Mental representation of morphologically complex

Mental representation of morphologically complex1 words and lexical access JUAN SEGUI and M AR l A- LU ISA ZU BIZARR ETA

Abstract We present in this article a certain number of linguistic and psycholinguistic arguments supporting the hypothesis whereby derived lexical items are listed in the lexicon as autonomous but related lexical items. Some experimental data obtained in French concerning the recognition of derivational affixed words reveal a difference in processing between prefixed and suffixed words. This difference in processing is attributed to the linear organization of these words and to the left-to-right nature of the perceptual parser. It is suggested that only suffixed words are accessed via the lexical entry for the root. In sentence context, accessing the root means that the end part of suffixed words is treated as a unit. We propose that such a procedure of affix detection can be very useful by way of facilitating syntactic parsing and lexical-syntactic integration in on-line language processing. This hypothesis provides a plausible explanation for why syntactically relevant affixes tend to appear at the end of words, Introduction

In this article we will present and discuss some experimental results that address the question of how morphologically complex words (more precisely, prefixed and suffixed words) are accessed in word recognition. We shall also discuss another closely related question concerning how morphologically complex words are represented in the mental lexicon. It seems clear that the answer to the first question relies heavily on the answer to the second. Two radically different views can be taken concerning the representation of affixed words: 1. The combination of an affix and a root constitutes a lexical entry. Like any other lexical item, it encodes its surface properties: i.e. it encodes the phonological and syntactic output-result of combining the affix and the root. Linguistics 23 (1985), 759-774

0024-3949/85/0023-0759 $2.00 © Mouton Publishers

Brought to you by | Université Paris Descartes - Interuniversita Authenticated Download Date | 3/1/16 8:25 AM

760 /. Segui and M.-L. Zubizarreta 2. The combination of an affix and a root does not constitute a lexical entry, which encodes its syntactic and phonological surface properties. There are several variants of this hypothesis. One of the most popular in the current studies of morphology and lexical access was put forward by Taft and Forster (1975). Within this view, only roots constitute lexical entries. Each lexical entry contains a list of the affixes with which it combines. These two points of view concerning the representation of affixed words in the mental lexicon make different predictions concerning the manner in which this type of word is accessed. The second hypothesis clearly entails that morphological decomposition must take place prior to lexical access in the case of prefixed words. This is so because it is the root (and not the combination aff + root) which constitutes the lexical entry. Lexical access must therefore take place via the root. In order to access the root in a prefixed word, the prefix must be previously stripped. This means that prefixes must be listed or encoded in a peripheral access file. Whether there is or is not suffix stripping before lexical access within this theory depends on whether lexical recognition is conceived of as a strictly left-toright procedure or not. If this is the case, then the root will be accessed before the suffix recognition procedure is applied. According to this perspective, morphological decomposition in the case of suffixed words will take place 'after' lexical access (unlike the case of prefixed words). Conversely, in a system which is not strictly left-to-right, the lexical recognition device can look up the end of the word 'before' lexical access, accomplishing suffix stripping before lexical access. This would mean that the peripheral access file will have a list not only of prefixes but also of suffixes. On the other hand, the first hypothesis concerning the representation of affixed words does not entail any kind of morphological decomposition before lexical access. Experimental data concerning lexical access of prefixed and suffixed words in French, which will be discussed in section 2, support this view of lexical access. If morphological decomposition prior to lexical access need not be assumed, then no morphological information need be encoded in the peripheral access file. According to this view, all morphological information can be confined to the central lexicon, where words in general are encoded. The hypothesis that derived lexical items are listed in the lexicon has often been criticized as putting an unnecessary load on memory. It should be noted, however, that if morphologically derived words are listed in the lexicon as autonomous but related lexical items, there is no reason to believe that they are costly in terms of memory. If morphologically derived words are entered in the lexicon as members of a family of


Mental representation and lexical access 761 morphologically related lexical items, their presence in the family can be considered as aiding, rather than hindering, memory. Before we present the experimental data, we will examine more precisely how affixed words are formed and represented in the lexicon and we will provide some general linguistic motivation for this view. The notion of morphological family will also be made explicit, and readers are warned that we will not discuss the distinction between derivational and inflectional morphology (nor the two types of derivational affix in English). The general view presented here will probably have to be refined when these finer distinctions are taken into consideration. Linguistic considerations

An affix defines what we can call an affix frame. An affix frame encodes the morphological, syntactic, and phonological properties of an affix. Let us consider an example, the affix -able. It combines with transitive verbal roots to produce an adjective. Moreover, -able affects the argument structure of the root to which it is attached, namely it turns a transitive verb into an intransitive, as shown by the minimal pair: χ reads books/books are readable. The exact nature of this process does not concern us here. For our present purpose we can simply assume that -able carries the following specification: Delete Subject, Object— > Subject. The affix frame defined by -able is as in (1): (1)

A: Delete Subj, Obj— > Subj

V ( + trans)

able

We can construct the following plausible hypothesis concerning the acquisition of affixed words: when the language learner acquires a new affix, he/she 'enters' into the lexicon the affix frame defined by the new affix. This affix frame automatically scans the lexicon. Every item that satisfies the selectional requirements defined by the affix frame is inserted as a sister of the affix in the frame. The properties of the derived word are then determined. The features of the affix take precedence over the features of the root in determining the properties of derived words. This property is already encoded in the affix frame: as shown in example (1) the mother node of the frame carries the features of the affix. But if the root is specified for some property for which the affix is not, then this property of the root becomes relevant in determining the properties of the whole.


762 J. Segui and M.-L. Zubizarreta (Compare the percolation conventions for word formation proposed by Lieber (1980), which are a slightly modified version of those proposed by Williams 1981.) So, for example, in the case of -able, the affix determines the syntactic category of the derived word (as specified by the affix frame which it defines). On the other hand, -able has no argument structure. The root with which it combines does have an argument structure, which becomes the argument structure of the derived word, modified in the manner specified by the -able affix frame. The phonological properties of the derived word are also determined, including its accentual pattern. Brackets are then erased. The derived word is entered in the lexicon as a new lexical item (with all its surface syntactic, morphological, and phonological specifications) and can then serve as the basis for further affixation. THUS, NEW LEXICAL ENTRIES ARE AUTOMATICALLY GENERATED, WITHOUT HAVING TO AWAIT ENCOUNTERS IN THE LINGUISTIC ENVIRONMENT.

This view of how words are formed is basically the one advocated in Pesetsky (1979), who suggested applying SPE's bracket erasure convention during the process of word formation. As pointed out by Pesetsky, this view of word formation makes an interesting prediction: affixation is sensitive only to the surface properties of the item with which it combines. The prediction seems to be correct. Indeed, this property of word formation has often been attributed to the cyclical nature of the phonological, syntactic, and morphological rules that apply within the domain of a lexical item. The principle of cyclicity can be defined as follows: (2) In order to determine the surface properties of a lexical item B which has the following morphological structure: (..Y..(A)..X..), it is necessary and sufficient to take into account the surface properties of A, in addition to the relevant properties of X and Y. (2) implies that when an affix is added to a morphologically complex lexical item A, it is the properties of the item A as a whole and not the properties of its constituent parts that are relevant in computing the properties of the newly derived item B. If lexical items derived via affixation are encoded in the lexicon as described above, and if word formation takes as input only lexical items that are encoded in the lexicon, then the effect of cyclicity follows directly. Cyclicity need not be postulated as a principle of grammar, it is derived from the properties of word formation. Below we briefly illustrate the evidence in favor of cyclicity with a few examples. The word formation process known as affixation seems to be cyclical in the sense defined above. Most affixes can only combine with lexical items of a given categorical type: verbal, nominal, or adjectival. As noted by Allen (1978), an affix is sensitive only to the categorical features of the


Mental representation and lexical access 763 root to which it attaches, whether derived or underived. Therefore, we do not find affixes that only attach, for example, to deverbal nouns or deadjectival verbs. Morphophonological processes also seem to be sensitive to the properties of the derived form as a whole but not to the properties of morphemes which compose it. Dell and Selkirk (1978) noticed that morphemes in French can be divided into two classes according to whether they can undergo and trigger backing of a low vowel. They distinguish these two classes with the feature + / — L. If a + L affix is attached to a + L root with low vowel, the vowel undergoes backing. If the root is — L, the vowel does not undergo backing. Now consider a case in which a lexical item is formed by attaching a -L affix onto a +L root. Given that the affix takes precedence over its sister in determining the properties of the affix, being the head of the derived lexical item, the derived root wil be - L. If a + L affix is in turn attached to the derived — L root, the backing rule will not apply, in conformity with (2). Another example of a cyclical phonological rule is the rule which assigns primary stress to a word in English (cf. Halle and Vergnaud i.p.). This rule interacts in an interesting way with the rule of secondary stress assignment. The latter applies after the former and may be formulated as in (3): (3) Stress every other syllable counting from right to left and starting with a syllable that bears main stress. The stress pattern of the following word illustrates (2): Äppalächicola ('indicates primary stress, ' indicates subsidiary stress). An interesting pattern arises when several stresses have already been assigned as a consequence of the cyclical application of the primary stress rule. Essentially, the assignment of secondary stress by rule (3) will respect the preliminary pattern that results from the cyclical combination of primary stresses. Consider, for example, the morphologically complex word instrumentality, which has primary stress on the syllable preceding -ity. Its secondary stress pattern is not * instrumentality, as it would be if it were a monomorphemic word. Its stress pattern is rather instrumentality. This is due to the stress pattern of instrumental, which has main stress on ment, and from which instrumentality is derived. Clearly, it is the stress pattern of the derived lexical item instrumental, and not that of the basic morpheme instrument (which has no stress on ment) which is relevant in determining the stress pattern of instrumentality, in conformity with (2). Certain affixes modify in certain specified ways the argument structure of the lexical item to which they attach. These morphosyntactic processes are also cyclical in nature, as Muysken (1981) has shown on the basis of

Brought to you by | Université Paris Descartes - Interuniversitai Authenticated Download Date | 3/1/16 8:25 AM

764 /. Segui and M.-L. Zubizar reta data from Quechua dialects of central Peru. Both reflexivization and causativization are morphosyntactic rules in this language. The reflexive affix ku attaches to a verb and referentially binds the object to the subject. The causative affix chi adds a new subject argument to the verb to which it attaches and the old subject becomes an object argument. In the dialects that Muysken considers, two orders of affixation are possible: (a) v-chi-ku and (b) v-ku-chi. In conformity with principle (1), the process triggered by Affl must apply before the process triggered by Aff2 in [[v-affl] aff2]. Therefore in (a) the reflexive ku can only bind the Subject of the verb derived by causativization (i.e. it cannot bind the Subject of K). In (b) ku must bind the Subject of the embedded verb (i.e. it cannot bind the Subject of the verb derived by causativization). Baker (1984) discusses many comparable examples which clearly show the cyclical nature of morphosyntactic processes. It is perhaps worthwhile noting that some of the phenomena described above have been attributed in the literature to a principle other than cyclicity. Indeed, Williams (1981) has proposed that the strictly local nature of the selectional restrictions of affixes is due to the atom condition, which states the following: a restriction on the attachment of an affix to y can only refer to features realized on y. But the principle of cyclicity (which, as we have already said, need not be stated in the grammar because it can be derived from the way that words are formed in the lexicon) is of a much more general nature. The atom condition cannot account for the properties of English stress described above, the reason being that a stress pattern cannot be described in terms of features on a node. As suggested above, each morphologically derived form constitutes a lexical entry of its own, but crucially it is not an isolated lexical entry. It is linked to all those lexical items to which it is morphologically related. What defines a morphological family is the common root that all the members of the family share. This common root, whether bound or free, also constitutes a lexical entry, which perhaps functions as the 'head' of the morphological family as suggested by Cutler (1983). At this point we do not have evidence as to the shape of the internal structure of a morphological family. A possible hypothesis is that lexical items within a morphological family are directionally related. That is, the linking relations that a lexical item entertains with other members of the family could reflect its derivational history. Adopting this hypothesis would be useful in order to avoid infinite recursivity within the lexicon. This could be achieved by disallowing loops. Words like rereread will then not be encoded in the lexicon. To illustrate this, Figure 1 shows a subpart of a morphological family.


Mental representation and lexical access 765

A — > jincollablel—ι y\

in

lindecollablelj

A A

V

V — > |collabilis(er)| >Hcollable

able

|decollabilis(er)| is |incollabilis(er) | [indecollabilis(er)|

Figure 1. A subpart of the colle 'glue' family

Strictly speaking, the members of a family are the set of lexical items which are related by the affix frames. The frames themselves are not members of any family. They belong to the class of grammatical vocabulary, and they 'cut across' different families. Note that while the derived lexical items themselves do not contain morphological information, their morphological structure can be inferred from the structure of the family. In the next section we will present and discuss some experimental data concerning the recognition of derivational affixed words in French. The data reveal an interesting difference in processing between prefixed and suffixed words. We think that the view that we outlined above concerning the representation of morphologically complex words allows us to account for these differences while maintaining a uniform representation for both types of affixed lexical items. Our conclusion will be that in accessing both prefixed and suffixed words no morphological decomposition is accomplished at a preaccess level. We will suggest that the difference found between prefixed and suffixed words can be attributed to the left-to-right nature of the parser. Given this hypothesis, prefixed words will be accessed in the same manner that monomorphemic words are. We suggest that suffixed words, on the other hand, are accessed via the lexical entry of the root, assuming Cutler's (1983) suggestion that the root constitutes the 'head' of the morphological family. Accessing the root activates all the members of the family and a matching procedure then applies between the end part of the input signal and the end part of all the members of the family in order to determine which is the right candidate. Since affixed words are encoded in the lexicon with their surface properties, no computation needs to be accomplished on-line in order to determine the properties of the whole. This conclusion, of course, does not exclude the possibility that when accessing morphologically


766 /. Segui and M.-L. Zubizarreta complex words of very low frequency we might resort to a different strategy: a computational strategy that exploits the affix frames rather than, or as well as, the lexical representation. Nor does it exclude the possibility that at a postaccess level the parser might make use in certain cases of information contained in the affix frames in order to achieve higher efficiency at the level of syntactic parsing. Some general speculative suggestions will be made in the conclusion concerning the latter point. In the next section we ask to what extent morphological structure is exploited in word recognition. 2.

Psycholinguistic considerations

Numerous experiments in psycholinguistics provide support for the hypothesis that members of the same family, defined by the base root, are strongly related in the mental lexicon. In particular, it has been demonstrated on several occasions that the prior presentation of a word corresponding to the root or to one of the members of a morphological family facilitates generally the subsequent identification of the other members, for example by Murrell and Morton (1974); Stanners et al. (1979); Henderson et al. (1984); Jarvella and Meijers (1983). These results appear to indicate that accessing one of the words of the family 'activates' the other members' representations in the internal lexicon, at least for regular related forms (Kempley and Morton 1982). As previously noted, the fact of postulating the existence of a level of representation where morphologically complex words are stored as global units does not prevent us from proposing that the lexicon contains information concerning their internal morphological structure. This information should be available from the internal organization of the morphological family. In fact, morphological structure information MUST be incorporated in some way in the lexical representation. Different types of psychological evidence indicate that such information has a significant role in production or perception performance. All the same, we know very little about the precise nature of the processing of this type of word in lexical access. Traditionally, this problem has been studied by stating it in terms of processing complexity. The research carried out in this perspective has attempted to relate the morphological complexity of words to their processing complexity. However, as noted by Cutler (1983), who has critically examined this area of research, no serious empirical evidence exists at present which supports this hypothetical relation. A second hypothesis of complexity has been derived from Taft and


Mental representation and lexical access 767 Forster's (1975) theory of morphological decomposition. As pointed out above, according to this theory, morphologically complex words are decomposed into their constituent morphemes (root plus affixes) during perceptual processing in order for lexical access to take place. In this case, morphological decomposition occurs 'prior' to lexical access, which is conducted on the basis of the root. Therefore, the access code for a morphologically complex word will be the root (more precisely the BOSS of the root). Within this theory, an adequate segmentation of the word relies on the detection and stripping of segments which are likely to be affixes. One might expect, therefore, that pseudoaffixed words would induce the incorrect application of this stripping procedure. The processing system would 'assimilate' the pseudoaffixed word to an affixed word, thus complicating access to the correct lexical representation. Consequently, the processing of pseudoaffixed words must be generally more complex than for matched affixed words. The empirical evidence provided on this point is not conclusive and depends on the morphological type considered. While in certain experimental conditions the processing of pseudoprefixed words proved to be more complex than that of truly prefixed words (Taft 1979, 1981), other experiments have found no difference (Rubin et al. 1979; Henderson et al. 1984). Moreover, no difference in processing is observed between pseudosuffixed and truly suffixed words (Taft 1979; Manelis and Tharp 1977; Henderson et al. 1984). Below we will discuss two recent lexical access experiments conducted in French, using both prefixed and suffixed words. They show a clear processing asymmetry between these two types of affixes. Cole et al. (n.d.) have conducted an experiment using an experimental procedure analogous to the one used by Manelis and Tharp (1977). In their experiment, Manelis and Tharp presented pairs of strings of letters to subjects who were instructed to respond 'yes' if BOTH of the letter strings were English words. Some of the experimental pairs consisted of words with the same morphological characteristics (homogeneous pairs). These were pairs of derived suffixed words (e.g. PRINTER-DRIFTER) and pairs of pseudosuffixed words (e.g. SLANDER-BLISTER). Other experimental pairs were composed of words with different morphological characteristics (mixed pairs), a suffixed word and a pseudosuffixed word (e.g. PRINTER-SLANDER). Their results show that reaction times are faster for the homogeneous pairs than for the mixed pairs. Manelis and Tharp interpret this difference between the homogeneous and mixed pairs in terms of a hypothetical semantic priming transfer. The homogeneous pairs were processed faster 'because' they shared a common semantic element, the meaning of the suffix. Such interpretation does not explain,


768 /. Segui and M.-L. Zubizarreta however, the observed differences between the mixed pairs and the homogeneous pairs composed of two pseudosuffixed words. It is not clear what semantic features could be shared by the two pseudosuffixed words. However, these results show that this experimental paradigm can measure a difference in processing due to the type of morphological relation holding between the two members of a pair. In order to study more precisely the role of the internal morphological structure of a context word in the processing of the test word, Cole et al. used a sequential presentation of the elements of the pair with a lexical decision task on the second member of the pair (i.e. the test word). The first member of the pair (the context word) had merely to be read by the subject. The experimental material thus consisted of pairs of items (context and test) characterized by two properties: 1. the category of the test word (prefix, pseudoprefix, suffix, pseudosuffix); 2. whether or not the context word shared the same morphological structure as the test word (homogeneous and mixed pairs). The four groups of test words were matched with each other on frequency and length. Each test word was associated with two separate context words so that it was part of both a mixed and a homogeneous pair. These complementary pairs were associated to two different experimental blocks. Each block consisted of 100 pairs of items: 40 experimental pairs (5 χ 8), 20 fillers containing two monomorphemic words, and 40 pairs in which the second item was a nonword. Nonwords were derived from words by changing one letter; half of them were derived from affixed and pseudoaffixed words. The morphologically complex words used were derivational affixes with a free root. They were semantically 'transparent' words. Table 1 gives examples for each of the eight types of experimental pairs. The context word was presented visually on a computer screen for 800 msec, and then replaced by the test word. Two groups of 20 subjects participated in the experiment in order to counterbalance the order of presentation of the two blocks. Table 2 shows the mean RTs (in msec) for the eight types of experimental pairs. Table 1. Examples of experimental pairs Test word

Homogeneous pair

Mixed pair

Prefix Pseudoprefix Suffix Pseudosuffix

preface — prenom requin — recette rosier — potier vantard — bagnard

prefet — prenom rechute — recette entier — potier guepard — bagnard


Mental representation and lexical access 769 Table 2.

Mean RTs for experimental pairs

Test word

Homogeneous pair

Mixed pair


661 646 627 636

658 652 664 649

Globally the results indicated that homogeneous pairs are responded to faster than mixed pairs (642 msec, and 656 msec, respectively), but the difference varies greatly according to the type of test word considered. An analysis of variance shows that the interaction between the category of the test word (prefix, pseudoprefix, suffix, pseudosuffix) and the type of context (mixed or homogeneous) is significant (F (3-114) = 3.22 p < 0.05). In fact, specific comparisons indicated that it is only for pairs with a suffix test word that the difference between mixed and homogeneous pairs is significant (F (1-38) = 19.9 p < 0.001). A t test shows that this difference is significant with items as a random factor (t (9) = 3.58 p < 0.01). These results indicate an asymmetry in the processing of prefixed and suffixed words. While the processing time of suffixed words varied significantly according to the morphological characteristics of the word context, this was not the case for prefixed words. Such a result is difficult to reconcile with Manelis and Tharp's semantic priming hypothesis. According to this hypothesis we should have observed a comparable priming effect for the two types of word. Taft and Forster's prelexical decomposition hypothesis predicts that the pseudoaffixed words would be more sensitive to context than the truly affixed words. The observed results do not confirm this prediction. Before suggesting an interpretation of the above results we shall look very briefly at some recent experimental data that confirms the differential sensitivity of prefixed and suffixed words to context. In a second experiment, Cole et al. studied the access of prefixed and suffixed words by using a set induction technique. In this experiment the subjects made lexical decisions on prefixed or suffixed test words which occurred at the end of a list of 10 items. For half of the test words the context list was composed of words having the same morphemic structure as the test word, and the remaining lists were composed of nonaffixed words. We wish to stress here that, in the context lists with words sharing the same morphemic structure as the test word, the words did not have a particular affix in common but simply shared the property of being prefixed or suffixed words. The question asked was whether the response


770 /. Segui and M.-L. Zubizarreta times to affixed test words would be shorter when they were preceded by 10 other words of the same morphemic structure than when the context list words were not affixed. The results indicate that this is true for the suffixed test words (659 vs 696 msec) but no difference between the two types of context list was observed for prefixed items (711 vs 718 msec). In other words, the access of prefixed words does not appear to be affected by the morphological properties of the context items. In order to explain the prefix-suffix asymmetry we propose the following hypothesis: 1. Processing takes place essentially from left to right. 2. As proposed in section 1, roots and the forms derived from them constitute lexical entries organized in the form of a 'family'. The root — bound or free — constitutes the 'head' of a family. 3. As soon as the parser hits a form which corresponds to the head of the family, namely the root, the morphological family in question is accessed and a matching procedure is applied in order to select the right candidate. Processing from left to right requires the subject to access the root first in the case of suffixed words. Accessing the root morpheme entails accessing a morphological family of words. Subsequently, a new matching of the terminal properties of the signal with one of the members of this family takes place. We may hypothesize that it is this matching procedure which is speeded up in the case where words of comparable morphological structure root + affix are presented successively. It goes without saying that nothing comparable will arise with pseudosuffixed words. As for prefixed words, it is our contention that no matching procedure will be called for in lexical acces, given that the affix in this case precedes the root: i.e. there is no root accessing that precedes access of the whole word. In order to explain the results of the experiments presented here we hypothesized that different access procedures operate for suffixed and prefixed words. This difference in access procedure is linked to the linear organization of the components of these words, root + affix and affix + root, and to the fact that the perceptual analysis of words proceeds from left to right. It remains to be determined whether this interpretation of the asymmetry in processing of prefixed and suffixed words may allow a better understanding of why in certain languages like French and English the affixes that carry relevant syntactic information are more often suffixes than prefixes. In particular, suffixes, but not prefixes, determine generally the syntactic category of the derived word. Taking this into account, in sentence processing, suffix detection is required in order to assure the efficiency of the syntactic parsing. It is our contention that this


Mental representation and lexical access 111 Table 3. Substitution errors


Original word

Error

repris regard verdure torture

ropris rogard verduze tortuze

efficiency can be obtained if the access to the root triggers a suffix detection process; i.e. if the last part of the word can be treated by the system as a morphemic unit. Some recent results obtained by Pavard (1983) suggested that this hypothesis may be correct. Using an error detection task Pavard studied the probability of detecting an error in the first or last syllable of words in a text as a function of the morphological value of this syllable. The errors to be detected are orthographic substitution errors and they concern the first syllable of prefixed or pseudoprefixed words and the last syllable of suffixed or pseudosuffixed words. Table 3 presents an example of each type of substitution error. Pavard shows in a comprehension-correction task that while the detection of errors in an initial syllable is independent of its morphemic value (prefixed or pseudoprefixed), a clear difference is observed in the probability of detecting errors in the final syllable depending on its suffixed or pseudosuffixed status. In a very rapid reading situation (50 to 60 letters/sec) detection of errors that appear in the last syllable is better for suffixed words than for pseudosuffixed words. In another experiment the author shows that the probability of error detection is negatively correlated with the depth of processing (inferred from the number of recalled sentences) for the pseudoaffixed words. This parameter is irrelevant in the case of suffixed words: the detection of errors in suffixed words is stable and independent of the depth of processing of the text. This result is observed in the case of suffixed words with a productive root as well as in the case of suffixed words with a nonproductive root. In a third experiment Pavard shows that THE PROBABILITY OF DETECTING AN ERROR IN A SYLLABLE WHICH IS A SUFFIX IS NOT AT ALL AFFECTED BY THE CONTEXTUAL PREDICTABILITY OF THIS SUFFIX.

To account for these results Pavard suggests that in the case of suffixed words, it is the preliminary identification of the root morpheme that triggers the process of morphological analysis. This interpretation is compatible with our hypothesis that in sentence perception the access to


772 /. Segui and M.-L. Zubizarreta the root triggers a suffix detection process. In this case the end part of the suffixed word is treated by the subject as a perceptual unit. The behavioral consequences of this processing may seem apparently contradictory. In some cases the processing of the last syllable as a unit may lead the subject to 'ignore' the particular elements of that unit while in other cases this type of processing may lead the subject to 'focus' on the elements of that unit. These behavioral consequences can be determined by the particular demands of the experimental situation (see Drewnoski and Healy 1980; Smith and Sterling 1982). Anyway, the important point to note is the special status attributed by the system to the end part of the suffixed words. While for isolated suffixed words the access to the root triggers a matching procedure in order to detect the particular member of the morphological family corresponding to the stimulus, in sentence context the end part of the suffixed words is automatically treated as a morphemic unit. Such a procedure of affix detection can be very useful in order to facilitate syntactic parsing and lexical-syntactic integration in on-line sentence processing. To illustrate this speculative proposal, consider agreement features, passive morphology (-ed), derivational affixes like -able and -ize. The agreement features of a verb must be checked to agree with the subject. Formally speaking, we can say that these features, carried by an inflectional suffix, must percolate up to the verb phrase: i.e. the sister node of the noun phrase subject. Agreement is established by coindexing the noun phrase and the verb phrase. If the suffix is identified separately from the root and if furthermore the closed class items are in some sense easier to integrate to the sentence than the open class items, then the agreement subject-verb phrase relation will be sanctioned before the verb is attached to the verb phrase. That is, the syntactic subject-verb agreement relation will be established before integrating the verb with previously parsed material. Passive morphology, -able, -ize, etc., is also rich in syntactic information. Processing these suffixes separately allows the determination of a lot of the syntactic structure even before integration of the verb with the remainder of the sentence. Passive morphology will give information about the status of the sentence subject. It will flag the following information: interpret the surface subject as the logical object of the sentence, -able will indicate that the subject must be interpreted as the theme of the sentence and that no object is expected; -ize will indicate that the subject must be interpreted as the agent of the sentence and that an object must be expected; and so on. Consequently, taking the suffix as a unit makes it possible for the language processor to efficiently exploit the syntactically relevant infor-


Mental representation and lexical access 773 mation carried by these morphemes in on-line sentence parsing. Although the details of this proposal are yet to be fully worked out, it offers a plausible explanation for why syntactically relevant affixes tend to appear in suffix position. Received 20 February 1985 Revised version received 10 August 1985

Universite Rene Descartes

Note Preparation of this paper was supported by A.T.P. 5207 C.N.R.S. to J. Segui. M.-L. Zubizarreta acknowledges support from the Fondation Fyssen, France. Correspondence address: Laboratoire de Psychologie Experimentale, Universite Rene-Descartes, 28 rue Serpente, 75006 Paris, France.

References Allen, M. A. (1978). Morphological investigations. Unpublished Ph.D. thesis, University of Connecticut. Baker, M. (1984). The mirror principle and morphosyntactic explanation. Unpublished manuscript. Cambridge, Mass.: MIT. Cole, P., Beauvillain, C., Pavard, B., Segui, J., and Zubizarreta, M. L. (n.d.). L'acces au lexique pour les mots morphologiquement complexes. Rapport A.T.P. CNRS (in preparation). Cutler, A. (1983). Lexical complexity and sentence processing. In G. B. Flores d'Arcais and R. J. Jarvella (eds.), The Process of Language Understanding. New York: Wiley. Dell, F., and Selkirk, E. (1978). On a morphologically governed vowel alternation in French. In J. S. Keyser (ed.), Recent Transformational Studies in Generative Grammar. Linguistic Inquiry Monograph. Cambridge, Mass. Drewnowski, A., and Healy, A. F. (1980). Missing -ing in reading: letter detection errors on word endings. Journal of Verbal Learning and Verbal Behavior 19, 247-262. Halle, M., and Vergnaud, J. R. (i.p.). Aspects of the Theory of Stress. Cambridge, Mass.: MIT Press. Henderson, L., Wallis, J., and Knight, J. (1984). Morphemic structure and lexical access. In H. Bouma and D. G. Bouwhuis (eds.), Attention and Performance X. Hillsdale, N.J.: Erlbaum. Jarvella, R. J., and Meijers, G. (1983). Recognizing morphemes in spoken words: some evidence for a stem-organized mental lexicon. In G. B. Flores d'Arcais and R. J. Jarvella (eds.), The Process of Language Understanding. Chichester. Wiley. Kempley, S. T., and Morton, J. (1982). The effects of priming with regular and irregularly related words in auditory word recognition. British Journal of Psychology 73,441-454. Lieber, R. (1980). The organisation of the lexicon. Unpublished Ph.D. dissertation, MIT (distributed by IULC). Manelis, L., and Tharp, D. A. (1977). The processing of affixed words. Memory and Cognition 5, 690-695.


774 /. Segui and M.-L. Zubizarreta Murreil, G. A., and Morton, J. (1974). Word recognition and morphemic structure. Journal of Experimental Psychology 102,963-968. Muysken, P. (1981). Quechua causatives and logical form. In A. Belletti, L. Brandi, and L. Rizzi (eds.), Theory ofMarkedness in Generative Grammar. Proceedings of the 1979 Glow Conference, Pisa. Pavard, B. (1983). Traitement perceptif des mots affixes: mise en evidence d'un contröle cognitif. L'Annee Psychologique 83,443-464. Pesetsky, D. (1979). Russian morphology and lexical theory. Unpublished manuscript. Cambridge, Mass.: MIT. Rubin, G. S., Becker, C. A., and Freeman, R. H. (1979). Morphological structure and its effect on visual word recognition. Journal of Verbal Learning and Verbal Behavior 18, 757-767. Smith, P. T., and Sterling, C. M. (1982). Factors affecting the perceived morphemic structure of written words. Journal of Verbal Learning and Verbal Behavior 21, 704-721. Stanners, R. F., Neiser, J. J., Hernon, W. P., and Hall, R. (1979). Memory representation for morphologically related words. Journal of Verbal Learning and Verbal Behavior 18, 399-412. —, Neiser, J. J., and Painton, S. (1979). Memory representation for prefixed words. Journal of Verbal Learning and Verbal Behavior 18, 733-743. Taft, M. (1979). Recognition of affixed words and the word frequency effect. Memory and Cognition 7, 263-272. —(1981). Prefix stripping revisited. Journal of Verbal Learning and Verbal Behavior 20, 289-297. —, and Forster, K. I. (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behavior 14, 638-647. Williams, E. (1981). On the notion of 'lexically related' and 'head of a word'. Linguistic Inquiry 12, 245-274.
