicited phonemes follow a vowel, but this is not the case in the morphological conditions in which sometimes the relevant morpheme follows a consonant. What if children have problems with -s following a consonant? If children had this type of difficulty, i.e. a
purely phonological problem, for instance, with clusters, the test would not be able to detect it. Children showing difficulties with the production of word final inflection morphemes with this test may receive a diagnosis of SLI. However, since the test does not take into account the morphophonological context in which the final phonemes are elicited, children failing to perform within the norms in the test may actually just have difficulties with phonological clusters: in fact, in this test, phonological clusters only appear in the morphological condition, creating a bias that is very difficult to control. . Table 5.2 Distribution of word final clusters in the Test of Early Grammatical Impairment, (TEGI)(Rice & Wexler, 2001).
Word final cluster
Third person probe
Third person probe
Past tense probe
Past tense probe
The distribution of phonological clusters in word final position (Table 5.2) was analysed with the Fischer’s exact-test. The test is significant, χ (2) = 11.72, p = .002. This means that word final clusters are significantly unbalanced, i.e. they appear
morphophonological distribution may have serious impact in the assessment, since the test is not designed to tease apart whether problems in the morphological
conditions are due to inflected form generation (i.e. rule application) or to phonological complexity (i.e. cluster articulation). A final note regards tests of speech perception. One may notice that our research focused on speech perception, while our clinical implications concern production. This is mainly due to a substantial lack of purely perceptual tasks in clinical assessments, even if perception can be impaired (and also selectively impaired in people with disorders, Ramus, 2010). As Vance, Rose and Coleman (2009, pag 708) point out, Few materials are available to assess speech perceptual skills in young children without hearing impairments. However, children with a range of developmental conditions are at risk of speech discrimination deficits. Tasks that reliably assess speech perception skills are thus necessary for research and clinical practice. An exception in this sense is the task from Newton, Chiat and Hald (2008). The task, based on minimal pairs discrimination, represents a substantial innovation if compared to classic tasks used to assess speech perception, such as picture selection tasks with minimal pairs. The task is based on the preferential looking paradigm. Two items generating a phonological minimal pair in each trial were shown on the screen, but only one of the items is named through the speakers. Gaze was measured for two seconds. Participants were English speaking children aged 2 to 7. Results showed that if participants were looking at the wrong target with their first gaze, they were quicker at switching target than if they were looking at the right target in the first place. A significant overall effect of duration was also observed, with targets being observed for longer than distractors. These effects did not however vary with age. Further, the test showed a phonological effect, with voice and manner
contrasts in the minimal pairs leading to slower times in gaze changes compared to place of articulation contrasts. The task, thus, is informative not only in terms of discrimination ability, but also in terms of relative attention to differences that are discriminated. Further research on clinical samples with similar paradigms may offer new relevant insights into the perceptual phonological deficits of children with dyslexia and those with SLI.
5. 6 Clinical data analysis In order to investigate the potential impact of the unbalanced distribution reported in the CNRep (see section 5.5.1), we analysed the assessment data based on clinical reports of 22 children who were all assessed for language difficulties in a previous set of research conducted in our department, whose information was available under the agreements of the Linguistic Assessment Clinic programme. All of the children were either currently undergoing speech and language therapy, or had had speech and language therapy at some point in their childhood. Data from 22 children were used. There were 12 boys and 10 girls, aged between 5;08 and 14;06. Being tested by different clinicians and also due to their slightly different presentation in terms of speech and language profiles, children were assessed with various tests. All children included in our sample, however, were assessed with CNRep and at least with one other standardised language assessment. All children performed at least 1SD below the mean in at least one of the assessments (excluding the CNRep), and all of them had normal non-verbal abilities. As such, all children in our sample can be defined as “language impaired”.
The sample can be divided in 3 subgroups: when children presented with a language impairment due to causes external to language, such as hearing loss or stroke, they were classified as non-SLI. When they presented with a linguistic deficit that could not be explained by any other factor, they were classified as SLI. When, they presented with a linguistic deficit and with reading and/or phonological difficulties, they were classified as dyslexia+SLI. More specifically, the criteria for assigning participants to one of the two groups were the following: when a child was reported as having phonological and/or reading difficulties and performed below their expected performance in at least one of the tasks (other than CNRep) they were assessed with, they were assigned to dyslexia+SLI. If the child performed poorly on a linguistic task but reading/phonological performance was not flagged as problematic, the child was classified as SLI. Information was sometimes limited and children were assessed by different clinicians. As such, the classification has to be considered non-conclusive. Further research, assessing each child with the same set of tests, could better disentangle word position effects in different disorders, in the sense that the starting classification of the various groups may be more reliable than the one adopted in this study. Detailed information about each participant is presented in appendix 4. Descriptive statistics for the two clinical groups together are presented in Table 5.3. Descriptive statistics for the three groups separately about performance in 5 syllable nonwords are presented in.
Table 5.3 Descriptive Statistics for children with a clinical condition. Disorder N Mean Age Age SD Mean CNRep SD CNRep SLI+dyslexia 9 9;00 2;09 22.44 7.33 SLI 9 10;07 2;06 25.00 12.09
Note: The Mean CNRep refers to the mean number of correct items. Table 5.4 Descriptive Statistics (based on raw number of errors) group 1. SLI + dyslexia. N Minimum Maximum Mean SD
Statistic Statistic 5 3.00 5
Table 5.5 Descriptive Statistics (based on raw number of errors) group 2. SLI. N Minimum Maximum Mean SD Statistic
Table 5.6 Descriptive Statistics (based on raw number of errors) controls. N
Statistic Statistic 5 1.66 4
SD SE .44
We claimed above that non-initial clusters are not systematically distributed across item-lengths in CNRep, and that this creates risks considering that also nonclinical groups tend to find it difficult to process consonant clusters in non-initial position. To test whether a word position effect was present (and overlooked) in the clinical assessment, we analysed the distribution of errors in the CNRep reports of
the 22 children with language impairment (Table 5.3), and also compared their performance to that of 12 age-matched controls (age was matched at group level, mean age of the control group is 9;9, SD, 1.64). As discussed above, non-initial clusters are not present at all in the CNRep in words composed of 2 and 3 syllables. In 4 syllable words their distribution is unbalanced, which makes 4 syllable words an unpractical ground to analyse word position effects: in 4 syllable words, 4 words out of 10 contain non initial clusters, and 6 words out of 10 do not. Thus, there is a statistical bias, and children are ante-eventum more likely to make more errors in the condition in which there are no clusters. Five syllable words, however, offer a clear ground for analysis, since half of the tokens contain a non-initial cluster, and half do not. Two conditions (presence VS absence of non-initial clusters when there was an error) were compared with a paired samples t-test (a chi-square could not be used because the expected count is smaller than 5 in all cells). For a detailed description of the stimuli see Table 5.7 9. Table 5.7 Five syllable nonwords used for the clinical data analysis. Condition 5 syllable nonwords Nonwords containing a non-initial cluster Sepretennial, detratapillic, confrantually, underbrantuand, versatrationist Nonwords not containing a non-initial cluster Defermication, reutterpation, altupatory, pristoractional, voltularity
The t-test is significant, t (21) = -3.8, p = .001. The children with Language Impairment in our sample, thus, made a significantly larger number of errors in 5 syllable nonwords when the nonwords contained non-initial clusters. This pattern is 9
Pristoractional stands in the group without clusters for syllabification reasons, since most speakers consider the phonemes creating the cluster as belonging to two different syllables, as in the real word “reaction”, syllabified re – ac – tion (hyphenation24.com)
predicted by non-clinical data; for instance, a similar effect is reported in Experiment 2 in this PhD. Performance of the clinical samples was compared to the performance of 12 agematched controls. Raw data for controls were provided by Lisa Archibald and Susan Gathercole (2006), who personally sent us their file. In order to check whether children with dyslexia+SLI, and children with SLI and TDs performed differently, a repeated measures ANOVA was performed. The repeated measures ANOVA was performed having group (SLI vs dyslexia+SLI vs TDs) as between-subjects factor, and cluster (whether there was a cluster or not when the child made an error) as within-subjects factor. The ANOVA revealed a significant effect of cluster, F (1) = 8.71, p = .006 and a significant effect of group, F (2) = 3.98, p = .030. The interaction was not significant, F (2) = 0.34, p > .05. An ANOVA analysing the overall performance in the test, rather than 5 syllable words only, shows a significant groups effect: F (2) = 4.41, p = .022. When using the overall test performance, TD children significantly outperformed both groups: t (19) = 9.56, p = .009 when compared to dyslexia+SLI children; and t (19) = 7.00, p = .049 when compared to children with SLI only. The results are represented visually in Figure 5.4. Values referring to the condition in which the cluster is present are reported in blue, values referring to the condition in which the cluster is absent are reported in red. Bars are divided according to group (SLI+dyslexia, SLI, controls).
Number of errors
2 1.5 1 0.5 0
Figure 5.4 Accuracy (number of errors) in 5 syllable words.
An analysis of the correlation between age and performance in the test was also performed. The correlation is significant when using a one-tailed hypothesis, r (18) = .39, p = .054. Given this finding, an independent samples t-test was performed comparing the ages of the two groups to assess whether age could act as a confound in the two groups (for instance, if children with dyslexia+SLI are younger than children with SLI, age may be the cause of their lower performance). The analysis suggests that this is not the case, t (16) = .006, p > .05. Further post-hoc analyses were carried out. The results show that, when comparing all groups, the only significant difference is between controls and children with dyslexia+SLI (Table 5.8).
Groups 1 2 3
Table 5.8 Post-hoc analysis between groups. 95% Confidence Interval Mean Difference Lower Upper (I-J) SE Sig. Bound Bound 1.38 .61 .099 -.18 2.96
Key: group 1 stands for dyslexia+SLI, group 2 stands for SLI, group 3 stands for aged-matched controls.
The effect of cluster was investigated within each group, using paired samples t-tests, presented in Table 5.9. The analysis shows that the only significant difference between nonwords with and without noninitial clusters is observed in children with SLI, t (8) = -3.00, p = .017 (Table 5.9). Overall, the results obtained with the post-hoc tests suggest that the pattern observed in the ANOVA is driven by a very limited number of contrasts. It should be noted that the sample used is relatively small. Further research with a larger sample may significantly improve the analysis conducted in this part of the project. Table 5.9 Post-hoc analysis within groups.
Group Dys+SLI SLI Controls
Paired Differences 95% Confidence Interval of the Difference SE SD Mean Lower Upper
In summary, a main effect of cluster was observed, but the Post-hoc analysis shows that the effect is driven by the SLI group only. A main effect of group was also observed: the post-hoc analysis shows that the effect of group is driven by the comparison between children with dyslexia+SLI and controls: children with dyslexia+SLI made more errors than typically developing children. These results suggest that children with dyslexia+SLI and controls can be distinguished using phonological measures, as already suggested in Marshall and van der Lely (2009). The analysis of the performance presented above, combined with the analysis of the distribution presented in section 5.5.1 may raise some questions about the nature of the phenomena measured with CNRep. If children made more errors when nonwords contained a non-initial cluster in 5 syllable words, the length effect observed may be due to the unbalanced distribution of non-initial clusters and not (only) by length. This is because non-initial clusters are all positioned in 4- and 5syllable nonwords in this test, as shown by the analysis of distribution reported in the chi-square in section 5.5.1. The possible contribution of clusters in the patterns of performance observed in CNRep is an issue already discussed and reported by the authors of the test themselves in a recent reanalysis of the test (Archibald & Gathercole, 2006). The authors reported and acknowledged a clusters effect, but did not make an association between the pattern of performance and the distribution of clusters. Our results suggest that the difficulties that children have with long nonwords in this test may be influenced by an effect of word position that is found in typical populations and does not represent a clinical factor. The validity of the assessment as an important contributor in the battery is not questioned, especially because the test
was evaluated on a large number of participants and revealed to be a powerful diagnostic tool (Conti-Ramsden & Hesketh, 2003). As a further confirmation, children with dyslexia+SLI performed worse than controls also in our analysis, and when using overall performance, TDs outperformed both clinical groups. However, the unbalanced distribution of clusters, as well as the absence of a natural way of controlling for it, requires attention. The results obtained analysing our clinical data confirm the predictions of our non-clinical tasks, and without questioning the quality of the test in detecting a clinical condition, they suggest that part of the word length effects observed may be a consequence of phonological complexity rather than shortterm memory. Nonword repetition tasks rely on what is known as the phonological loop (Figure 5.5), a subsystem of working memory (Baddeley, 2003). According to Baddeley (1992, Page 556): The term working memory refers to a brain system that provides temporary storage and manipulation of the information necessary for such complex cognitive tasks such as language comprehension, learning, and reasoning. This definition has evolved from the concept of a unitary short-term memory system. Working memory has been found to require the simultaneous storage and processing of information. It can be divided into the following three subcomponents: (i) the central executive, which is assumed to be an attentional-controlling system, is important in skills such as chess playing and is particularly susceptible to the effects of Alzheimer's disease; and two slave systems, namely (ii) the visuospatial sketch pad, which manipulates visual images and (iii) the phonological loop, which stores and rehearses speech-based information and is necessary for the acquisition of both native and second-language vocabulary.
Figure 5.5 Baddeley’s (1992) graphic representation of the phonological loop.
Our analysis shows, together with analyses from previous studies, that phonological complexity may have an important effect in nonword repetition, and this effect is virtually independent from the effects related to short-term memory. For this reason, it is important to assess whether the word length effects observed in CNRep are a consequence of short-term memory effects, or if they are a consequence phonological complexity.
5. 7 Conclusion This chapter presented the relation between sublexicon and reading and the relevance of sublexical word position effects in the assessment of language disorders. Sublexical word position effects are a relatively overlooked phenomenon, but this chapter shows that they can be informative in the interpretation of performance, and can be useful in making predictions about children with and without a clinical condition. Most importantly, they can be useful in the interpretation of data obtained with clinical assessments. Since many assessments are based on the use of nonwords -tapping into sublexical representations (specifically, assessment of reading, and, more precisely, of decoding), the word position effects we explored in the previous chapters can be used to interpret the contribution of phonological complexity in
performance of subjects with and without a clinical condition, completing the claims that can be made about short-term memory when using a nonword task. Our data analysis shows that the word length effect reported in CNRep may be influenced by the fact that non-initial clusters are contained only in long words in the test. The analysis shows that, when comparing words of the same length, children of all groups are more likely to make an error if the word contains a non-initial cluster. This finding is important because it shows that the length effect observed may be a consequence of phonological complexity rather than a consequence of short-term memory. This finding does not question the reliability of CNRep as a contributor in the assessment of language disorders: Children with dyslexia+SLI performed worse than children with SLI only, and children with SLI performed worse than controls when considering the overall test performance. The analysis simply suggests attention in the interpretation of the performance on CNRep. Word length effects may in fact be a consequence of word position effects.
CHAPTER 6. DISCUSSION 6. Discussion
6. 1 Summary This final chapter is organised as follows: the first part is a critical evaluation of the stimuli and of the methods used in this project. Following the critical evaluation, , the hypotheses advanced at the beginning of the project are presented. Then, the findings of the experiments are summarised, and limitations of the studies are presented. Implications of our findings for research and for clinical practice are discussed, followed by ideas for future research. Finally, a concluding reflection on the subject is provided.
6. 2 Critical evaluation of stimuli and methods Before summarising and discussing the findings of this PhD, it is important to present a critical evaluation of the stimuli and of the methods used in this project. This discussion is fundamental for two reasons: on the one hand, it introduces a sceptical gaze to the findings reported, and it is important to have this attitude when pursuing scientific research. On the other hand, it offers the ground for possible developments of this study, in which some of the confounds we are going to mention could be taken into account. A list of possible criticalities is presented below: 1. Recording of stimuli: The words used in all the tasks were recorded in spoken form, and not obtained using software. A clear advantage of using stimuli recorded by linguistis is that they sound more natural, even when the stimuli recorded are nonwords. However, this choice also presents some problems. First of all, even when being very careful, there are differences in the length of the stimuli,
and they are extremely difficult to control for in the data analysis. Second, in experiments in which the disambiguating point is non-inital, it is crucial for the rest of the word to be identical in the various conditions. For instance, in the discrimination of the words “side” and “size”, it is fundamental that the two words are identical until the uttering of the final phoneme. As much as every attention was used to ensure that this was the case, there may have been very subtle differences that analogic linguistic analysis could not detect, and this may have had influenced the results. In the future, it would be interesting to run again these tasks using nonwords generated by computer software, such as Ivona. 2. Variability in length: the words in these experiments do actually vary in length. The analysis assumes that this variation is not influencing the result, since it is assumed that the variation is so small that it did not affect processing. One cannot, however, be absolutely sure about it. In the future, again, this problem could be solved using stimuli obtained with software rather than recorded by linguists. 3. Reaction times as post-hoc: Reaction times are used in experiment 3 as evidence for the claim that word beginnings are salient in English adults. However, reaction times calculated in Experiment 3 are a post-hoc analysis. They were not part of the hypothesis and the measure that was supposed to be used as dependent variable, accuracy, was rejected because at ceiling. These are relatively arbitrary choices. Even if both choices are motivated (the rejection of accuracy and the use of RTs), it is important to check whether the results obtained in Experiment 3 will be consistently replicated in English speakers, since the hypothesis appeared to be poorly predictive and the result found is a consequence of a post-hoc analysis. Given
the (little) evidence obtained with this experiment, caution is suggested in the generalisation of the finding. 4. Words used in the fourth experiment: the words used in Experiment 4 are problematic for a number of reasons. First of all, elements in lexical pairs are semantically unrelated, while elements in morphosyntactic minimal pairs are semantically related. Although the problem of this confound is solved when looking at Experiments 5 and 6, Experiment 4 appears quite weak in this regard when considered in isolation. Second of all, some of the words have homophones, and some homophones even belong to different word classes. For example, the word “eight” is a homophone of the word “ate”, the word “side” is an homophone of the word “sighed” (although a short lengthening of the vowel is observed when uttering it as a verb) and the word “place” can have different meanings, it can be a noun, referring to a physical place, or it can be a verb (to place) that can refer to the moving of objects but also to the finding of jobs. Word length was only considered in syllabic terms, i.e. all words were monosyllabic. The actual length of the words was not considered, and, similarly, the orthographic length of these words was not considered. Finally, final clusters are more frequently present in the morphological condition, and this is by its own a very severe confound. Most of these problems were solved with Experiments 5 and 6, but, again, they remain present when considering Experiment 4 in isolation, and as a consequence the interpretation of Experiment 4 in isolation remains very problematic.
5. Stimuli used in Experiment 5: being a selection of the stimuli in Experiment 4, stimuli in Experiment 5 could retain some of their problems. Several of them are, however, solved. The phonological problem is controlled for: only CVC words are used in the entire task. The length problem is controlled for: ERPs are time-locked to the disambiguating points, making differences in length irrelevant. The semantic problem is controlled for: the nature of MMN negativity led to a precise prediction in terms of the effect of semantics, and the result obtained clearly indicates that the difference cannot be related to semantics (for an in depth discussion see section 4 of chapter 4). One of the pairs used contains homophones: “side” vs “size”. Arguably, this problem is controlled for by the absence of vowel lengthening in the presentation of these stimuli. The absence of vowel lengthening should induce in listener the perception of these words as being nouns and not verbs. A new, unsolved problem, however, emerges in this task: in order to control for all these confounds the selection of stimuli ended up being very small, thus the number of items used in this task (eight) may be considered too small. 6. Vowel lengthening is not considered in Experiments 4, 5 and 6: vowel lengthening associated with verbs is acknowledged in the three experiments, but never measured. Further, it is not considered in detail in the discussion. Could we safely say that the different reaction times are due to the presence of morphology or to the presence of vowel lengthening associated with verbs? How intimately related are morphology and vowel lengthening? These questions deserve appropriate discussion, and for the moment are overlooked in this project. Further research may address the effects of vowel lengthening in morphological processing by using artificial stimuli that contain morphology but lack vowel lengthening.
7. Interpretation of Experiment 6: In Experiment 6 a large contrast was detected between the non-morphological condition and the phonological control condition. This result is not explained by the theory, it is not predicted by the presence of morphology (there is no morphology in both conditions) and may undermine the conclusions reached by the rest of the analysis. In fact, the reason for the difference between non-morphological condition and phonological control condition is likely to be phonological, for instance the difference could be due to the fact that the contrast in voicing feature is present only in the non-morphological condition. As a consequence, one could argue that the difference between the morphological and the non-morphological conditions is strongly influenced by the “same” effect. At the actual state of the research, it is not clear how large the impact of phonology is and how large the impact of morphology is in the generation of the difference between morphological and non-morphological conditions. 8. Clinical data analysis: in the clinical data analysis, the focus is on the distribution of non-initial clusters, while the distribution of initial clusters is completely overlooked. This is potentially a problem. In fact, if we observe also the distribution of initial clusters, we can see that words that do not contain non-initial clusters contain initial clusters only in one case (out of five). In other words, words that contain non-initial clusters contain more clusters overall, independently to their position. Given this situation, there is no way with this data to understand with certainty whether the effects we observed are a consequence of non-initial clusters or if they are the consequence of the presence of clusters overall. A possible way to
solve this problem would be to test children with the same nonwords containing noninitial clusters and with a set of new nonwords that always contain an initial cluster. Given the critical evaluation presented in this section, caution is advised in the interpretation of the result, and while reading any claim made in the experimental chapters and in the following sections, one should consider also the acknowledged limits (and the unacknowledged ones) of the experimental paradigms used.
6. 3 Main findings from the PhD The aim of the current PhD project was to investigate word position effects in sublexical representations. Word position effects have been broadly studied by researchers, and some phenomena are widely and cross-linguistically observed. Word beginnings are salient and important for lexical access, and this was demonstrated by psycholinguistic experiments as well as by connectionist models on lexical access (McClelland & Elman, 1986). Word final positions are, instead, prone to different phenomena, and their saliency depends on the presence of morphological information. Disruptive phonological phenomena, such as deletion, are blocked when morphological information is transmitted (Pater, 2006). A very limited amount of research concerning word position effects, however, has been conducted on sublexical representations, and this project addressed this problem. Further, in the last part of the dissertation, the implication of word position effects in clinical assessment were investigated on a sample of children with developmental language disorders. The first three experiments concentrated on word beginnings, with tests involving children and adults, and speakers of English and speakers of Italian. The
second three experiments involved word endings, and were carried out on English speaking adults, tested with psycholinguistic and neurolinguistic methods. The implications of our findings were then evaluated on a clinical sample of English children with developmental language disorder. The hypotheses, summarised in chapter 1, were the following: At the input sublexical level: Hypothesis 1: Word beginnings are inherently salient Hypothesis 2: Word endings are optionally salient The hypotheses were based on the analysis of the linguistic and psycholinguistic literature available on lexical word position effects, combined with an analysis of the knowledge available about phonological representations. Hypothesis 1 was confirmed on Italian for children and adults, and similar effects were observed on English adults: Italian participants (children and adults) were more accurate in detecting word initial contrasts than in detecting word medial contrasts. English subjects were slower in detecting word initial contrasts than word medial contrasts (although the analysis presents some important limitations, presented in section 6.2). For what concerns Italian, the pattern was only observed in unstressed syllables. Hypothesis 2 was confirmed using different types of stimuli and different methodologies: the discrimination of elements carrying morphological information took longer in terms of Reaction Times and generated brain related potentials (specifically MMN) with larger amplitudes than the discrimination of elements without morphology.
All experiments were based on variations of the same methodology, which consisted in the discrimination of elements differing only on one phonological unit (one phoneme). Given that the sublexicon is the focus of this dissertation, most tasks were created using nonwords. However, this was not the case for Experiments 4 and 5. An important remark, thus, needs to be made about the use of lexical items in the study of sublexical processes. In Experiments 4 and 5, the stimuli used were real words, because these are the type of elements stored in the lexicon. However, the claim made in this work is that the tasks used are still addressing sublexical processes. This is related to the type of model of phonology this research is based on (in other words, on our definition of sublexicon) and on the type of tasks that were performed with these real words. Concerning the model, the reference used is the model of phonological acquisition developed by Ramus and colleagues (2010). The model, as the authors claim, is a psycholinguistically plausible development of the classic model of Chomsky and Halle (1968) in which phonological representations are divided in underlying and surface representations. While surface representations contain fully specified description of phonemes, in a given word and in given contexts, underlying representations are more abstract concerning certain features of the phonemes. For instance, the phoneme /k/ can appear in two forms in English: with aspiration, as in car, and without aspiration, as box. At the level of underlying representations this difference is not specified, but at the level of surface representations the difference is specified (with a rule). Surface representations correspond to sublexical representations in Ramus et al. (2010). At this level, phonemes are fully specified. Our task consisted in the discrimination of fully specified phonemes. The participants were asked to compare between surface forms
of phonemes (/t/ vs /s/ and /z/ vs /d/). Thus, even if using lexical items, the contrasts performed were taking place sublexically in Ramus et al. (2010) terms. This does not mean that the lexicon could not have had an effect at all in the task, for instance in terms of the time needed to access these surface phonemes. For this reason, we also measured item-based correlations between lexical frequency and reaction times.
6. 4 Theoretical contribution 6. 4. 1 Findings consistent with our hypotheses The results observed largely confirmed our hypotheses and were congruent with the literature our hypotheses were based on. The nature of phonological representations (Ramus et al., 2010) suggests that the word position effects observed at the lexical level are a consequence of sublexical constraints operating at the level of perception. Thus, the substantially coherent finding of sublexical saliency of word beginnings is in line with what the available literature would predict (Ramus et al., 2010, Beckman, 1998, Smith, 2004). Our predictions were not expressed explicitly in the literature, but resulted from the combination of the predictions made by other studies on the topic. Particularly, Beckman (1998) defined the word beginning saliency principle as a phonological constraint applying in production, and Ramus et al. (2010) suggested that output and input phonological rules influence each other. The combination of these two proposals led to the generation of the first hypothesis: a word beginning saliency principle may be active in sublexical perception. Similarly, the evidence for higher complexity in morphological word endings is in line with a large body of evidence coming from theoretical linguistics and crosslinguistic analysis (for example, Pater, 2006, Harris, 2011). In this case, again, our
hypothesis was not already present in the literature, but resulted from findings in different studies. Crucial for our prediction is the work of Grainger and Ziegler (2011) and Pater (2006). In a connectionist model of reading, Grainger and Ziegler (2011) suggest that bound morphemes, such as -ed in English, may be detected by readers before the word is actually accessed. They suggest the existence of a finegrained parser that detects grapheme chunks representing morphological information, as in, for instance, the string “-ed”. Pater (2006) shows that, in production, morphological chunks receive particular saliency as well. This led to the generation of the hypothesis according to which what Grainger and Ziegler (2011) proposed for reading may be active in speech perception as well. The pattern observed in the clinical sample is consistent with some literature, for instance with Marshall and van der Lely (2009). The interpretation we give to this result is, however, not entirely consistent with the literature, since we claim that the phenomenon observed is not the symptom of a clinical condition. Marshall and van der Lely (2009) showed that children with dyslexia and/or SLI are less accurate in repeating nonwords when they contain non-initial clusters than when they contain initial clusters. They failed to observe the same pattern in TD samples (although they observed a tendency for accuracy to be slightly lower for medial syllables in all TD age groups tested in their study. Analysing only TDs with one ANOVA may have given a significant result). The current study, instead, showed that similar effects are present in neurotypical samples as well (experiments 1, 2 and 3). For this reason, interpreting the results from our clinical sample in light of the interpretation given in Marshall and van der Lely (2009) would not be consistent with our findings.
6. 4. 2 Findings not consistent with our hypotheses The word beginning saliency principle was observed only in unstressed syllables. This result was rather surprising. A possible interpretation of this result was provided at the end of Chapter 2. In short, this finding could be explained if we assume phonological rules (or constraints) to be organised in a hierarchical system, as suggested for instance by Optimality Theory (Prince & Smolensky, 1997, Beckman, 1998). If we make this type of assumption, we could suggest that the saliency generated by stress is more relevant (or more powerful) than the saliency generated by word position effects. As such, stressed syllables are always salient with no relation to their position in the word. Word position effects are thus not detected in stressed syllables. Unstressed syllables, instead, are non-salient, and word position effects apply to them. Since word position effects apply, unstressed word beginnings are salient and unstressed word medial positions are non salient. Another surprising result regards the lack of correlations between lexical frequency and item-based reaction times in Experiment 4. Even if our results support the claim that a large amount of variation is due to morphology, it is hard to believe that lexical frequency is not playing a role at all, considering that several studies in the past found this correlation (Stemberger & McWhinney, 1988, Bertram et al., 2000, Kirjavainen, Nikolaev & Kidd, 2012). From a statistical point of view, absence of a significant result does not indicate absence of the effect (one can make a claim only out of significant results, but no claims out of non significant results). In other words, the absence of correlations may be due to experimental factors rather than to linguistic factors (more on this in section 5.6). Another possible explanation for the absence of frequency based correlations is that the lack of these correlations is due to
the nature of the task performed. As discussed in the introduction to this chapter, the task was built in a way that the representations involved were sublexical (Ramus et al., 2010). For this reason, lexical factors such as lexical frequency may have had little to no effect on the task.
6. 5 Implications of our findings for research The main advancement for research regards the extension of word position effects to the sublexicon. With our set of experiments, we showed that word initial saliency and word final optional saliency apply sublexically. The activation of morphophonological phenomena in an isolated fashion from semantics contributes with relevant data to debates about interfaces (Selkirk, 2002, 2008, 2011) between different domains of grammar (grammar in a broad sense, i.e. as the system that includes syntax, morphology, phonology, semantics and pragmatics). Our results suggest that morphological phenomena, such as morpheme stripping, apply sublexically, with no activation of the lexicon, and, as a consequence, with no activation of (lexical) meaning. The tests also confirm the dissociation between different types of saliency: word position saliency is different from stress saliency, as already suggested in Marshall and van der Lely (2009). Our set of experiments adds new insights about the relationship between these two types of saliency, showing that they operate independently and that they operate hierarchically. Future research on acoustic saliency may find our data useful as a ground for psycholinguistic/acoustic differences in saliency in perception. Future research in morpheme stripping may find our data on sublexical stripping useful. Our data add important information for the debate between rule-
based vs whole-form derivation of morphology, suggesting that certain rule-like phenomena take place, and they take place sublexically (for information on the debate see Pinker & Ullman, 2002a, Stemberger & McWhinney, 1988, Diessel, 2012). Our experiments with EEG-ERPs offer some neurolinguistic evidence of the process, and this may be useful considering that there is a relative imbalance between linguistic, psycholinguistic and neurolinguistic evidence for the processing of inflected forms (morpheme stripping), most of the evidence being based on psycholinguistic data.
6. 6 Implications for clinical practice The word position effects we detected seem to have been ignored in two commonly used standardised assessments of phonological short-term memory (CNREP) and production of tense markers (TEGI). In Chapter 5, we analysed data from a sample of children with developmental language disorders, and we showed the disadvantages that may arise in assessment as a consequence of overlooking word-position effects. Our data show that word length effects observed in nonword repetition tasks such as CNRep (Gathercole & Baddeley, 1996) may be influenced by word position effects. Children with and without language disorders in our sample were more likely to make errors in long words (4 and 5 syllable words) when these contained non-initial clusters. This pattern of performance is in line with what typical performance would predict. Nonword repetition tasks such as CNRep (Gathercole & Baddeley, 1996) are useful tools in the assessment of dyslexia and/or SLI (Norbury, Bishop & Briscoe, 2001, Gallon & Marshall, 2009), and this claim is not challenged
by the present research project. The relation between poor performance in nonword repetition tasks and linguistic performance has been demonstrated both for dyslexia (Gallon & Marshall, 2009) and for SLI (Norbury, Bishop & Briscoe, 2001). What is questioned here, however, is the word length effect observed in CNRep. The statistical analysis we performed on the distribution of clusters shows that these are unbalanced, i.e. that they are always in 4- and 5- syllable words. The analysis we performed on our clinical sample and on TDs shows that children are more likely to make an error if the word contains a non-initial cluster. This suggests that the unbalanced distribution of clusters observed in the statistical analysis may have an effect on the clinical assessment. Particularly, it suggests that the word length effect observed A: may not be due to word length, or at least may be influenced by another factor (non-initial clusters) B: it may be non-clinical in nature, but just present as a processing constraint in human phonology, as suggested by the results in the experiments 1, 2 and 3 in this dissertation. The main implication for clinical practice is that, where possible, clinicians should measure the amount of variability due to word position effects and consider that the word length effect detected in the CNRep may be due to word position effects (or, at least, influenced by them). Speech and language therapists can divide the 5-syllable words in the CNRep between words with and without non initial clusters, and then compare the number of errors in the two conditions. When a larger number of errors is detected in the condition in which non-initial clusters are present, caution must be used in claiming that there is a word length effect in the CNRep. In those cases, the word length effect may well be due to word-position
effects. Alternatively, the test may be adapted to produce a more balanced set of items in terms of consonant cluster distribution.
6. 7 Limitations and future directions The lack of item-based correlations (with frequency) in Experiment 4 was surprising. One possible explanation for this result is lack of power. However, the power analysis shows the study has a power of 77% (assuming a medium effect size). Another possible explanation is that the type of variability in terms of frequency between the items played a role. Frequencies were not normally distributed (unfortunately the choice of items was very constrained because the number of lexical minimal pairs ending in /s/, /z/, /t/, and /d/ is very small). There could be, thus, a limitation in the experimental paradigm. A larger and more distributed set of items may have improved the likelihood of finding frequency effects. In the future, it could be worth repeating the experiment with a larger set of items and/or with items differently distributed in terms of frequency, possibly including only one of the two categories (lexical or morphological). However, as suggested in section 5. 2, the reason may well be that frequency effects were not present at all because the task was tapping into the sublexicon. If this is the case, correlations would not be found even if the task was improved. Another limitation is in the samples used. For the set of experiments on word beginnings, for instance, it may have been interesting to look at English children as well, drawing a parallel with Italian children. Similarly, it would be interesting to observe how children with clinical conditions would perform in the tasks developed for this project, and whether word position effects are larger when a clinical
condition is present (as suggested by Marshall & van der Lely’s data (2009)). Further research could extend the studies to include samples of English speaking TD children and children with Dyslexia, SLI or both. Another interesting possible development could consist in the application of the EEG-ERP paradigm to sublexical items containing morphological information. Specifically, it would be interesting to observe what would happen combining Experiments 5 and 6, i.e. using nonwords with morphological information in the Mismatch Negativity paradigm. The MMN experiment conducted in this PhD was based on the use of real lexical and morphosyntactic minimal pairs. Contrasts like cared-cares, carrying morphological information, revealed to generate a larger MMN than contrasts like side-size, not carrying morphological information. The task needed to be done with real words because the physiological measure was meant to disentangle contribution of semantics to the RTs effects observed in Experiment 4. If lexical contrasts led to a bigger MMN, a semantic explanation of Experiment 4 would have been more likely. Since morphological contrasts led to a bigger MMN, a morphological explanation of Experiment 4 is more likely. Once this is acknowledged, a better understanding of the neurophysiological processes associated with sublexical morpheme stripping can only be achieved performing the task with nonwords. MMNs generated by the contrast of items carrying potential morphological information and items without morphological information could be compared. The prediction is that the MMN will be larger in amplitude when elements carrying potential morphological information are contrasted. The items are already available from Experiment 6, so the task would be relatively straightforward to conduct.
Another possible development of this research could consist in the application of these findings for developing clinical tools. For instance, it would be interesting to develop an Android or iOS app for children with dyslexia and/or SLI taking into account word initial or word final position effects. At the moment, there are no Android or iOS applications that specifically address the deficits of children with SLI. A few apps exist for dyslexia, and these, at least partially, can be inspiring. Rello et al. (2014) developed an app for iPad in Spanish, in which children are asked to perform a series of games with graphemes, such as completing words, or finding wrongly spelled words, using the touch interface. Another app developed by Trancreative LLC by Kline (2012) for English suggest the correct spelling of words when children type in a phonologically plausible alternative (for instance, if they write "resipi" the software would suggest the spelling "recipe"). These apps are useful for improving reading abilities. Considering the imbalance between dyslexia and SLI apps, even if the two disorders affect the same percentage of the population, it would be interesting to develop an app for Android for children with SLI aimed at developing sensibility to certain critical features of their language. The app has to be child-friendly, and presented as a game, so that the training will look entertaining to the child, rather than scholastic. A big effort should be made to profit nicely from the touch interface of the tablet. This will make the app much more entertaining, which is of central importance for its success. For instance, for English, one possible game could consist in touching, as fast as possible, the "puppet that speaks well" on the screen. Two puppets talking will be shown on the screen, one correctly inflecting the verbs, and one incorrectly inflecting the verb. Winning a touch videogame is an extremely rewarding practice for children, and as such it is likely that they will
engage in the game until they play well and are able to finish the game. This would be very useful to them for developing their morphosyntactic system. It would be interesting to deal with morphophonological effects as well: some verbs should generate a word final cluster, some should not. The contribution of the app would then be twofold: on the one hand, the number of errors of children could be measured in both conditions, offering a large amount of data useful for research. On the other hand, a properly designed app will help children in improving their performance in these tasks, and thus the app may have positive implications for speech and language intervention.
6. 8 Final reflections The reference dictionary defines a word as “a unit of language, consisting of one or more spoken sounds or their written representation that functions as a principal carrier of meaning”. The idea of words as “units” is also part of the theoretical framework of several academics working in psycholinguistics. According to several researchers, the storage and the processing of complex word forms consists of the storage and processing of those forms as units (see Stemberger & McWhinney, 1988). Words are described as entities similar to bricks that speakers combine to create sentences. Our data suggest that this is true only to an extent. A parallel can be made with physics. In a sense, atoms are units, and they combine to create bigger entities (molecules). However, we are well aware that atoms themselves have a complex structure, and subatomic phenomena are part of the complex processes that make atoms interact with each other. Words are like atoms. Their sub-structure is sometimes overlooked, but the nature of sub-word processes is crucial not only for
words themselves, but also for what concerns the interaction of words in the generation of sentences. Tense inflections are a good example of this idea. Tense inflections belong to the word, usually in the form of bound morphemes. However, their very nature is that of elements that generate dependencies with other words in the sentence, and contribute to the generation of complex structures. The principle can be explained referring to the example below: 1) Mary, the girl you met yesterday at the pub, plays football every Wednesday In the sentence in (1), the bound morpheme –s, belonging to the word plays, is not only part of this word, but is as well-linked to the noun Mary, the subject of the sentence. Word structure, thus, is not important only for words themselves, but is important also from a syntactic point of view. The experiments in this PhD, and others in the literature, show that complex processes take place at the sub-word level, and that words are not always treated as units. The last experiment in this PhD shows how these sub-word processes are active sublexically, with no relation to the meaning of words (i.e. with no relation to semantics). Thus, it appears that a form of morpheme stripping can take place in isolation from semantics. This finding may be important in the debate about interfaces. Particularly, while Chomsky (1995) proposed a model in which syntax and morphosyntax are independent from semantics, Tomasello (2006) proposes a model in which morphosyntax is strictly related to semantics. While, for Chomsky, the application of rules is independent from the meaning of the lexical items, for Tomasello the two aspects cannot be separated: stem, affix and their associated meanings are treated as holistic entities. In other words, while the first proposal suggests that inflected forms are derived by applying rules (Pinker & Ullman, 2002a), the second proposal suggests that inflected
forms are stored as units with their verbs, and as such are not separable from the meaning of the verbs they are attached to (Bertram et al., 2000). While the results in this PhD may seem to fit better with the first proposal, a more articulated explanation may be closer to the right answer. Interface problems remain very relevant in psycholinguistic research, since processing entails connections between speech sounds and meaning, so it necessarily involves interfaces. There is some evidence that morpheme processing is not optimal but rather redundant (Caramazza et al., 1988). Our data show that morpheme stripping can operate in a separate fashion from the lexicon, but it does not exclude that a second form of morpheme processing may take place at the lexical level. Further, it does not exclude that some forms may be stored as units (similar flexibility is also offered by Pinker and Ullman, 2002a, but it remains to be explained the extent of its application and it is not clear the number of inflected forms stored as units in their model). If a large number of forms are not derived with a sublexical rule, but rather stored in their inflected form, it may be argued that both theories, the one from Tomasello (2004) and the one from Pinker and Ullman, (2002a) are valid, because the system is redundant. Some recent research on auditory perception shows the existence of redundant systems (Pieszek, Widmann, Gruber & Schröger, 2013). In their study, conducted with ERPs in a MMN paradigm, the authors showed that our brain stores partially redundant (but also contradictory) predictions about sound patterns. The task was performed with low level stimuli such as tones varying in pitch. The linguistic knowledge available, however, suggests that the brain may make redundant predictions on grammatical structures as well (grammatical in broad sense), and thus that similar effects may also be elicited by morphologically complex words. In accepting that linguistic rules have
to be simple, or we would not be able to explain the paradox of language acquisition, scientists often believed that these rules have to be elegant as well. However, it may be the case that nature found it more adaptive to have logical and non-logical systems coexist in our brain, creating a non-elegant but very efficient tool which is able to build structures we still barely understand. It remains for future research to address this issue.
REFERENCES Archibald, L. M., & Gathercole, S. E. (2006). Nonword repetition: A comparison of tests. Journal of Speech, Language, and Hearing Research,49(5), 970-983. Akamatsu, T. (1995). The commutation test and minimal pairs. Contextos, (25), 1340. Ambridge, B., & Lieven, E. V. (2011). Child language acquisition: Contrasting theoretical approaches. Cambridge University Press. Angrilli, A., Dobel, C., Rockstroh, B., Stegagno, L., & Elbert, T. (2000). EEG brain mapping of phonological and semantic tasks in Italian and German languages. Clinical Neurophysiology, 111(4), 706-716. Astorkiza, J. R. C. (2007). Minimal contrast and the phonology-phonetics interaction. University of Southern California. PhD dissertation. ProQuest. Baayen, R. H., Shaoul, C., Willits, J., and Ramscar, M. (2015). Comprehension without segmenta-tion: A proof of concept with naive discriminative learning. Manuscript submitted for publication. Baddeley, A. (1992). Working memory. Science, 255(5044), 556-559. Baddeley, A. (2003). Working memory: looking back and looking forward.Nature Reviews Neuroscience, 4(10), 829-839.
Barlow, J. A., & Gierut, J. A. (2002, February). Minimal pair approaches to phonological remediation. In Seminars in Speech and Language (Vol. 23, No. 1, pp. 57-68). Theime medical publishers inc. Bakker, I., MacGregor, L. J., Pulvermüller, F., & Shtyrov, Y. (2013). Past tense in the brain's time: Neurophysiological evidence for dual-route processing of past-tense verbs. Neuroimage, 71, 187-195. Beckman J. N. (1998). Positional faithfulness. Amherst, MA: University of Massachusetts Amherst, PhD dissertation. Beckman, J. N. (2013). Positional faithfulness: an optimality theoretic treatment of phonological asymmetries. Routledge. Benasich, A. A., Choudhury, N., Friedman, J. T., Realpe-Bonilla, T., Chojnowska, C., & Gou, Z. (2006). The infant as a prelinguistic model for language learning impairments:
behavior. Neuropsychologia, 44(3), 396-411. Berko J. (1958), The child learning of English morphology, Massachusetts Institute of Technology. Bertram, R., Schreuder, R., & Baayen, R. H. (2000). The balance of storage and computation in morphological processing: The role of word formation type, affixal homonymy, and productivity. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(2), 489.
Benasich A. A. , & Tallal P., (2002). Infants discrimination of rapid auditory cues predicts later language impairment. Behavioural Brain Research, 136, 31-49. Bishop D. & Snowling M. (2004). Developmental Dyslexia and Specific Language Impairment: same or different?, Psychological Bullettin, 130-6, 858-886. Bishop, D. V., & McArthur, G. M. (2005). Individual differences in auditory processing in specific language impairment: A follow-up study using event-related potentials and behavioural thresholds. Cortex, 41(3), 327-341. Blache, S. E., Parsons, C. L., & Humphreys, J. M. (1981). A minimal-word-pair model for teaching the linguistic significance of distinctive feature properties.Journal of Speech and Hearing Disorders, 46(3), 291-296. Blevins, J. 2004. Evolutionary phonology: The emergence of sound patterns. Cambridge: Cambridge University Press. Blöte, A. W., Resing, W. C., Mazer, P., & Van Noort, D. A. (1999). Young children's organizational strategies on a same–different task: A microgenetic study and a training study. Journal of Experimental Child Psychology, 74(1), 21-43. Boersma, P. (1998). Functional phonology: Formalizing the interactions between articulatory and perceptual drives. Holland Academic Graphics/IFOTT. Bouton, S., Serniclaes, W., Bertoncini, J., & Cole, P. (2012). Perception of speech features by French-speaking children with cochlear implants. Journal of Speech, Language, and Hearing Research, 55(1), 139-153.
Bouton, S., Colé, P., & Serniclaes, W. (2012). The influence of lexical knowledge on phoneme discrimination in deaf children with cochlear implants. Speech Communication, 54(2), 189-198. Brancalioni, A. R., Bonini, J. B., Gubiani, M. B., & Keske-Soares, M. (2012). Ambientes favorecedores para produção dos fonemas plosivos/k/e/g.. Distúrbios da Comunicação. ISSN 2176-2724, 24(1). Browman, C. P. (1978). Tip of the tongue and slip of the ear: implications for language processing. University of California, UCLA Working Papers in Phonology, 42, 1–159. Brown, J. D. (1995). The elements of language curriculum: A systematic approach to program development. Heinle & Heinle Publishers, 20 Park Plaza, Boston, MA 02116.. Brown, R., & McNeill, D. (1966). The “tip of the tongue” phenomenon. Journal of Verbal Learning and Verbal Behavior, 5(4), 325–337. Caramazza, A., Basili, A. G., Koller, J. J., & Berndt, R. S. (1981). An investigation of repetition and language processing in a case of conduction aphasia. Brain and Language, 14(2), 235-271. Caramazza, A., Laudanna, A., & Romani, C. (1988). Lexical access and inflectional morphology. Cognition, 28(3), 297-332. Carnie, A. (2007; 2013). Syntax: a generative introduction. 2nd, 3rd edition. Blackwell, London.
Cheng, X., Schafer, G., & Riddell, P. M. (2014). Immediate auditory repetition of words and nonwords: An ERP study of lexical and sublexical processing.PloS one, 9(3), e91988. Chiat S. , (2001). Mapping theories of developmental language impairment: Premises, predictions and evidence, Language and Cognitive Processes, 16 (2/3), 113–142. Chiat, S., Roy, P. (2007). The preschool repetition test: An evaluation of performance in typically developing and clinically referred children. Journal of Speech, Language, and Hearing Research, 50(2), 429-443. Chomsky, N. (1957), Syntactic Structures, The Hague/Paris: Mouton Chomsky, N., & Halle, M. (1968). The sound pattern of English. New York: Harper and Row. Chomsky, N. (1995). The minimalist program (Vol. 1765). Cambridge, MA: MIT press. Cilibrasi, L., Stojanovik, V., & Riddell, P. (2015). Word Position and Stress Effects in Consonant Cluster Perception and Production. Dyslexia, 21(1), 50-59. Clahsen, H. (2008). Behavioral methods for investigating morphological and syntactic processing in children. Language Acquisition and Language Disorders, 44, 1. Cole, R. A., & Jakimik, J. (1980). How are syllables used to recognize words? The Journal of the Acoustical Society of America, 67(3),965–970.
Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: a dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108(1), 204. Connine, C. M., Blasko, D. G., & Titone, D. (1993). Do the beginnings of spoken words have a special status in auditory word recognition?. Journal of Memory and Language, 32(2), 193-210. Contemori C. , Garraffa M. (2010). A cross modality study on syntax in SLI: the limits of computation as a measure of linguistic abilities. Lingua, 8 (120) , 19401955 Conti-Ramsden G. , Botting N. , Faragher B. (2001), Psycholinguistic markers for Specific Language Impairment, Journal of child Psychology and Psychiatry, (42-6), 741-748. Conti-Ramsden, G., & Hesketh, A. (2003). Risk markers for SLI: a study of young language-learning children. International Journal of Language & Communication Disorders, 38(3), 251-263. Cornell, S. A., Lahiri, A., & Eulitz, C. (2011). “What you encode is not necessarily what you store”: Evidence for sparse feature representations from mismatch negativity. Brain Research, 1394, 79-89 Crosbie, S., Holm, A., & Dodd, B. (2005). Intervention for children with severe speech disorder: a comparison of two approaches. International Journal of Language & Communication Disorders, 40(4), 467-491.
Darcy, I., Peperkamp, S., & Dupoux, E. (2007). Bilinguals play by the rules: Perceptual
Phonology, 9, 411-442. Darcy, I., Ramus, F., Christophe, A., Kinzler, K., & Dupoux, E. (2009). Phonological knowledge in compensation for native and non-native assimilation. Variation and Gradience in Phonetics and Phonology, 14, 265. Dehaene S. (2009), Reading in the brain, the new science of how we read, Penguin Books, New York. Dehaene-Lambertz G. , Baillet S. (1998). A phonological representation in the infant brain. NeuroReport, 9:1885–8. Dehaene-Lambertz, G., Dupoux, E., & Gout, A. (2000). Electrophysiological correlates
Neuroscience, Journal of, 12(4), 635-647. Diependaele, K., Grainger, J., & Sandra, D. (2012). Derivational morphology and skilled reading. The Cambridge Handbook of Psycholinguistics., 311-332. Diessel, H. (2013). Construction grammar and first language acquisition. In T. Hoffmann & G. Trousdale (Éds.), The Oxford Handbook of Construction Grammar, Oxford, UK. Dietrich, C., Swingley, D., & Werker, J. F. (2007). Native language governs interpretation of salient speech sound differences at 18 months. Proceedings of the National Academy of Sciences, 104(41), 16027-16031.
Dollaghan, C. (2008). Diagnosis of specific language impairment. Encyclopedia of Language and Literacy Development (pp. 1-5). London, ON: Canadian Language and Literacy Research Network. Dufour, S. , Nguyen, N. , & Frauenfelder, U. H. (2007). The perception of phonemic contrasts in a non-native dialect, J. Acoust. Soc. Am. 121, EL131–EL136. Earl, C. M. (2011). A critique of minimal pair materials: A guide for speechlanguage pathologists. PSHA, 5. Echols C. H. , (1996). A role for stress in early speech segmentation. In: J.L. Morgan & K. Demuth (eds.), Signal to Syntax . Mahwah, NJ: LEA. Eimas, P. D., Siqueland, E. R., Jusczyk, P., & Vigorito, J. (1971). Speech perception in infants. Science, 171(3968), 303-306. Eisner, J. (2002). Comprehension and compilation in Optimality Theory. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics (pp. 56-63). Association for Computational Linguistics. Elliott, E. A., Braun, M., Kuhlmann, M., & Jacobs, A. M. (2012). A dual-route cascaded model of reading by deaf adults: evidence for grapheme to viseme conversion. Journal of Deaf Studies and Deaf Education, 17(2), 227-243. Feldman, N. H., Griffiths, T. L., & Morgan, J. L. (2009). Learning phonetic categories by learning a lexicon. In Proceedings of the 31st annual conference of the cognitive science society (Vol. 2208, p. 213). Austin, TX: Cognitive Science Society.
Field, A. (2009). Discovering statistics using SPSS. Sage publications. Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1-2), 3-71. Foote, J. A., Holtby, A. K., & Derwing, T. M. (2011). Survey of the teaching of pronunciation in adult ESL programs in Canada, 2010. TESL Canada journal, 29(1), 1-22. Fonteneau, E., & van der Lely, H. K. (2008). Electrical brain responses in languageimpaired children reveal grammar-specific deficits. PLoS One, 3(3), e1832. Friedmann, N., & Lukov, L. (2008). Developmental surface dyslexias.Cortex, 44(9), 1146-1160. Friedmann N. , Novogrodsky R, (2008) Subtypes of SLI: sy-SLI, pho-SLI, le-SLI, and pra-SLI, Proceedings of GALA-2008. Friedmann N. , Rizzi L. , Belletti A. (2009) Relativized relatives, Lingua 119. 67– 88. Friedmann, N., & Gvion, A. (2002). Modularity in developmental disorders: Evidence from Specific Language Impairment and peripheral dyslexias.Behavioral and Brain Sciences, 25(06), 756-757. Frisch, S. A., Large, N. R., & Pisoni, D. B. (2000). Perception of wordlikeness: Effects of segment probability and length on the processing of nonwords.Journal of memory and language, 42(4), 481-496.
Fudge, E. (1987). Branching structure within the syllable. Journal of Linguistics, 23(02), 359-377. Galaburda, A. M., Menard, M. T., & Rosen, G. D. (1994). Evidence for aberrant auditory anatomy in developmental dyslexia. Proceedings of the National Academy of Sciences, 91(17), 8010-8013. Gallon Stuart, N., and Marshall, C. (2009) Using Non-word Repetition Tasks with Children with Dyslexia, Dyslexia Review, 21, 1, 4-8. Gaskell, M. G., & Marslen-Wilson, W. D. (1996). Phonological variation and inference in lexical access. Journal of Experimental Psychology: Human perception and performance, 22(1), 144. Gathercole, S. E., & Baddeley, A. D. (1996). The children's test of nonword repetition. Psychological Corporation. Gierut, J. A. (2001). Complexity in Phonological Treatment Clinical Factors. Language, Speech, and Hearing Services in Schools, 32(4), 229-241. Gleason, J. Berko & Ratner, N. B. (Eds.) (2009) The Development of Language, 7th Edition. Boston: Pearson/Allyn & Bacon Goldrick, M., & Rapp, B. (2007). Lexical and post-lexical phonological representations in spoken production. Cognition, 102(2), 219-260.
Goldsmith, J., & Laks, B. (2011). Generative phonology: its origins, its principles, and its successors. In Waugh, L. & Joseph, JE, The Cambridge History of Linguistics. Grainger J. &Ziegler J. C. (2011), A dual route approach to orthographic processing, Frontiers in Psychology, 2-54, 1-13. Guasti, M. T. (2004). Language acquisition: The growth of grammar. MIT Press, Cambridge, MA. Gunawardena, A., Krikeb, J., Moelijker, N., Puppi, M., & Witteveen, S. (2011). Speech production and speech processing observed, UVA SSC 191, 1-9. Guttorm (2010), Newborn event-related potentials predict poorer pre-reading skills in children at risk for dyslexia, Journal of Learning Disabilities vol. 43 no. 5 391401. Hage, S. R. D. V., Cendes, F., Montenegro, M. A., Abramides, D. V., Guimarães, C. A., & Guerreiro, M. M. (2006). Specific language impairment: linguistic and neurobiological aspects. Arquivos de neuro-psiquiatria, 64(2A), 173-180. Handy, T. C. (2005). Event-related potentials: A methods handbook. MIT press. Harris, J. (2011) Deletion. In: van Oostendorp, M and Ewen, CJ and Hume, EV and Rice, K, (eds.) The Blackwell Companion to Phonology. (1597 - 1621). WileyBlackwell: Oxford, UK.
Hayes, B. (2012). How predictable is Italian word stress. National Chiao Tung University, May, 11. Hayes-Harb, R. (2007). Lexical and statistical evidence in the acquisition of second language phonemes. Second Language Research, 23(1), 65-94. Hirsh-Pasek, K., & Golinkoff, R. M. (1999). The origins of grammar: Evidence from early language comprehension. MIT Press, Boston. Hoonhorst, I., Medina, V., Colin, C., Markessis, E., Radeau, M., Deltenre, P., & Serniclaes, W. (2011). Categorical perception of voicing, colors and facial expressions: a developmental study. Speech Communication, 53(3), 417–430. Hoover, W. A., & Gough, P. B. (1990). The simple view of reading. Reading and Writing, 2(2), 127-160. Hulslander, J., Talcott, J., Witton, C., DeFries, J., Pennington, B., Wadsworth, S., ... & Olson, R. (2004). Sensory processing, reading, IQ, and attention.Journal of Experimental Child Psychology, 88(3), 274-295. Iverson, P., Hazan, V., & Bannister, K. (2005). Phonetic training with acoustic cue manipulations: A comparison of methods for teaching English/r/-/l/to Japanese adults. The Journal of the Acoustical Society of America, 118(5), 3267-3278. Jacquemot, C., Dupoux, E., & Bachoud-Lévi, A. C. (2007). Breaking the mirror: Asymmetrical disconnection between the phonological input and output codes. Cognitive Neuropsychology, 24(1), 3-22.
Jakobson, R. (1962). Selected Writings Vol 1 Phonological Studies. Mouton & Company. Jakubowicz, C., Nash, L., Rigaut, C., & Gerard, C. L. (1998). Determiners and clitic pronouns in French-speaking children with SLI. Language Acquisition, 7(2-4), 113160. Joanisse M. F. (2004), Specific Language Impairments in children. Phonology, semantics and the English past tense, Current Directions in Psychological Science, (13-4), 156-160. Joanisse M. F. , Manis F. R. , Keating P. , Seidenberg M. S. (2000), Language deficits in dyslexic children: speech perception, phonology and morphology, Journal of Experimental Child Psychology, (77), 30-60. Jobard, G., Crivello, F., & Tzourio-Mazoyer, N. (2003). Evaluation of the dual route theory of reading: a metanalysis of 35 neuroimaging studies. Neuroimage,20(2), 693712. James, L. E., & Burke, D. M. (2000). Phonological priming effects on word retrieval and tip-of-the-tongue experiences in young and older adults. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(6), 1378. Jusczyk, P. W., Cutler, A., & Redanz, N. J. (1993). Infants' preference for the predominant stress patterns of English words. Child Development, 64(3), 675-687.
Jusczyk, P. W., Smolensky, P., & Allocco, T. (2002). How English-learning infants respond to markedness and faithfulness constraints. Language Acquisition, 10(1), 3173. Kager, R. (1999). Optimality theory. Cambridge: Cambridge University Press. Kahn, D. (2015). Syllable-based generalizations in English phonology. Routledge. Katz, J. S., Wright, A. A., & Bachevalier, J. (2002). Mechanisms of same-different abstract-concept learning by rhesus monkeys (Macaca mulatta).Journal of Experimental Psychology: Animal Behavior Processes, 28(4), 358. Kenstowicz, M. J. (1994). Phonology in generative grammar (pp. 195-249). Cambridge, MA: Blackwell. Kirjavainen M. , Nikolaev A. , Kidd E. (2012), The effect of frequency and phonological neighbourhood density on the acquisition of past tense verbs by Finnish children, Cognitive Linguistics, (23-2), 273-315. Kirk, C., & Demuth, K. (2005). Asymmetries in the acquisition of word-initial and word-final consonant clusters. Journal of Child Language, 32(04), 709-734. Kovelman, I., Norton, E. S., Christodoulou, J. A., Gaab, N., Lieberman, D. A., Triantafyllou, C., ... & Gabrieli, J. D. (2012). Brain basis of phonological awareness for spoken language in children and its disruption in dyslexia.Cerebral Cortex, 22(4), 754-764.
Krott, A., & Lebib, R. (2013). Electrophysiological evidence for a neural substrate of morphological rule application in correct wordforms. Brain research, 1496, 70-83. Kuhl, P. K., Williams, K. A., Lacerda, F., Stevens, K. N., & Lindblom, B. (1992). Linguistic experience alters phonetic perception in infants by 6 months of age. Science, 255(5044), 606-608. Kurisu, K. (2001). The phonology of morpheme realization, PhD dissertation, University of California Santa Cruz, Santa Cruz. LaBerge, D. (1983). Spatial extent of attention to letters and words. Journal of Experimental Psychology: Human Perception and Performance, 9(3), 371 Lavric, A., Pizzagalli, D., Forstmeier, S., & Rippon, G. (2001). A doubledissociation of English past-tense production revealed by event-related potentials and low-resolution electromagnetic tomography (LORETA). Clinical Neurophysiology, 112(10), 1833-1849. Law F. , Strange W. (2010), Perception of Canadian French word-final vowels in lexical and morphosyntactic minimal pairs by English learners of French, Journal of the Acoustic Society of America, 128, 2488. Leminen, A., Leminen, M., Kujala, T., & Shtyrov, Y. (2013). Neural dynamics of inflectional and derivational morphology processing in the human brain. Cortex, 49(10), 2758-2771. Leonard, C. M., Lombardino, L. J., Walsh, K., Eckert, M. A., Mockler, J. L., Rowe, L. A., ... & DeBose, C. B. (2002). Anatomical risk factors that distinguish dyslexia
from SLI predict reading skill in normal children. Journal of Communication Disorders, 35(6), 501-531. Levis, J. & Cortes, V. (2008). Minimal pairs in spoken corpora: Implications for pronunciation assessment and teaching. In C. A. Chapelle, Y.-R. Chung, & J. Xu (Eds.), Towards adaptive CALL: Natural language processing for diagnostic language assessment (pp. 197-208). Ames, IA: Iowa State University. Liberman, A. M., Harris, K. S., Hoffman, H. S. & Griffith, B. C. (1957). The discrimination of speech sounds within and across phoneme boundaries. Journal of Experimental Psychology 54 (5): 358–368. Liegeois-Chauvel, C., Musolino, A., & Chauvel, P. (1991). Localization of the primary auditory area in man. Brain, 114(1), 139-153.Luck S. J. (2005), An introduction to the event related potential technique, MIT press, Cambridge. Macmillan, N. A., & Creelman, C. D. (2005). Detection Theory: A User’s Guide Lawrence Erlbaum Associates. New York. Männel, C., & Friederici, A. (2008). Event-related brain potentials as a window to children’s language processing. Developmental Psycholinguistics. On-line Methods in Children’s Language Processing, 29-72. Marcus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J., Xu, F., & Clahsen, H. (1992). Overregularization in language acquisition. Monographs of the society for research in child development, i-178.
Marshall C. R. , Harcourt-Brown S. , Ramus F. , Van der Lely H. K. J. (2009), The link between prosody and language skills in children with Specific Language Impairment and/or Dyslexia, International Journal of Language Communication Disorders, (44-4), 466-488. Marshall C. R. , Van der Lely H. K. J. (2007), The impact of phonological complexity on past tense inflection in children with G-SLI, Advances in SpeechLanguage Pathology, (9-3), 191-203. Marshall, C. R., & Van der Lely, H. K. (2009). Effects of word position and stress on onset cluster production: Evidence from typical development, specific language impairment, and dyslexia. Language, 85(1), 39-57. Marshall C. R. , Van der Lely H. K. J. (2012), Irregular past tense forms in English: how data from children with specific language impairment contribute to models of morphology, Morphology, (22), 121-141. Maslen R. J. C. , Theakston A. L. , Lieven E. V. M. , Tomasello M. (2004), A dense corpus study of past tense and plural overregularization in English, Journal of Speech, Language and Hearing Research, (47-d), 1319-1333. Marslen-Wilson, W., & Zwitserlood, P. (1989). Accessing spoken words: the importance of word onsets. Journal of Experimental Psychology, 15(3), 576. Mason, K., Rowley, K., Marshall, C. R., Atkinson, J. R., Herman, R., Woll, B., & Morgan, G. (2010). Identifying specific language impairment in deaf children
acquiring British Sign Language: Implications for theory and practice. British Journal of Developmental Psychology, 28(1), 33-49. Matchmann V. A. , Wulfeck B. , Weismer S. E. (1999), Morphological productivity in children with normal language and SLI: a study of the English past tense, Journal of Speech, Language and Hearing Research, (42), 206-219. McArthur, G. M., Hogben, J. H., Edwards, V. T., Heath, S. M., & Mengler, E. D. (2000). On the “specifics” of specific reading disability and specific language impairment. Journal of Child Psychology and Psychiatry, 41(07), 869-874. McArthur, G. M., & Hogben, J. H. (2001). Auditory backward recognition masking in children with a specific language impairment and children with a specific reading disability. The Journal of the Acoustical Society of America,109(3), 1092-1100. McCandliss, B. D., Fiez, J. A., Protopapas, A., Conway, M., & McClelland, J. L. (2002). Success and failure in teaching the [r]-[l] contrast to Japanese adults: Tests of a Hebbian model of plasticity and stabilization in spoken language perception. Cognitive, Affective, & Behavioral Neuroscience, 2(2), 89-108. McCarthy, J. J. (1979). On stress and syllabification. Linguistic Inquiry, 443-465. McClelland, J. L. & Elman, J. L. (1986). The TRACE Model of Speech Perception. Cognitive Psychology, 18, 1-86. McClelland, J. L., & Patterson, K. (2002). Rules or connections in past-tense inflections: what does the evidence rule out?. Trends in Cognitive Sciences,6(11), 465-472.
McClelland, J. L. & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. McGregor, K. K., & Leonard, L. B. (1994). Subject Pronoun and Article Omissions in the Speech of Children With Specific Language ImpairmentA Phonological
Research, 37(1), 171-181. Mehler, J., Dupoux, E., Nazzi, T., & Dehaene-Lambertz, G. (1996). Coping with linguistic diversity: The infant’s viewpoint. Signal to syntax: Bootstrapping from speech to grammar in early acquisition, 101-116. Miller C. A. , Leonard L. B. , Finneran D. (2008), Grammaticality judgments in adolescents with and without language impairment, International Journal of Language Communication Disorders, (43-3), 346-360. Moura, O., Simões, M. R., & Pereira, M. (2014). WISC-III Cognitive Profiles in Children with Developmental Dyslexia: Specific Cognitive Disability and Diagnostic Utility. Dyslexia, 20(1), 19-37. Mora, J. C. (2005). Lexical knowledge effects on the discrimination of non-native phonemic contrasts in words and nonwords by Spanish/Catalan bilingual learners of English. In ISCA Workshop on Plasticity in Speech Perception. Moreland M. L. (2011), The status of the post-initial jod: evidence from British English wordlikeness judgements, PhD dissertation, University of Reading, UK.
Morton, J. (1969). Interaction of information in word recognition. Psychological Review, 76(2), 165. Munson, B., Edwards, J., & Beckman, M. E. (2012). Phonological representations in language acquisition: Climbing the ladder of abstraction. The Oxford handbook of laboratory phonology, 288-309. Näätänen, R., Paavilainen, P., Rinne, T., & Alho, K. (2007). The mismatch negativity (MMN) in basic research of central auditory processing: a review. Clinical Neurophysiology, 118(12), 2544-2590. Navarrete, E., & Costa, A. (2005). Phonological activation of ignored pictures: Further evidence for a cascade model of lexical access. Journal of Memory and Language, 53(3), 359-377. Naeser, M. A., & Hayward, R. W. (1978). Lesion localization in aphasia with cranial computed tomography and the Boston Diagnostic Aphasia Exam. Neurology, 28(6), 545-545. Nazzi, T. (2005). Use of phonetic specificity during the acquisition of new words: Differences between consonants and vowels. Cognition, 98(1), 13-30. Nazzi, T., Bertoncini, J., & Mehler, J. (1998). Language discrimination by newborns: toward an understanding of the role of rhythm. Journal of Experimental Psychology: Human Perception and Performance, 24(3), 756.
Newman, A. J., Ullman, M. T., Pancheva, R., Waligura, D. L., & Neville, H. J. (2007). An ERP study of regular and irregular English past tense inflection. Neuroimage, 34(1), 435-445. Newton, C., Chiat, S., & Hald, L. (2008). Evaluation of a novel technique for assessing speech discrimination in children. Clinical Linguistics & Phonetics,22(45), 325-333. Nicolson, R. I., Fawcett, A. J., & Dean, P. (2001). Developmental dyslexia: the cerebellar deficit hypothesis. TRENDS in Neurosciences, 24(9), 508-511. Nithart C. , Demont E. , Majerus S. , Leybaert J. , Poncelet M. , Metz-Lutz M. (2009), Reading disabilities in SLI and Dyslexia result from distinct phonological impairments, Developmental Neuropsychology, (34-3), 296-311. Nootebom, S. G. (1981). Lexical retrieval from fragments of spoken words: beginnings vs. endings. Journal of Phonetics, 9(4),407–424. Norbury, C. F., Bishop, D. V., & Briscoe, J. (2001). Production of English Finite Verb MorphologyA Comparison of SLI and Mild-Moderate Hearing Impairment. Journal of Speech, Language, and Hearing Research, 44(1), 165-178. Norris, D., McQueen, J. M., & Cutler, A. (2000). Merging information in speech recognition: Feedback is never necessary. Behavioral and Brain Sciences,23(03), 299-325. Novogrodsky, R. & Friedmann, N. (2010). Not all dependencies are impaired in Syntactic-SLI: Binding in children with a deficit in whmovement. Proceedings of the
34th Boston University Child Language Development. Somerset, MA: Cascadilla Press. Oetting J. B. , Rice M. L. (1993), Plural acquisition in children with Specific Language Impairment, Journal of Speech, Language and Hearing Research, (36), 1236-1248 (abstract) Padovan A. (2013). Why Does Generative Linguistics Resist The Concept of Protolanguage? - 19 ICL Geneva. Pallier, C., Colomé, A., & Sebastián-Gallés, N. (2001). The influence of nativelanguage phonology on lexical access: Exemplar-based versus abstract lexical entries. Psychological Science, 12(6), 445-449. Pater, J. (2006). The locus of exceptionality: Morpheme-specific phonology as constraint indexation. University of Massachusetts Occasional Papers in Linguistics, 32, 259-296. Paulesu, E., Démonet, J. F., Fazio, F., McCrory, E., Chanoine, V., Brunswick, N., ... &
unity.Science, 291(5511), 2165-2167. Pennington, B. F. (2006). From single to multiple deficit models of developmental disorders. Cognition, 101(2), 385-413. Pieszek, M., Widmann, A., Gruber, T., & Schröger, E. (2013). The human brain maintains contradictory and redundant auditory sensory predictions. PloS one,8(1), e53634.
Pinker, S. & Ullman M. T. (2002a), The past and future of past tense, Trends in Cognitive Sciences, (6-11), 456-462. Pinker, S., & Ullman, M. T. (2002b). Combination and structure, not gradedness, is the issue. Trends in Cognitive Sciences, 6(11), 472-474. Pitt, M. A., & Samuel, A. G. (1995). Lexical and sublexical feedback in auditory word recognition. Cognitive Psychology, 29(2),149–188. Prasada, S., & Pinker, S. (1993). Generalisation of regular and irregular morphological patterns. Language and Cognitive Processes, 8(1), 1-56. Poll, G. H., Betz, S. K., & Miller, C. A. (2010). Identification of clinical markers of specific language impairment in adults. Journal of Speech, Language, and Hearing Research, 53(2), 414-429. Prince, A., & Smolensky, P. (1997). Optimality: From neural networks to universal grammar. Science, 275(5306), 1604-1610. Pulvermüller, F., Kujala, T., Shtyrov, Y., Simola, J., Tiitinen, H., Alku, P., ... & Näätänen, R. (2001). Memory traces for words as revealed by the mismatch negativity. Neuroimage, 14(3), 607-616. Pulvermüller, F., & Shtyrov, Y. (2003). Automatic processing of grammar in the human brain as revealed by the mismatch negativity. Neuroimage, 20(1), 159-172.
Pulvermüller, F., Shtyrov, Y., Kujala, T., & Näätänen, R. (2004). Word-specific cortical activity as revealed by the mismatch negativity. Psychophysiology, 41(1), 106-112. Raphael, L. J. (1972). Preceding vowel duration as a cue to the perception of the voicing characteristic of word-final consonants in American English. The Journal of the Acoustical Society of America, 51(4B), 1296-1303. Ramus, F., Rosen, S., Dakin, S. C., Day, B. L., Castellote, J. M., White, S., & Frith, U. (2003). Theories of developmental dyslexia: insights from a multiple case study of dyslexic adults. Brain, 126(4), 841-865. Ramus, F., Peperkamp, S., Christophe, A., Jacquemot, C., Kouider, S., & Dupoux, E. (2010). A psycholinguistic perspective on the acquisition of phonology. Laboratory phonology, 10(3), 311-340. Raven, J. C. (1995). Coloured progressive matrices. San Antonio, Texas: Pearson Assessment. Reis, A., & Castro-Caldas, A. (1997). Illiteracy: A cause for biased cognitive development. Journal of the International Neuropsychological Society, 3(05), 444450. Rello, L., & Baeza-Yates, R. (2014, April). Evaluation of Dyswebxia: A reading app designed for people with dyslexia. In Proceedings of the 11th Web for All Conference (p. 10). ACM.
Rice M. L. , Wexler K. (1996), Toward tense as a clinical marker of Specific Language Impairment in English speaking children, Journal of Speech and Hearing Research, (39-dec), 1239-1257. Rice, M. L., & Wexler, K. (2001). Rice/Wexler test of early grammatical impairment. Psychological Corporation. Rinker, T., Kohls, G., Richter, C., Maas, V., Schulz, E., & Schecker, M. (2007). Abnormal frequency discrimination in children with SLI as indexed by mismatch negativity (MMN). Neuroscience Letters, 413(2), 99-104. Rispens J. , Been P, (2007), Subject-verb agreement in phonological processing in developmental dyslexia and specific language impairment (SLI): a closer look, International Journal of Language Communication Disorders, (42-3), 293-305. Roach P. (1983, 1991, 2000), English Phonetics and phonology, Cambridge University press, UK. Robertson E. K. , Joanisse M. F. , Desroches A. S. , Terry A. (2012), Past-tense morphology and phonological deficits in children with Dyslexia and children with Language Impairment, Journal of Learning Disabilities, (20-10), 1-11. Rost, G. C., & McMurray, B. (2009). Speaker variability augments phonological processing in early word learning. Developmental Science, 12(2), 339-349. Rugg, M. D., & Curran, T. (2007). Event-related potentials and recognition memory. Trends in Cognitive Sciences, 11(6), 251-257.
Saben, C. B., & Ingham, J. C. (1991). The effects of minimal pairs treatment on the speech-sound production of two children with phonologic disorders.Journal of Speech, Language, and Hearing Research, 34(5), 1023-1040. Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-monthold infants. Science, 274(5294), 1926-1928. Sartori, G., Job, R., & Tressoldi P. E. (2006). DDE-2, Batteria per la valutazione della dislessia e della disortografia evolutiva. Firenze: Giunti O.S. Scharinger, M., Lahiri, A., & Eulitz, C. (2010). Mismatch negativity effects of alternating
Neurolinguistics, 23(4), 383-399. Scheer, T., & Szigetvári, P. (2005). Unified representations for stress and the syllable. Phonology, 22(01), 37-75. Schuele, C. M. (2004). The impact of developmental speech and language impairments on the acquisition of literacy skills. Mental Retardation and Developmental Disabilities Research Reviews, 10(3), 176-183. Seidenberg, M. S. (2005). Connectionist models of word reading. Current Directions in Psychological Science, 14, 238-242. Selkirk, E. (2002). The syntax-phonology interface. International Encyclopedia of the Social and Behavioral Sciences,. Section 3.9,Article 23. Elsevier, 2002
Selkirk, E., (2008), The prosodic structure of function words, in Optimality Theory in Phonology: A Reader, Blackwell. Selkirk, E. (2011). The syntax-phonology interface. The Handbook of Phonological Theory, 2, 435-483. Shaywitz, S. E., & Shaywitz, B. A. (2005). Dyslexia (specific reading disability). Biological Psychiatry, 57(11), 1301-1309. Shen, X. S., & Lin, M. (1991). A perceptual study of Mandarin tones 2 and 3.Language and Speech, 34(2), 145-156. Shtyrov, Y. and Pulvermüller, F. (2002), Memory traces for inflectional affixes as shown by mismatch negativity. European Journal of Neuroscience, 15: 1085–1091. doi: 10.1046/j.1460-9568.2002.01941.x Shtyrov, Y., & Pulvermüller, F. (2007). Language in the mismatch negativity design. Journal of Psychophysiology, 21(3), 176-187. Shtyrov, Y., Kujala, T., & Pulvermüller, F. (2010). Interactions between language and attention systems: early automatic lexical processing?. Journal of Cognitive Neuroscience, 22(7), 1465-1478. Shtyrov, Y., Pulvermüller, F., Näätänen, R., & Ilmoniemi, R. J. (2003). Grammar processing outside the focus of attention: an MEG study. Journal of Cognitive Neuroscience, 15(8), 1195-1206.
Smith, J. (2002). Phonological augmentation in prominent positions. University of Massachusetts, Amherst. PhD dissertation. Smith, J. L. (2004). Phonological augmentation in prominent positions. Routledge, Abingdon-on-Thames, UK. Snowling, M. J. (2005). Language skills and learning to read: the dyslexia spectrum. In Dyslexia, speech and language: A practitioner’s handbook, 1-14. Swingley, D., & Aslin, R. N. (2000). Spoken word recognition and lexical representation in very young children. Cognition, 76(2), 147-166. Stein, J., & Walsh, V. (1997). To see but not to read; the magnocellular theory of dyslexia. Trends in Neurosciences, 20(4), 147-152. Stein, J. F., Riddell, P. M., & Fowler, S. (1988). Disordered vergence control in dyslexic children. British Journal of Ophthalmology, 72(3), 162-166. Steinberg, J., Truckenbrodt, H., & Jacobsen, T. (2011). Phonotactic constraint violations in German grammar are detected automatically in auditory speech processing: A human event-related potentials study. Psychophysiology, 48(9), 12081216. Stemberger J. P. & MacWhinney B. (1986), Frequency and the lexical storage of regularly inflected forms, Memory & Cognition, 14 (1), 17-26.
Steriade, D. (1995). Underspecification and markedness. In Selkirk, E., & Goldsmith, J. A. (1995). The handbook of phonological theory.Oxford/Cambridge, Blackwell, 114-174. Stojanovik V. , Riddell P. M. (2008), Expressive versus receptive language skills in specific reading disorder, Clinical Linguistics Phonetics, 22, 4-5, 305-92. Strauss, T. J., Harris, H. D., & Magnuson, J. S. (2007). jTRACE: A reimplementation and extension of the TRACE model of speech perception and spoken word recognition. Behavior Research Methods, 39(1), 19-30. Szenkovits, G., & Ramus, F. (2005). Exploring dyslexics' phonological deficit I: lexical vs sub-lexical and input vs output processes. Dyslexia, 11(4), 253-268. Tanaka, H., Black, J. M., Hulme, C., Stanley, L. M., Kesler, S. R., WhitfieldGabrieli, S., ... & Hoeft, F. (2011). The brain basis of the phonological deficit in dyslexia is independent of IQ. Psychological Science, 22(11), 1442-1451. Thomas, M. S., & McClelland, J. L. (2008). Connectionist models of cognition. The Cambridge handbook of computational psychology, 23-58. Tomasello M. (2006), First stepts toward a usage-based theory of language acquisition, Cognitive Linguistics, (11), 1-2, 61-82. Tomasello, M. (2000). Do young children have adult syntactic competence?. Cognition, 74(3), 209-253.
Tomasina M. O. , Tan K. L. , Ng P. , Beme Y. I. , Graham S. (2011), The past tense debate: Is phonological complexity the key to the puzzle?, Neuroimage, (57), 271280. Torgesen, J. K., & Burgess, S. R. (1998). Consistency of reading-related phonological processes throughout early childhood: Evidence from longitudinalcorrelational and instructional studies. In Word recognition in beginning literacy, 161-188. Treiman, R., Kessler, B., Knewasser, S., Tincoff, R., & Bowman, M. (2000). English speakers’ sensitivity to phonotactic patterns. Papers in laboratory phonology V: Acquisition and the lexicon, 269-282. Trubetzkoy, N. S. (1969). Principles of phonology. Berkeley: University of California Press Tyler, L. K., Randall, B., & Marslen-Wilson, W. D. (2002). Phonology and neuropsychology of the English past tense. Neuropsychologia, 40(8), 1154-1166. Vance, M., Rosen, S., & Coleman, M. (2009). Assessing speech perception in young children
Audiology, 48(10), 708-717. Van der Lely, H. K. J. (2005), Domain Specific cognitive systems: insight from Grammatical-SLI, Trends in Cognitive Science, (9-2), 53-59.
Van der Lely H. K. J. , Ullman M. (1996), The computation and representation of past tense morphology in specifically language impaired and normally developing children, BUCLD 20 Proceedings, (20), 804-815. Van der Lely H. K. J. , Ullman M. T. (2001), Past tense morphology in specifically language impaired and normally developing children, Language and cognitive processes, (16-2/3), 177-217. Vender, M. (2009). Scalar implicatures and developmental dyslexia. InProceedings of the 2nd International Clinical Linguistics Conference. Vidyasagar, T. R., & Pammer, K. (2010). Dyslexia: a deficit in visuo-spatial attention, not in phonological processing. Trends in Cognitive Sciences, 14(2), 57-63. Vitevitch, M. S., & Luce, P. A. (1998). When words compete: Levels of processing in perception of spoken words. Psychological Science, 9(4), 325-329. Vitevitch, M.S. & Luce, P.A. (2004) A web-based interface to calculate phonotactic probability for words and nonwords in English. Behavior Research Methods, Instruments, and Computers, 36, 481-487. Warren, J. E., Wise, R. J., & Warren, J. D. (2005). Sounds do-able: auditory–motor transformations and the posterior temporal plane. Trends in Neurosciences, 28(12), 636-643. Wedel, A. (2012). Lexical contrast maintenance and the organization of sublexical contrast systems. Language and Cognition. 4(4), 319–355.
Werker, J. F., & Tees, R. C. (1984). Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant behavior and development, 7(1), 49-63. Williams, D. M., & Lind, S. E. (2013). Comorbidity and diagnosis of developmental disorders. In Current issues in developmental disorders, 19-45. Zoccolotti, P., & Friedmann, N. (2010). From dyslexia to dyslexias, from dysgraphia to dysgraphias, from a cause to causes: A look at current research on developmental dyslexia and dysgraphia. Cortex, 46(10), 1211-1215. Zwitserlood, P. (1989). The locus of the effects of sentential-semantic context in spoken-word processing. Cognition, 32(1), 25–64. Zlatin, M. A., & Koenigsknecht, R. A. (1975). Development of the voicing contrast: Perception of stop consonants. Journal of Speech, Language, and Hearing Research, 18(3), 541-553.
APPENDICES 7. Appendices
Appendix 1 Reading test:
Table 7.1 Orthographic productive rules used
Rules Rule 1
giu = /dʒu/
sce = /ʃe/
gn = /ɲ/
gli = /ʎi/
Table 7.2 Examples of words used in DDE.
Conditions Examples Real words, highly concrete and frequent:
i.e. vino (wine), bambino (child), letto (bed)
Real words, highly concrete and i.e. insetto (bug), cero (wax), margine infrequent: (edge) Real words, highly abstract and frequent: i.e. pace (peace), ragione (reason), successo (success) Real words, highly abstract and infrequent: i.e. dominio (domination), sciopero (strike), simbolo (symbol) Nonwords, shallow and short: i.e. fosto, prisi, tonca Nonwords, shallow and long:
i.e. locostato, tacipaca
gnoba, pronounced pronounced cɔdʒu
Table 7.3 Coloured Progressive Matrices standardised scores.
Id s1s1 s2s1 s3s1 s4s1 s5s1 s6s1 s7s1 s8s1 s9s1 s10s1 s11s1 s12s1 s13s1 s14s1 s15s1 s16s1 s17s1 s18s1 s19s1 s20s1 s21s1 s22s1 s23s1 s24s1 s25s1 s26s1 s27s1 s28s1 s29s1 s30s1 s31s1 s32s1 s33s1 s34s1
Age ~ 9;0 9;3 8;10 9;3 9;9 9;2 8;10 9;2 8;11 9;0 8;7 10;0 9;6 9;1 9;9 ~ 8;6 8;2 8;3 9;2 9;11 9;3 8;4 8;4 8;5 8;6 8;3 9;0 8;6 8;10 8;9 8;7 8;3
Score 34 29 30 31 30 29 26 19 34 21 31 28 22 25 31 32 25 34 26 26 32 29 30 27 26 28 29 33 32 30 21 27 25 23
Standardised score ~ 108 100 115 100 100 90 70 120 75 105 100 75 85 105 110 ~ 130 90 90 110 100 100 95 90 100 105 125 110 110 75 95 85 80
Table 7.4 Specifically designed stimuli: nonwords.
tra:kata pla:kata pra:kata kla:kata kra:kata
dra:kata bla:kata bra:kata gla:kata gra:kata
traka:ta plaka:ta praka:ta klaka:ta kraka:ta katra:ta kapla:ta kapra:ta kakla:ta kakra:ta ka:trata ka:plata ka:prata ka:klata ka:krata
draka:ta blaka:ta braka:ta glaka:ta graka:ta kadra:ta kabla:ta kabra:ta kagla:ta kagra:ta ka:drata ka:blata ka:brata ka:glata ka:grata
Table 7.5 Real words used in experiment 4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lexical bud slight mat side kit tact loud sort fate eight plate court rate right great grout mate node bait pride
Lexical buzz slice mass size kiss tax louse sauce face ace place course race rice grace grouse mace nose bass prize
Morphological asked called cared chewed formed helped hoped joined killed learned liked lived looked moved died dropped played proved saved seemed
Morphological asks calls cares chews forms helps hopes joins kills learns likes lives looks moves dies drops plays proves saves seems
Table 7.6 Nonwords used in experiment 6.
Stem number 1 2 3 4 5
Morpho 1 vɪld 3 vɛld 5 væld 7 vɔld 9 vʌld
Morpho 2 vɪlz 4 vɛlz 6 vælz 8 vɔlz 10 vʌlz
Non-morpho 41 vɪlt 43 vɛlt 45 vælt 47 vɔlt 49 vʌlt
Non-morpho 42 vɪls 44 vɛls 46 væls 48 vɔls 50 vʌls
Control 81 vɪlb 83 vɛlb 85 vælb 87 vɔlb 89 vʌlb
Control 82 vɪlm 84 vɛlm 86 vælm 88 vɔlm 90 vʌlm
6 7 8 9 10
11 nɪld 13 naɪld 15 næld 17 nɔld 19 nʌld
12 nɪlz 14 naɪlz 16 nælz 18 nɔlz 20 nʌlz
51 nɪlt 53 naɪlt 55 nælt 57 nɔlt 59 nʌlt
52 nɪls 54 naɪls 56 næls 58 nɔls 60 nʌls
91 nɪlb 93 naɪlb 95 nælb 97 nɔlb 99 nʌlb
92 nɪlm 94 naɪlm 96 nælm 98 nɔlm 100 nʌlm
11 12 13 14 15
21 θɪld 23 θaɪld 25 θæld 27 θɔld 29 θʌld
22 θɪlz 24 θaɪlz 26 θælz 28 θɔlz 30 θʌlz
61 θɪlt 63 θaɪlt 65 θælt 67 θɔlt 69 θʌlt
62 θɪls 64 θaɪls 66 θæls 68 θɔls 70 θʌls
101 θɪlb 103 θaɪlb 105 θælb 107 θɔlb 109 θʌlb
102 θɪlm 104 θaɪlm 106 θælm 108 θɔlm 110 θʌlm
16 17 18 19 20
31 dʒɑld 32 dʒɑlz 71 dʒɑlt 72 dʒɑls 111 dʒɑlb 112 dʒɑlm 33 dʒaɪld 34 dʒaɪlz 73 dʒaɪlt 74 dʒaɪls 113 dʒaɪlb 114 dʒaɪlm 35 dʒæld 36 dʒælz 75 dʒælt 76 dʒæls 115 dʒælb 116 dʒælm 37 dʒɔld 38 dʒɔlz 77 dʒɔlt 78 dʒɔls 117 dʒɔlb 118 dʒɔlm 39 dʒʌld 40 dʒʌlz 79 dʒʌlt 80 dʒʌls 119 dʒʌlb 120 dʒʌlm Note: Vowels used: ɪ aɪ æ ɔ ʌ (ɛ ɑ) last block used ɑ instead of ɪ, first block used ɛ instead of aɪ, starting consonants: v n θ dʒ.
Phonotactic probabilities calculated with the Vitevich and Luce (2004) software. In tables from Table 7.7 to Table 7.11words are transcribed in a form that is readable by the software. Detail of the transcription can be found in the appendix of Vitevich and Luce (2004). For each word, the first raw represents positional segment frequency, the second raw represent biphone frequency. None of these measures equals zero, indicating that all words in this task are phonotactically legal.
Table 7.7 Block 1. vIlz
.0224 .0962 .0737 .0121
.0224 .0729 .0737 .0121
.0224 .0794 .0737 .0121
.0224 .0165 .0737 .0121
.0039 .0090 .0002
.0035 .0087 .0002
.0025 .0086 .0002
.0002 .0035 .0002
SUM (+1) 1.2044 1.0130 vIld
.0224 .0962 .0737 .0403
.0224 .0729 .0737 .0403
.0224 .0794 .0737 .0403
.0224 .0165 .0737 .0403
.0039 .0090 .0040
.0035 .0087 .0040
.0025 .0086 .0040
.0002 .0035 .0040
SUM (+1) 1.2326 1.0168 vIlt
.0224 .0962 .0737 .0894
.0224 .0729 .0737 .0894
.0224 .0794 .0737 .0894
.0224 .0165 .0737 .0894
.0039 .0090 .0039
.0035 .0087 .0039
.0025 .0086 .0039
.0002 .0035 .0039
SUM (+1) 1.2817 1.0167 vIls
.0224 .0962 .0737 .0501
.0224 .0729 .0737 .0501
.0224 .0794 .0737 .0501
.0224 .0165 .0737 .0501
.0039 .0090 .0018
.0035 .0087 .0018
.0025 .0086 .0018
.0002 .0035 .0018
SUM (+1) 1.2424 1.0146 vIlb
.0224 .0962 .0737 .0179
.0224 .0729 .0737 .0179
.0224 .0794 .0737 .0179
.0224 .0165 .0737 .0179
.0039 .0090 .0005
.0035 .0087 .0005
.0025 .0086 .0005
.0002 .0035 .0005
SUM (+1) 1.2103 1.0134 vIlm
.0224 .0962 .0737 .0295
.0224 .0729 .0737 .0295
.0224 .0794 .0737 .0295
.0224 .0165 .0737 .0295
.0039 .0090 .0010
.0035 .0087 .0010
.0025 .0086 .0010
.0002 .0035 .0010
SUM (+1) 1.2218 1.0139
Table 7.8 Block 2. v^lz
.0224 .0392 .0737 .0121
.0238 .0962 .0737 .0121
.0238 .0343 .0737 .0121
.0238 .0794 .0737 .0121
.0007 .0046 .0002
.0019 .0090 .0002
.0015 .0026 .0002
.0019 .0086 .0002
SUM (+1) 1.1474 1.0055 v^ld
.0224 .0392 .0737 .0403
.0238 .0962 .0737 .0403
.0238 .0343 .0737 .0403
.0238 .0794 .0737 .0403
.0007 .0046 .0040
.0019 .0090 .0040
.0015 .0026 .0040
.0019 .0086 .0040
SUM (+1) 1.1756 1.0093 v^lt
.0224 .0392 .0737 .0894
.0238 .0962 .0737 .0894
.0238 .0343 .0737 .0894
.0238 .0794 .0737 .0894
.0007 .0046 .0039
.0019 .0090 .0039
.0015 .0026 .0039
.0019 .0086 .0039
SUM (+1) 1.2247 1.0092 v^ls
.0224 .0392 .0737 .0501
.0238 .0962 .0737 .0501
.0238 .0343 .0737 .0501
.0238 .0794 .0737 .0501
.0007 .0046 .0018
.0019 .0090 .0018
.0015 .0026 .0018
.0019 .0086 .0018
SUM (+1) 1.1854 1.0071 v^lb
.0224 .0392 .0737 .0179
.0238 .0962 .0737 .0179
.0238 .0343 .0737 .0179
.0238 .0794 .0737 .0179
.0007 .0046 .0005
.0019 .0090 .0005
.0015 .0026 .0005
.0019 .0086 .0005
SUM (+1) 1.1532 1.0059 v^lm
.0224 .0392 .0737 .0295
.0238 .0962 .0737 .0295
.0238 .0343 .0737 .0295
.0238 .0794 .0737 .0295
.0007 .0046 .0010
.0019 .0090 .0010
.0015 .0026 .0010
.0019 .0086 .0010
SUM (+1) 1.1648 1.0064
Table 7.9 Block 3. nclz
.0238 .0165 .0737 .0121
.0238 .0392 .0737 .0121
.0068 .0962 .0737 .0121
.0068 .0343 .0737 .0121
.0003 .0035 .0002
.0014 .0046 .0002
.0011 .0090 .0002
.0002 .0026 .0002
SUM (+1) 1.1261 1.0040 ncld
.0238 .0165 .0737 .0403
.0238 .0392 .0737 .0403
.0068 .0962 .0737 .0403
.0068 .0343 .0737 .0403
.0003 .0035 .0040
.0014 .0046 .0040
.0011 .0090 .0040
.0002 .0026 .0040
SUM (+1) 1.1543 1.0077 nclt
.0238 .0165 .0737 .0894
.0238 .0392 .0737 .0894
.0068 .0962 .0737 .0894
.0068 .0343 .0737 .0894
.0003 .0035 .0039
.0014 .0046 .0039
.0011 .0090 .0039
.0002 .0026 .0039
SUM (+1) 1.2034 1.0076 ncls
.0238 .0165 .0737 .0501
.0238 .0392 .0737 .0501
.0068 .0962 .0737 .0501
.0068 .0343 .0737 .0501
.0003 .0035 .0018
.0014 .0046 .0018
.0011 .0090 .0018
.0002 .0026 .0018
SUM (+1) 1.1641 1.0055 nclb
.0238 .0165 .0737 .0179
.0238 .0392 .0737 .0179
.0068 .0962 .0737 .0179
.0068 .0343 .0737 .0179
.0003 .0035 .0005
.0014 .0046 .0005
.0011 .0090 .0005
.0002 .0026 .0005
SUM (+1) 1.1319 1.0043 nclm
.0238 .0165 .0737 .0295
.0238 .0392 .0737 .0295
.0068 .0962 .0737 .0295
.0068 .0343 .0737 .0295
.0003 .0035 .0010
.0014 .0046 .0010
.0011 .0090 .0010
.0002 .0026 .0010
SUM (+1) 1.1435 1.0048
Table 7.10 Block 4. [email protected]
.0068 .0794 .0737 .0121
.0068 .0165 .0737 .0121
.0068 .0392 .0737 .0121
.0138 .0605 .0737 .0121
.0002 .0086 .0002
.0002 .0035 .0002
.0004 .0046 .0002
.0011 .0059 .0002
SUM (+1) 1.1720 1.0090 [email protected]
.0068 .0794 .0737 .0403
.0068 .0165 .0737 .0403
.0068 .0392 .0737 .0403
.0138 .0605 .0737 .0403
.0002 .0086 .0040
.0002 .0035 .0040
.0004 .0046 .0040
.0011 .0059 .0040
SUM (+1) 1.2001 1.0128 [email protected]
.0068 .0794 .0737 .0894
.0068 .0165 .0737 .0894
.0068 .0392 .0737 .0894
.0138 .0605 .0737 .0894
.0002 .0086 .0039
.0002 .0035 .0039
.0004 .0046 .0039
.0011 .0059 .0039
SUM (+1) 1.2493 1.0127 [email protected]
.0068 .0794 .0737 .0501
.0068 .0165 .0737 .0501
.0068 .0392 .0737 .0501
.0138 .0605 .0737 .0501
.0002 .0086 .0018
.0002 .0035 .0018
.0004 .0046 .0018
.0011 .0059 .0018
SUM (+1) 1.2100 1.0106 [email protected]
.0068 .0794 .0737 .0179
.0068 .0165 .0737 .0179
.0068 .0392 .0737 .0179
.0138 .0605 .0737 .0179
.0002 .0086 .0005
.0002 .0035 .0005
.0004 .0046 .0005
.0011 .0059 .0005
SUM (+1) 1.1778 1.0094 [email protected]
.0068 .0794 .0737 .0295
.0068 .0165 .0737 .0295
.0068 .0392 .0737 .0295
.0138 .0605 .0737 .0295
.0002 .0086 .0010
.0002 .0035 .0010
.0004 .0046 .0010
.0011 .0059 .0010
SUM (+1) 1.1894 1.0098
Table 7.11 Block 5. JYlz
.0138 .0343 .0737 .0121
.0138 .0794 .0737 .0121
.0138 .0165 .0737 .0121
.0138 .0392 .0737 .0121
.0004 .0026 .0002
.0011 .0086 .0002
.0002 .0035 .0002
.0014 .0046 .0002
SUM (+1) 1.1338 1.0032 JYld
.0138 .0343 .0737 .0403
.0138 .0794 .0737 .0403
.0138 .0165 .0737 .0403
.0138 .0392 .0737 .0403
.0004 .0026 .0040
.0011 .0086 .0040
.0002 .0035 .0040
.0014 .0046 .0040
SUM (+1) 1.1620 1.0070 JYlt
.0138 .0343 .0737 .0894
.0138 .0794 .0737 .0894
.0138 .0165 .0737 .0894
.0138 .0392 .0737 .0894
.0004 .0026 .0039
.0011 .0086 .0039
.0002 .0035 .0039
.0014 .0046 .0039
SUM (+1) 1.2111 1.0069 JYls
.0138 .0343 .0737 .0501
.0138 .0794 .0737 .0501
.0138 .0165 .0737 .0501
.0138 .0392 .0737 .0501
.0004 .0026 .0018
.0011 .0086 .0018
.0002 .0035 .0018
.0014 .0046 .0018
SUM (+1) 1.1718 1.0048 JYlb
.0138 .0343 .0737 .0179
.0138 .0794 .0737 .0179
.0138 .0165 .0737 .0179
.0138 .0392 .0737 .0179
.0004 .0026 .0005
.0011 .0086 .0005
.0002 .0035 .0005
.0014 .0046 .0005
SUM (+1) 1.1397 1.0036 JYlm
.0138 .0343 .0737 .0295
.0138 .0794 .0737 .0295
.0138 .0165 .0737 .0295
.0138 .0392 .0737 .0295
.0004 .0026 .0010
.0011 .0086 .0010
.0002 .0035 .0010
.0014 .0046 .0010
SUM (+1) 1.1512 1.0040
Appendix 3 Professor Harald Baayen at the University of Tuebingen ran a few simulations using his connectionist model with our experimental paradigms (including the stimuli used in this project). For a detailed description of the functioning of his connectionist model see Baayen, Shaoul,Willits and Ramscar (2015). Below, a short summary and a short commentary on the outputs obtained with his software are reported. Introduction The stimuli used in our experiments were used by Baayen to run a simulation of our tasks (experiments IV and V) with his model. First of all, it should be noted that he claims that his predictions are in line with our findings. We found what he would expect us to find, even if the model he is using is substantially different from the theoretical framework we are in. Description of Baayen’s model Before going in the detail of his prediction, a brief description of Baayen´s model (2015) will be given: Most language processing models assume words to be stored as units. Speech is, however, continuous (or, at least, pauses do not correspond to word boundaries), and all models thus have to explain how speech segmentation takes place. Baayen et al. (2015) suggests that words may not be stored as units and speech segmentation may actually not take place. According to his model, communication does not take place by segmenting speech and combining the meaning of words. According to his model, speech is mapped on a long term network trained on full sentences, and thus perceiving means discriminating experiences mapped in the network.
Perception operates through phonological items (sentences are treated as sequences of phonological items only, not as sequences of words). Online processing corresponds to discriminating what is perceived from the rest of the conceivable experiences. According to his model, experiences (i.e. meaning) and sounds are not separate, but part of the same level or analysis (although, he specifies that the experiences and sounds do belong, or form, different levels of representaion). Instead of relying on the concept of words, he uses a very specific notion of lexeme, later renamed LEXOME. From his paper: Lexemes are not form units, nor are they semantic units, but rather they represent the points of contrast that both form and meaning serve to mediate in lexical systems. In his paper, he shows how his model is able to predict performance obtained in other studies, for instance, performance in the classic study from Saffran et al (1996). Over time, the model analyses chunks of phonological items and activates lexemes generated by previous experiences. Again, speech is not decomposed in words, but processed in a continuous fashion. Lexemes that are most active during processing correspond to the bit of experience (i.e. meaning) selected from and contrasted to all the possible conceivable meanings. Simulation with our stimuli Now the simulation with our stimuli will be presented: Experiment 1: in this experiment we compared RTs in the discrimination of elements belonging to lexical and morphosyntactic minimal pairs. We found that the second type of discrimination takes more time (in the thesis, this is referred to as Experiment IV).
Baayen shows that his model makes the same prediction. It is interesting to note that the model learns comparing outcomes to stored lexemes (i.e. experienced lexemes – experience values are calculated using information extracted from corpora). Lexemes are divided between lexical and grammatical: for instance there is a grammatical lexeme for PAST and a lexical lexeme for the verb PLAY. From what I understood he compared differences in activation of the two elements in the pair, and found that only contrasting elements in a morphosyntactic pair generates a significant difference. Elements in morphosyntactic pairs are more different, according to his model, and he explains why: For what concern lexical lexemes there is no difference in the two conditions, but for what concerns grammatical lexemes, morphosyntactic pairs generated a larger contrast, since the two elements in the pair activate different grammatical lexemes.
Experiment 2: in this experiment we elicited MMN with lexical and morphosyntactic pairs, and obtained a larger negativity in the second condition (in the thesis, this is referred to as Experiment V). Again, Baayen shows that his model makes a similar prediction. In the lexical condition standard stimuli (i.e. “side”) do not activate the grammatical lexeme PAST, but, he says, the deviant stimuli (i.e. “size”) activate the grammatical lexeme PRESENT. This information, he claims, clashes with the drop in activation of the lexical lexeme. In the morphosyntactic condition there is no clash, because the drop in activation of the lexical lexeme corresponds to the drop in activation of the grammatical lexeme
PAST (because this lexeme is very active with the standard stimuli (i.e. “cared”)), see Figure 7.1.
Figure 7.1 Activation of the pairs cared/cares and side/size. Note: Blue arrows indicate the change in activation for the stems, and the red arrows indicate the change in activation for tenses
Insights from the model It is interesting to notice that, even proposing something completely different from classic linguistic models, Professor Baayen did not get rid of the idea of grammatical information operating somehow separately from the rest of the linguistic information. Professor Baayen did not get rid of the idea of grammatical lexemes, and it is the activation of these lexemes that predicts his finding with our stimuli. When comparing his approach to our approach there may be terminological issues and also some major differences in the idea of stripping but the idea of a separate grammatical (or at least morphological) system is maintained in both approaches.
Potential issue His predictions on the MMN task may seem a bit puzzling. My first concern was about the activation of morphological lexemes with lexical items. Specifically, it may be surprising to see that “size” activates PRESENT more than “side” activates PAST. One second concern could rise about the interpretation of the graph. Professor Baayen claims that the absence of MMN in the lexical condition is due to the activation clash discussed above. In his words “my hypothesis is that under such conflicting changes in activation, no MMN is generated”. This is however not completely in line with what we know about MMN: several studies show that the bigger the conflict, the bigger the MMN (as explained in this review: Näätänen et al., 2007). However, this contradiction could be solved if we observe his prediction from a different perspective: lexical lexemes are not playing a role. Grammatical lexemes are generating the MMN. If we make this assumption, the highest amount of conflicting information is in the morphosyntactic condition, and his model would nicely predict the brain correlates we found. Conclusion Baayen’s predictions are in line with our findings, and this is substantially due to the fact that his model incorporates the idea of grammatical lexemes. It is extremely interesting that two very different approaches end up needing the concept of grammatical information and that in both approaches this information operates, to a certain extent, autonomously from lexical lexemes/items.
Appendix 4 Table 7.12 Information on children with a clinical condition.
SLI + dyslexia
SLI + dyslexia
language impairment associated with moderate, bilateral, sensorineural hearing loss language impairment associated with fluctuating hearing loss due to infections language impairment associated with speech, language and communication difficulties language impairment associated with word finding and motor coordination problems deficit in phonological short-term memory and complex syntactic structures processing problems in vocabulary comprehension, phonological processing and reading language impairment associated with phonological short-term memory disorder language production deficit, particularly lexical retrieval and
CNRep raw score 23/40
structure generation word finding difficulties, but solid syntactic competence and phonological processing language disorder affecting primarily syntax and morphosyntax linguistic deficit particularly evident in discourse comprehension (event recall) past history of hearing difficulties and reading difficulties expressive language delay
specific language difficulties
language impairment associated particularly with phonological difficulties language impairment associated with phonological and lexical retrieval disorder, dyslexia language expressive impairment due to left middle cerebral artery infarction (stroke) language impairment affecting primarily phonology and
dyslexia + SLI
dyslexia + SLI
morphology specific language impairment affecting primarily morphology, dyslexia language impairment and literacy difficulties
dyslexia + SLI
dyslexia + SLI
specific reading disorder (magnocellular dyslexia)
dyslexia + SLI
dyslexia + SLI
Clinical profiles: J. has a moderate, bilateral, sensorineural hearing loss. at the date of testing he was 8;11. He was assessed with the Token test, Test of word finding and CNRep and the Grammaticality Judgmenet task. His major area of difficulty seems to be phonological processing. Secondly, he seems to present difficulties in word finding. D. has a history of fluctuating hearing loss associated with repeated infections. at the time of assessment she was 6;0. She was assessed with the Reynell receptive scale, the CNREP, CELF, FROG. D. presents with phoneme discrimination difficulties, phoneme articulation difficulties and grammatical markers problems. She was diagnosed as (at risk for) dyslexia + SLI. E. has significant speech, language and communication difficulties. at time of assessment she was aged 10;4. She was assessed with TROG and CNREP and RAVENS. M. has difficulties with word finding and motor coordination. 7;4. he was assessed with TROG, CELF, CNREP. He presents severe language grammatical and phonological problems, being in the lower 10% in all tasks J. has a deficit in phonological short term memory and a deficit in certain complex syntactic structures. At time of testing he was 9;1. He was assessed with trog, celf, CNREP and PHAB. The diagnosis is of language impairment.
T. has difficulties with vocabulary comprehension, with phonological processing and reading. At time of testing he was 12;9. He was tested with the british picture vocabulary scale BPVS II, the CNREP and TROG. He performed in the low 10 percentile in all tasks but Trog, where he performed at average level. W. was 14;2 at the time of testing. She presents with a phonological short term memory disoder. She was assessed with CNREP, TROG, CELF, she performed below average in all tasks. E. has a language production deficit, particularly for what concerns lexical retrieval and structure generation. At time of testing she was 14;6. She was assessed with TOWK, CNREP, FROG. She shows a phonological and lexical deficit in production as well as difficulties with complex grammatical structures. S. has word finding difficulties, but relative solid syntactic competence and phonological processing. At time of testing she was 5;3. She was assessed with CNREP, EPVT, BPVS and LARSP. L. presents with a language disorder affecting primarily syntax. She only produces simple sentences and fails to comprehend complex structures such as wh questions. at time of assessment she was 7;4. She was assessed with CNREP, TROG, BPVS, LASRSP, PHAB G. has a linguistic deficit particularly evident in discourse comprehension (event recall). At time of testing he was 10;7. He was assessed with CNREP, TOKEN, ACE, ERRNI, LARSP, FROG. CNRep and LARSP were in the normal range. D. presents with a past history of hearing difficulties and with reading difficulties. At the time of testing he was 8;4. He was assessed with CNREP, and PHAB, presenting expected performance. (The therapy recommended a dyslexia assessment and suggests that past hearing difficulties may play an effect in phoneme grapheme association). C. has an expressive language delay. When assessed he was 6;8. He was assessed on CNREP, and BAS. Phonological performance is as expected. S. presents with specific language difficulties. At the time of assessment she was 10;00. She was assessed with CNrep, TROG, and ACE and PHAB. All scores are below average for her age.
O. presents with specific phonological difficulties. At the time of testing he was 7;8. He was assessed with CNREp, trog, BPVS, ACE. He performed below average on CNRep and in the lower average of TROG. C. presents with a phonological and lexical retrieval disorder. At the time of testing he was 10;2. He was assessed with CNREp, BPVS, TWF, CELF. He performed below average in CNREP and TWF. S. had a left middle cerebral artery infarction (stroke) secondary to a left internal carotid dissection. He presents with a language expressive impairment. He was assessed with Cnrep, TWF-II, and CELF 4 and TROG. Only TROG performance was in normal range. At the time of testing he was 9;4. S. has a linguistic impairment concerning phonology and morphology. At the time of assessment she was 10;5. She was assessed with CNrep, TROG, TEGI and CELF, showing poor performance in all tasks. L. presents with Specific Language Impairment due to syntactic difficulties. She was assessed with TROG, CNREP and PHAB. at time of assessment she was 14;3. She performed below average on TROG. T. was referred from the DST for literacy difficulties. At the time of assessment he was 7;5. His performance on assessments is as expected for his age, except on CNREP. He was assessed on TROG, CNREP, and PHAB. R. has a previous diagnosis of specific reading disorder. At the time of assessment he was 9;6. He was assessed with CNREP, PHAB, and BPVS. His performance is as expected for his age. Data suggest presence of magnocellular dyslexia. A. The child has a previous diagnosis of dyslexia from the DST. At the time of testing her age was 9;1. She was assessed with TROG, BPVS and CNREP. In TROG only 18 blocks were passed.