Character recognition among English- speaking L2 readers of Japanese

International Journal of Applied Linguistics

䉬

Vol. 21

䉬

No. 3

䉬

2011

Character recognition among Englishspeaking L2 readers of Japanese Etsuko Toyoda and Tim McNamara The University of Melbourne

The present study investigated the development of semantic processing skills in character recognition among English-speaking L2 readers of Japanese with different levels of knowledge of kanji (the morphographic script used in the Japanese writing system) by using a timed semantic processing task (a task involving the comparison of kanji). By analysing the results of the task (correct response rates and reaction times), the study described the changes in semantic processing skills at the different stages of knowledge of kanji. The overall findings of the study suggest that 1) L2 readers with different levels of target script knowledge approach the recognition of characters differently, and that 2) L2 semantic processing skills approximate those of L1 readers with increased L2 script knowledge. Keywords: Japanese, kanji, L2, semantic, recognition

Introduction Learning to read is a complex task, not only for children, but also for adult L2 learners. One of the most important abilities required for reading comprehension is efficient word recognition (e.g. Koda 1992; Grabe 2004), the process of retrieving orthographic, phonological and semantic information for the word (Harris and Coltheart 1986). Printed words may be recognized as © 2011 Blackwell Publishing Ltd

384

䉬

Etsuko Toyoda and Tim McNamara

whole words, but, when less familiar, or even unfamiliar, readers may rely on the recognition of components providing phonological, grammatical or semantic information (phonemes, morphemes, stems, affixes, etc.). Languages differ in the degree to which this information is transparent to the reader. Clearly, the more systematic the language, the more help is provided to the novice reader. Furthermore, when learning an L2, the larger the differences (such as orthographic differences, differences of the regularity of letter-sound correspondence) between L1 and L2, the more obstacles there are to overcome. For readers of alphabetic languages, in addition to the orthographic unfamiliarity, character-based languages, especially Japanese, present complex recognition tasks, since the systems through which this information is provided are not simple. Further, the systematic features, while present, are not always reliable guides to sound, meaning and hence recognition (note however that – in contrast to Japanese – the simplified characters used in Mainland China – but not Taiwan – tend to be relatively phonologically transparent). The ability to process and evaluate the complex information contained within characters as an aid to word recognition is the focus of this paper. Young L1 readers transit several developmental stages as they learn to process L1 script efficiently. Subsequently, when a literate adult L1 reader who has developed high level integrated word recognition skills in the L1 learns an L2 with orthographic features that are very different from those of the L1, it is reasonable to speculate as to whether such readers go through a similar developmental sequence. The resemblance between L1 and L2 morphosyntactic development which forms the basis for interlanguage theory (Selinker 1972), offers a precedent for speculation as to the possibility that there might be parallels in the development of the ability to read words in L1 and L2. In interlanguage theory, interlanguage forms are commonly held to be shaped not only by transfer from the L1 but by an independent process which resembles or (in its stronger form) recapitulates the process of L1 acquisition. Research on learning to read words in a second language does, in fact, suggest the existence of developmental processes that seem to reflect those found in L1 word recognition development. While there are crosslinguistic effects (resembling transfer in L2 morphosyntactic development), whereby readers’ L1 word recognition skills continue to affect L2 word recognition (Haynes and Carr 1990; Koda 1995; Akamatsu 2003), research suggests the existence of processes independent of such effects: readers gradually develop L2-specific word recognition skills through building knowledge of L2 orthographic features (Ke 1998; Koda 1999; Chikamatsu 2006). This is evident from the fact that L2 readers from the same orthographic background can be poor or good readers in the L2, and that differences in L1 orthographic experience do not always generate differences in © 2011 Blackwell Publishing Ltd

Character recognition among English-speaking L2 readers of Japanese

䉬

385

efficiency in reading in L2 (Koda 1999). It appears that L2 reading ability develops as L2 knowledge increases. Chikamatsu (2006) compared word recognition strategies of higher and lower proficiency L1 English readers of L2 Japanese, and found that L2 word recognition strategies are developmental and that L1 orthographic interference diminishes as learners gain L2 proficiency. This suggests that a developmental sequence will be observed regardless of L2 orthographic differences, the case of literate adult L1 readers learning an L2, with very different orthographic features, being no exception. Relevant studies of the existence of developmental sequences in word recognition in an L2 are scarce. The study by Chikamatsu (2006) focuses on L2 readers’ orthographic and phonological processing in word recognition; the present study looks at how L2 readers process semantic information in word recognition. The investigation of semantic processing is appropriate given that research suggests word recognition ability is closely related to the ability to use semantic-category information (Koda 2002). When asked to relate words, L1 children with poor reading comprehension are affected by concrete semantic relations that are associated in the real world (e.g,. shampoo-hair), while those with good reading comprehension are sensitive to more abstract semantic relations based on category coordinate knowledge (e.g. aeroplane-train) (Nation and Snowling 1999). L1 adult readers are therefore expected to have a well-developed ability to use semantic-category information. L2 readers’ semantic networks (which are supplemented by orthographic and phonological connections), have lower levels of linkage between words than those of L1 readers (Wilks and Meara 2002). However, it is plausible that L2 readers’ processing of the semantic information gradually approximates that of experienced L1 readers. The present study investigates the development of semantic processing skills in word recognition among English-speaking L2 readers of Japanese with different levels of knowledge of the Japanese writing system. In particular, it focuses on character recognition as a part of the word recognition process. As will be demonstrated below, semantic-category information is complex, and not always reliable. The development of the ability to recognize semantic information more or less explicitly available or implied in characters, and to evaluate its reliability, can be expected to be a gradual process, given this complexity. In the study, we investigate two research questions: 1. Do L2 readers with differing levels of knowledge of the script involved demonstrate differences in their ability to recognize semantic information contained within characters? 2. How does L2 readers’ processing of the semantic information compare with that of experienced L1 readers? © 2011 Blackwell Publishing Ltd

386

䉬


In this paper, these two questions will be investigated with reference to the ability of L1 English readers of L2 Japanese who have different levels of knowledge of kanji (the morphographic script used in the Japanese writing system), to recognize and evaluate the semantic information contained within the character. Kanji is one of the four types of script used in Japanese, and it has a pivotal role in Japanese sentences as most nouns, and stems of verbs and adjectives, are written in kanji. As a result, fast and accurate kanji character recognition is essential for successful reading comprehension in Japanese (Sayeg 1996). Although many processes are involved in reading, lower-level processing such as the recognition of characters has been found to be critical (Chikamatsu 2006; Koda 1992). The present paper therefore deals with L2 readers’ semantic processing skills in kanji recognition. When processing semantic information in kanji, readers may use two levels of information: one at the character level; and another at the subcharacter (component) level. The focus of the present paper is the processing of the component level semantic information. In order for readers who may be unfamiliar with Japanese scripts to follow the design of the present study, it is necessary to review briefly the orthographic features of kanji.

Features of Japanese morphographic script First, kanji can be grouped, roughly speaking, into four patterns according to their graphical configuration (Nomura 1984): non-separable, singlecomponent, simple kanji (e.g. ), and multiple-component or compound kanji which display three basic patterns of arrangement of the components: left-right (e.g. ), top-bottom (e.g. ), and others (e.g. ). The various patterns of arrangement are summarized in Figure 1. Although the exact number of single-component characters in Japanese appears never to have been formally estimated, the figure is likely to be low: for example, it is reported that in Chinese only about 10 per cent are single-component characters (Leck, Weekes and Chen 1995). The processing of compound characters is therefore of fundamental importance in word recognition in Japanese.

Figure 1. Graphical configuration © 2011 Blackwell Publishing Ltd


䉬

387

Figure 2. The bushu of kanji

In addition to having varying positional arrangements, the components of kanji display further functional organization. In compound characters, one of the components indicates a broad semantic category; these categories (and hence the components representing them), are called bushu. In kanji dictionaries, all kanji are classified under these semantic categories (Koda 2002). The bushu in a character often, but not always, indicates the semantic meaning of the character. This indeterminacy is crucial to this study, as the development of an ability to recognize the association between a bushu and the actual semantic meaning of a character is complex and gradual, reflecting the complexity and uncertainty of the degree of association involved. The second component of the kanji can be one of two things: a phonetic indicator in some cases and a merely non-functional component in other cases. The componential structure of kanji is illustrated in Figure 2. It is necessary to say a little more about the indeterminacy of the relationship between the bushu and the actual semantic meaning of a character, which causes difficulties for readers of Japanese and which forms the basis of this study. Although the bushu is usually a reliable semantic indicator in a character, the semantic relationships between a character and its bushu and between a bushu and the characters sharing it are not always consistent (Flores d’Arcais, Saito and Kawakami 1995), unless the generic meanings of characters and components are known. For example, as shown in ‘word’ and ‘story’ share a common bushu, ‘speech’, and are Figure 3, ‘avenge’ and semantically closely related (example 1). On the other hand, ‘visit’ also share the same bushu ‘speech’, but have little apparent semantic relation (example 3). A number of kanji containing this bushu ‘acknowledge’ and come between these two extreme cases, for example, ‘forgive’ (example 2). The indeterminacy of the relationship will be reflected in the judgements of L1 readers of Japanese. Two kinds of L1 reader judgement are used in this © 2011 Blackwell Publishing Ltd

388

䉬


Figure 3. The indeterminacy of bushu

study. First, L1 readers should be able to estimate the degree of association between the semantic category indicated by a bushu and the actual meaning of the character in which it is found, as in the above examples. Thus, we can expect that in general, L1 readers will broadly agree (with some variation), on the varying degrees of association, particularly where the association is present but modest (i.e. neither clearly present, as in example 1, nor clearly absent, as in example 3, but present to some degree as in example 2). Moreover, on a timed semantic judgement task where L1 readers have to make rapid and spontaneous evaluations of the degree of such relatedness, they will more readily agree that there is a relationship when that relationship is clear, than when it is not clear. Thus we can expect: (1) a high affirmative response rate with a fast reaction time for strongly related items; (2) a low affirmative response rate with a fast reaction time for strongly unrelated items; and (3) a medium affirmative rate with a slow reaction time for marginally related items. These kinds of L1 reader judgements are used in this study as benchmarks for evaluating the performance of learners of differing levels of knowledge of kanji script; the extent to which the different proficiency L2 reader groups approximate L1 reader behaviour is the basis for the comparison of the groups, and the basis for evidence of developmental stages in the mastery of this awareness among L2 readers. © 2011 Blackwell Publishing Ltd


䉬

389

Method Participants L2 readers A total of 77 L2 readers of Japanese participated in the study. They were graduate and undergraduate university students (aged 18–40) learning or having learnt Japanese. In all cases, English was the language in which they felt most comfortable both speaking and reading, regardless of their first language. Potential L2 participants were asked if they had had experience of learning Chinese characters prior to learning Japanese. Anyone who answered ‘yes’ to this question was excluded from the final participant group. All participants in the final group had seen characters on the street such as restaurant names, and had some basic knowledge about kanji and their constituent components through learning Japanese at a university in Australia. However, none had any specific training in analysing the internal features of kanji. Participants fell into one of three levels of kanji reading skill. Conventionally, learners of Japanese are classified into three categories of proficiency in reading kanji: beginner, intermediate and advanced. They were classified according to the results of an adaptive self-checklist kanji knowledge test. An adaptive self-checklist test is frequently used to assess vocabulary knowledge. Criteria for applying this sort of test are: (1) a sufficiently large sample of words; (2) limited test time; and (3) difficulty tailored to the testtaker (Nation 1993; Read 1993). The current situation met these criteria. The form of the test involves a dichotomous choice for each item: yes (= ‘I know the item’)/no (= ‘I don’t know the item’). Such tests are often referred to as yes/no tests. These tests have been subject to research over several decades, and their validity has been investigated in a number of studies (e.g. Mochida and Harrington 2006). The adaptive self-checklist kanji knowledge test was created using a kanji database that contained data from the standardized Japanese Proficiency Test ). The Database for the 1,945 Basic Japanese Kanji, 2nd edition ( (Tamaoka, Kirsner, Yanase, Miyaoka and Kawakami 2002) consists of 1,945 kanji, the difficulty of which was determined by the Japan Foundation and the Association of International Education in 1993. The current kanji knowledge test involved 423 of these kanji, organized into three lists drawn from the beginner (135 characters selected), intermediate (152 characters selected) and advanced (136 characters selected) levels;1 participants ticked any kanji that they knew. For efficiency of administration, and to adapt the test initially to the approximate level of each subject, participants were shown the lists one by one. They were shown the advanced list first, and asked to indicate by a tick for each item if they knew it or not; anyone ticking more than five kanji on this © 2011 Blackwell Publishing Ltd

390

䉬


list was not further required to look at the intermediate and beginner lists, and was categorized as an advanced reader. Those who ticked fewer than five kanji on the advanced kanji list were asked to look at the intermediate kanji list, and the same procedure was followed. Likewise, participants unable to tick more than five kanji on the intermediate kanji list were required to try the beginner kanji list. To establish the reliability of the test, an analysis of responses from subjects to items on each of the three lists using Rasch modelling was carried out, using the computer program Winsteps (Linacre 2009). The reliability (Cronbach’s alpha) of each test was high (beginners, 0.97; intermediate, 0.96; advanced, 0.99). Following Shillaw (1996, 1999), who used Rasch analysis with a yes/no vocabulary test, the fit of items and persons was examined to see if subjects had approached the task of responding to the test items in a serious and meaningful fashion, or were ticking randomly: high levels of misfit of items and persons would suggest that the task was not being taken seriously. The number of misfitting items and persons was low (Items: beginners, 3; intermediate, 6; advanced 5. Persons: beginners, 1; intermediate, 1; advanced 0), suggesting that the task had been taken seriously, and the characterization of subjects into beginner, intermediate and advanced was thus supported by the analysis. The final participant group consisted of 20 beginner, 40 intermediate, and 17 advanced readers. The length of time studying Japanese in a classroom situation varied within each group; 1–5 years for the beginner group, 3–7 for the intermediate group and 5–10 for the advanced group.

L1 readers In addition, two groups of L1 readers of Japanese who were also fluent in English (Japanese-dominant bilinguals), agreed to participate. Ages ranged between 18 and 40. One group (Group A, N = 45), participated in a preliminary study, investigating perceptions of the semantic characteristics of the materials to be used in the main study; the second group (Group B, N = 16), performed the same task (see the following section) as the L2 readers in the main study, and their performance was used as the basis for comparison with the three L2 reader groups. All L1 participants had completed at least nine years of compulsory education in Japan.

Materials A timed semantic processing task was prepared in order to investigate perceptions and evaluation of the semantic information contained within characters. Participants were asked to judge the semantic relationship between pairs of kanji (hereafter, the kanji-kanji task). During the selection of kanji, four types of stimulus properties that were reported to be likely to affect © 2011 Blackwell Publishing Ltd


䉬

391

Table 1. Sample kanji characters used in the kanji-kanji task

Types of items in the kanji-kanji task Pairs of kanji with a common bushu Pairs of kanji without a common bushu Fillers

Samples

FYI

1st kanji

2nd kanji

Common bushu

(face)

(head)

(above neck)

(hunt)

(steel)

–

(up)

(middle)

–

Note: The English translations in parentheses did not appear in the task. The right column is given for reference only.

processing speed (the number of required strokes for writing, frequency of occurrence in print materials, the number of kanji that share a common bushu and the number of kanji that share a common reading), were controlled.2 There were 120 pairs of kanji altogether: 40 pairs with a common bushu (common-bushu items); 40 pairs without a common bushu (no-common-bushu items); and 40 fillers. Note that the focus of the paper was on how readers judge the commonbushu items. The results of no-common-bushu have not been included in this paper due to space constraints, but are available in the larger work (Toyoda 2009). As fillers are distracters, they were not the target of analysis. The common-bushu items all had some potential semantic connection between the two kanji which was indicated in the semantic category indicated by the common bushu (see Table 1). Where the bushu was shared, the degree of semantic relationship between the characters varied from strong to weak. For (face) and (head) example, one can see a semantic relationship between in the sense that both ‘face’ and ‘head’ can be connected to the meaning of the , ‘neck and above’. Not all the semantic relationships were as clear; bushu sometimes, the link was a purely etymological one. For example, the (move) and (serve) through the shared bushu relationship between (power) may be difficult to perceive, unless one considers how things were moved, and work done, in earlier times. In the no-common-bushu items, the two kanji lacked a common bushu and were likely to be semantically unrelated. (hunt) and (steel) have (animal bushu) and (gold For example, bushu) respectively, and thus have no common bushu; they are also semantically unrelated. Thus, the task posed two demands: to recognize that a character shared a bushu with the other characters in the pair; and in such a case to determine the degree of shared meaning. The kanji-kanji task was programmed to automatically show the items on the computer monitor. As the task involved a large number of items, the positions of items might have affected the processing latencies of items. To © 2011 Blackwell Publishing Ltd

392

䉬


avoid any possible effect of a particular position of the item, the order in which items were displayed was randomized.

Procedure Given that determining the semantic relationships involved for the commonbushu items is a matter of judgement, it was important, as a preliminary, to elicit L1 reader judgements of the semantic characteristics of the stimulus items to be used in the study, which would then be used to classify the stimuli used in the study proper. To this end, 45 L1 readers of Japanese (L1 readers, Group A) were asked to rate the semantic relatedness of the common-bushu items in the kanji-kanji task (See Table 2). For the purpose of rating, a scale of 0–5 was provided next to each item; ‘0’ meaning ‘not related at all’ and ‘5’ meaning ‘very closely related’. The degree of semantic relatedness varies; the relationship is sometimes transparent, and sometimes less so, depending mainly on the inherent semantic information of each kanji and each bushu, but it may vary from one person to another depending on their social and personal experience. Nevertheless, L1 readers of a language should have relatively similar semantic networks. Five sets of common-bushu items emerged from this preliminary step, depending on the judged strength of semantic relationship (the mode rating values were used), which ranged from ‘strongly related’ to ‘strongly unrelated’ (Table 3). Following this preliminary process, which yielded sets of items of known semantic relatedness, the participants in the study proper (3 groups of L2 readers and one group of L1 readers (Group B)), were tested individually. They were instructed to judge, as quickly as possible, whether or not the two kanji were semantically related. Each of the 120 pairs was displayed sequentially on the computer monitor, one character at a time. Prior to the first kanji being displayed (for 500 ms), an eye-fixation mark (double circles) was displayed for 600 ms, and between the two kanji, a blank frame was inserted for 200 ms (see Figure 4). The second Table 2. Sample items used in the semantic relatedness rating survey

Types of items in the semantic relatedness rating survey 1st column (face) (snow)

2nd column (head) rain

5___4___3___2___1___0 5___4___3___2___1___0

Note: The English translations in parentheses did not appear in the task. © 2011 Blackwell Publishing Ltd


䉬

393

Table 3. Different strength of semantic relatedness (shown in mode values) rated by the L1 readers for the common-bushu items

Semantic relatedness

Number of items

Samples

Mode value of rating 5 (strongly related) 4 3 2 1 0 (not related)

Common-bushu items

7 9 9 7 8

(face) (chain) (plant) (sand) (treasure) (rise)

(head) (steel) (potato) (crush) (palace) (star)

Notes: The English translations of the kanji did not appear in the task. Small numbers of items sorting low on the scale led to the coalescing of scale steps 1 and 2.

Figure 4. The order of displaying slides

item was presented until a response was made, but the time for the response was limited to five seconds in order to encourage a quick decision. Before the start of the task, three practice items were provided. The participants were given feedback on the accuracy of their responses to the practice items, and were asked if they would like to move on, or redo the practice. The participants were required to respond as quickly as possible by pressing the ‘yes’ key for acceptance of semantic relatedness for each set of task items (kanji pairs), the ‘no’ key for rejection of relatedness or the ‘skip’ key if they could not make a decision. No feedback was given except for the first three practice items.

Analysis Given that participants’ judgement of the degree of semantic relationship of the common-bushu items of known relatedness is the focus of the current study, only the analysis of the responses to the common-bushu items will be reported. The preliminary study had defined sets of items having different degrees of semantic relatedness. The results for the kanji-kanji task were analysed in © 2011 Blackwell Publishing Ltd

394

䉬


order to determine how the participants in the three L2 reader groups and the comparison L1 reader group (Group B) responded to these sets of items. Would they similarly recognize the items as constituting distinct sets? For each set of items with a particular degree of semantic relatedness, the mean percentage of ‘yes’ responses and the mean reaction time were calculated for each individual in the group concerned; this yielded a mean for that group. Mean differences between the levels of semantic relatedness within each group were analysed using repeated measures ANOVA. A post hoc pairwise comparison of means was carried out to compare each pair of levels, with a Bonferroni correction applied in order to adjust significance levels. For response rates and reaction times, one-way ANOVA was used for all groups by levels of relatedness. For each variable, given the unequal numbers of participants among the groups, a test of homogeneity of variance was applied and the resulting Levene statistic was checked. Where a non-significant Levene test indicated equal variances, a Scheffe test was used for post hoc comparisons; where a significant Levene test indicated unequal variances, a Tamhane test was used.

Results Tables 4 and 5 show the mean response rates and the reaction times for all levels of semantic relatedness for all the groups. The response pattern of the L1 readers was as expected. The mean proportion of affirmative responses declined from a high of 95.5 per cent for the set of strongly related items to a mean of 16.4 per cent for the strongly unrelated items as the sets became less and less related. The result of repeatedANOVA showed that there was a significant difference between the Table 4. Affirmative Response Rates of All Groups (%)

Relatedness level N

0

1+2

3

4

5

L1

16

L2 Adv

17

L2 Int

40

L2 Beg

20

16.4 (11.8) 35.7 (26.7) 64.6 (29.4) 62.0 (29.0)

34.8 (26.1) 42.0 (18.2) 54.9 (25.7) 49.6 (25.7)

69.4 (20.5) 53.8 (24.4) 56.7 (30.9) 54.6 (29.9)

68.1 (16.7) 48.7 (24.3) 68.3 (26.4) 77.5 (23.6)

95.5 (8.6) 67.2 (18.1) 71.0 (24.7) 65.8 (26.1)

Note: 0 = Strongly unrelated, 1+2 = Mostly unrelated, 3 = Somewhat related, 4 = Mostly related, 5 = Strongly related; Figures in parentheses are SD. © 2011 Blackwell Publishing Ltd


䉬

395

Table 5. Reaction Times of All Groups (ms)

Relatedness level N

0

1+2

3

4

5

L1

16

L2 Adv

17

L2 Int

40

L2 Beg

20

1055 (259) 1581 (645) 1235 (557) 1512 (483)

1144 (393) 1679 (611) 1279 (551) 1581 (590)

1152 (313) 1576 (680) 1349 (584) 1604 (539)

1069 (287) 1551 (629) 1323 (644) 1524 612)

859 (229) 1340 (628) 1231 (601) 1624 (557)

Note: 0 = Strongly unrelated, 1+2 = Mostly unrelated, 3 = Somewhat related, 4 = Mostly related, 5 = Strongly related; Figures in parentheses are SD.

relatedness levels (F(4,60) = 75.43, p < 0.05). Different relatedness sets differed significantly from each other, except between some neighbouring sets (between ‘mostly related’ and ‘somewhat related’, as well as between ‘mostly unrelated’ and ‘strongly unrelated’). Regarding reaction times, there was also a significant difference between the relatedness levels (F(4,60) = 10.01, p < 0.05). The reaction time for the strongly related set (859 ms) was significantly shorter than that for the other sets. Responses to the strongly unrelated items were faster (1055 ms) than to weakly related items. However the differences were not significant, except between the two extremes. The response pattern of the advanced L2 readers showed the most similarity with the pattern of the L1 readers: there was a trend towards fewer affirmative responses as the degree of semantic relatedness declined. The difference between the advanced L2 readers and the L1 readers was that these readers distinguished less sharply between the most strongly related and the most strongly unrelated items than did the L1 readers. Nevertheless, the overall difference between the relatedness levels in the advanced L2 readers’ judgements was significant (F(4,64) = 10.55, p < 0.05). The participants’ affirmative response rate for the strongly related set (67.2%) was significantly higher than for the other sets (except the ‘somewhat related’ set). The affirmative response rate for the strongly unrelated set (35.7%) was significantly lower than the strongly related set and the somewhat related set. The results for the reaction time almost mirrored the results of the affirmative response rate. There was a significant difference between the relatedness levels (F(4,64) = 5.53, p < 0.05), and the strongly related set (1340 ms) was responded to significantly more quickly than were the other sets (except the mostly related set). The reaction time for the strongly unrelated set (1581 ms) was significantly faster than the strongly related set, but not the other sets. © 2011 Blackwell Publishing Ltd

396

䉬


Now let us examine the results of the less experienced L2 readers to see whether or not the disparity between their processing of the semantic information and that of the L1 readers is larger than that which we have seen between the advanced readers and L1 readers, By contrast, the response patterns of the intermediate L2 readers and beginner L2 readers did not show the same consistently descending curve as that for the L1 and the advanced L2 readers. The most notable results for these less-skilled readers were the relatively high affirmative response rates for the strongly unrelated set (64.6% for the intermediate readers and 62.0% for the beginner readers). These results all indicated that, unlike the more experienced readers, the less-experienced readers relied on component level information more than character level information. The characteristics of the response patterns of the intermediate and beginner L2 readers were best described in a lack of disparity between the responses for the strongly related set and for the strongly unrelated set. The intermediate L2 readers’ affirmative response rate for the strongly related (71%) set was not significantly higher than for the strongly unrelated set (64.6%) despite the fact that it was significantly higher than the mostly unrelated and somewhat related sets. The strongly unrelated set was not significantly different from the other relatedness levels. For reaction time as well, the difference between the relatedness levels was not significant (except where the somewhat related set was responded to significantly more slowly than the strongly unrelated set). The response pattern of the beginner L2 readers was mostly similar to that of the intermediate L2 readers. The beginner L2 readers’ affirmative response rates for the strongly related and strongly unrelated sets were both relatively high, and there was no significant difference between the two. Interestingly, the affirmative response rate for the mostly related set was significantly higher than the somewhat related and mostly unrelated sets, which could be affected by obvious differences in the visual features of the kanji used in the sets rather than by their semantic information. There was no significant difference in the reaction times between the relatedness levels. Overall, in respect to affirmative response rate, significant differences between the participant groups were found in all the relatedness levels except the somewhat related level. In particular, the results of the strongly related (F(3,89) = 6.82, p < 0.05) and strongly unrelated (F(3,89) = 15.47, p < 0.05) sets showed marked contrasts between the participant groups. For reaction time, there were significant differences between the participant groups for the strongly related (F(3,89) = 5.86, p < 0.05), mostly unrelated (F(3,89) = 3.99, p < 0.05) and strongly unrelated (F(3,89) = 4.05, p < 0.05) sets. For the strongly related set, the L1 readers demonstrated their welldeveloped character recognition ability. The L1 readers’ affirmative response rate (95.5%) was significantly higher than that of the L2 readers including that of the advanced L2 readers (67.2%), who showed a similar response pattern © 2011 Blackwell Publishing Ltd


䉬

397

(descending curve) to the L1 readers. In contrast, for reaction time, although the L1 readers (859 ms) gave affirmative responses more quickly than the L2 readers, the difference was only significant in relation to the beginner L2 readers (1624 ms). The quick reaction time of the L2 intermediate readers may be an indication of their reliance on the bushu. It is likely that they judged semantic relatedness by examining whether there was a common bushu between the two kanji, rather than by extracting and comparing the meanings of the two. For the strongly unrelated set, there was no significant difference between the L1 readers (16.4%) and advanced L2 readers (35.7%) in their affirmative response rates. Both groups mostly rejected the relations between the ‘strongly unrelated’ kanji pairs, and therefore their affirmative response rates were significantly lower than the other L2 reader groups whose affirmative response rates were higher than the chance level. These results clearly indicate that the advanced L2 readers executed character dominant semantic processing just as the L1 readers did. On the other hand, the reaction time of the L1 readers for the strongly unrelated set (1055 ms) was significantly shorter than that of the advanced L2 readers (1581 ms), which indicates that the advanced L2 readers required much more time for making judgements when there were inconsistencies between the semantic information displayed in the kanji and their bushu. On the contrary, there were no significant differences in reaction time between the L1 group and the intermediate L2 groups. The intermediate did not spend a long time in responding to the strongly unrelated set, which was an indication that they did not consider the discrepancy in the semantic information between the character and component levels. A significant difference was shown between the L1 readers and the L2 beginner readers probably due to the beginner readers’ slow identification of the kanji.

Discussion First, let us consider what the results suggest about character recognition in L1 readers, since their judgements serve as benchmarks for evaluating the performance of learners of differing levels of knowledge of kanji script. We expected: (1) a high affirmative response rate with a fast reaction time for strongly related items; (2) a low affirmative response rate with a fast reaction time for strongly unrelated items; and (3) a medium affirmative rate with a slow reaction time for marginally related items. The results were as expected except for the strongly unrelated items, which were found to involve a relatively long decision time. In order to account for this, we need to consider how L1 readers process the semantic information of characters. Do skilled readers’ ultimate judgements about the meaning of a character necessarily involve component level processing, or does a character tend to be recognized ‘all at once’, as a whole? On the one hand, if the L1 readers had © 2011 Blackwell Publishing Ltd

398

䉬


only relied on the information from the bushu, then they would have rated all the items as ‘semantically related’, which in fact they did not. However, the notion that L1 adult readers did not rely solely on the bushu does not necessarily mean that they did not process this information at all. In fact, there is evidence from the response time data that the L1 readers did seem to be affected by the semantic information conveyed by the bushu. Affirmative response times tended to be faster (and more uniformly so) in judging the more clearly related or clearly unrelated items than in judging vaguely related items. When semantic information conveyed by bushu is consistent with the meanings of kanji, as in the case of the most strongly related items, the L1 readers responded to them significantly more quickly than in any other case. The study by Yamada and Takashima (2001) may be of help in interpreting this phenomenon. These researchers presented phonetically-written (hiragana) words, one by one, to Japanese university students, and asked them to say aloud the name of the bushu of the kanji that corresponded to the phonetically-written words. The responses of the participants were quicker and more accurate when the bushu were semantically closely related to the target kanji, with slower responses and higher error rates observed where the meanings of the kanji and its bushu were not closely related. The researchers concluded that the inconsistency between the information from the whole kanji and from the bushu was likely to have been the cause of the slower responses (as in the present study), and higher error rates. This processing of component information does not necessarily involve conscious awareness. At the debriefing session held immediately after the task, only one of the L1 readers mentioned looking at the bushu. Instead, the majority of readers reported that they did not use this semantic information at all, but instead used the phonological information within the character to determine the meanings of the kanji. (Remember that both semantic and phonological information may be contained within a kanji, each offering a route to the determination of character meaning.) In general, these skilled readers identified and acknowledged the semantic relationship between pairs of highly related kanji, and rejected any relationship between pairs of non-related kanji, regardless of the bushu of the kanji. Although they seemed to have processed the semantic information at the bushu level, it is reasonable to hypothesize that information at the character level would have taken precedence over the information from the semantic component (bushu) level at the time of decision-making. Now let us consider the behaviour of the L2 readers. Generally, the results of the task showed that the advanced L2 readers had developed a semantic network which, at least partially, resembled that of the L1 readers. The higher percentage of affirmative responses to the more related items and the lower percentage to the less related items indicated that the semantic judgement of the advanced readers was not solely dependent on the bushu information, which they were likely to have processed, given their experience. © 2011 Blackwell Publishing Ltd


䉬

399

Despite these similarities, there were differences between the advanced L2 readers and the L1 readers. While the L1 readers clearly judged the strongly related items as related, the advanced L2 readers were more hesitant in saying that ‘yes, they were related’. The L1 readers’ affirmative response rate was significantly higher than that of the L2 readers, including the advanced L2 readers. On the other hand, the differences between the three L2 groups were not significant. This suggests that the semantic network of L2 readers is not strong enough to make decisions confidently. This was true even for advanced L2 readers who had been studying Japanese more than five years. Nevertheless, the advanced L2 readers responded significantly more accurately and quickly when the bushu were semantically closely related to the kanji (the strongly related items), then when the relationship was less clear, which indicated that these readers had considered semantic information from both character and sub-character levels. This point was also supported by the fact that the 95 per cent confidence interval for the strongly unrelated set was relatively long. When information from the kanji and their bushu were not consistent, the responses varied. At the briefing session after the task, some advanced L2 readers remarked that they looked at the bushu of the kanji, but hesitated to rely on them ‘because bushu were not always reliable’, thus indicating they were considering both character-level and component-level information with predominance of the former over the latter. Koda (2002) claims that what discriminates L1 readers from L2 readers is that L1 readers are aware of the validity of component information. The findings of the current study suggest that at least some very advanced L2 readers have attained a native-like processing ability in that respect. The responses of the intermediate and beginner L2 readers showed a pattern quite distinct from that of either the L1 readers or the advanced L2 readers. Their affirmative response rates for the strongly unrelated set were significantly higher than those of the L1 and advanced L2 readers. Moreover, these less-skilled L2 readers’ response speed was generally faster than that of the advanced readers, and was closer to that of the L1 readers. This was especially prominent in the case of the responses to the strongly unrelated set (one might have anticipated faster response times for the advanced readers, who were presumably more skilled than the intermediate readers; this was not in fact the case). The L1 readers’ reaction times to the most strongly related set was significantly quicker than that of the advanced L2 readers. However, there was no significant difference in reaction time between the L1 readers and the intermediate and beginner L2 readers. These results all suggest that these less-skilled readers relied heavily on the information at the bushu level, whereas the advanced readers considered (perhaps subconsciously) the information both from the characters and the bushu, resulting in slower response times. At the debriefing session after the task, the majority of the intermediate L2 readers explicitly reported that they relied mainly on bushu for their judgements. © 2011 Blackwell Publishing Ltd

400

䉬


The findings indicate that there were qualitative differences in the performances of participants according to the amount of their kanji knowledge. L1 readers’ responses showed the expected sensitivity to degrees of semantic relatedness; these judgements did not change even under the time pressure. Advanced L2 readers demonstrated semantic processing skills that were similar in certain important respects to those of the L1 readers although their responses were much slower than the L1 readers. Intermediate and beginner L2 readers seem to have relied heavily on the information conveyed by the semantic component. These findings suggest that semantic processing skills are developmental, and as knowledge of the L2 script increases the processing starts from an approach relying on component-level information, and eventually approximates L1 readers’ semantic processing in which information at the character level predominates. This predominantly character-level information processing is essential for word recognition, as it has been argued that higher-level processing cannot be performed without proficient lower-level processing (Haynes and Carr 1990). How do these developmental differences compare with what is known of the development of L1 reading skills in children? For L1 children, the development of word recognition skills shifts from partial to analytical and to more integrated processing. This is true both for alphabetic scripts (Harris and Coltheart 1986; Gough 1991), and for morphographic scripts, such as Chinese characters (Ho and Bryant 1997; Shu 2003). When learners are at the analytical processing stage, they often make errors due to overgeneralization of the rules that they have found (Ho and Bryant 1997). For L2 readers, probably because they are cognitively more developed than children, they seem to go straight into the transition period rather than starting from the partial processing period. It seems to be critical for readers to go through a transition period (where they rely on component-level information), before arriving at more sophisticated processing. Less-skilled readers test hypotheses that they have generated from their limited knowledge, and gradually with experience become more skilled in the continuous assessment and amendment of their hypotheses. Children learning to read in their L1 also go through an intermediate transition stage (e.g. Harris and Coltheart 1986; Ho and Bryant 1997; Shu 2003), regardless of different scripts. The present study suggests that L2 readers with an alphabetic background who learn to read morphographic script, also show a stage in their progression to more sophisticated processing. Given that literate L2 readers already have L1 knowledge prior to the study of L2, they may form hypotheses that are shaped by cross-linguistic effects. Concurrently however, L2 readers seem to make and test hypotheses that are part of an independent process of learning to read L2 script. The processes involved have a descriptive parallel in classical accounts of the development of ‘interlanguage’ in second language learning (e.g. Selinker 1972) which proposed a mix of cross-linguistic effects and processing stages which appear to be related to the complexity of the learning task itself. © 2011 Blackwell Publishing Ltd


䉬

401

Conclusion In this study, the semantic processing skills of L2 readers of Japanese from an alphabetic background were investigated, in order to address the following research questions: 1. Do L2 readers with differing levels of knowledge of the script involved demonstrate differences in their ability to recognize semantic information contained within characters? 2. How does L2 readers’ processing of the semantic information compare with that of experienced L1 readers? Three groups of L2 readers with different levels of knowledge of Japanese kanji characters, and a group of L1 readers, took a timed semantic processing task (a task involving the comparison of kanji). The kanji in the task had been chosen to represent differing degrees of semantic relatedness, as established by a prior norming study using a different group of L1 readers. By analysing the results of the task, the study described the changes in semantic processing skills at the different levels of kanji knowledge. The overall findings of the study suggest that: (1) L2 readers with different levels of target script knowledge approach the recognition of characters differently; and that (2) L2 semantic processing skills approximate those of L1 readers with increased L2 script knowledge. Character recognition initially relies on component-reliant processing and moves to more character-level dominant processing. The novice crude processing stage, as suggested in the existing models of word recognition development of L1 children, was not observed in the case of L2 readers. Instead, adult L2 readers seem to go straight into the transition period rather than starting from the partial processing period. These findings have provided a basis for insights into the types of issues that might be usefully investigated in future research into the reading development of L2 readers of Japanese and other character-based languages. The focus of the current study was on characters carrying both componentlevel and character-level information. However, characters are part of words, which offer character-level and word-level information. Japanese words written in kanji can appear as either single character or compound character words although most words written in kanji are compound character words consisting of two kanji characters (Hatta, Kawakami, and Tamaoka 1998). The characters that can be single-character words tend to have clear meanings of the although the degree of clarity varies. For example, the first character word /niNshiki/ ‘recognition’ can be a single-character word /mito-meru/ (the last two letters are written in hiragana script), which can be translated to ‘to accept’, ‘to acknowledge’ or ‘to recognize’ depending is used in many compound words, such on the context. This character /niNtei/ ‘acknowledgement’, and as /niNka/ ‘approval’, © 2011 Blackwell Publishing Ltd

402

䉬


/kakuniN/ ‘confirmation’. The meanings of these words are more or less related through the character . On the other hand, the characters that are used mostly as a constituent character of a compound word may not have a of the same word clear meaning. For example, the second character /niNshiki/ ‘recognition’ is not a single-character word (except in a very rare use). Therefore, what this character represents needs to be inferred by examining the meanings of several words containing this character. It /chishiki/ ‘knowledge’, /shikibetsu/ is used in words such as /keNshiki/ ‘insight’. However, it may not be simple to ‘distinction’, and determine that this character represents the concept ‘knowing’. Thus, a character has a potential semantic function in a word. However, the degree of semantic relatedness between a word and its constituent character(s) and between words sharing a common character varies. The relationship is sometimes transparent, and sometimes less so, depending mainly on whether or not the constituent character can be a stand-alone word. Whether L2 learners develop in kanji word semantic processing in a similar fashion as observed in this study may be an interesting research topic. The findings also offer a solid research foundation in the light of which proposals for the teaching of reading to such learners may be considered. The teaching of kanji was for long a neglected area in Japanese language teaching (Noguchi 1995). The most popular method of teaching was rote learning (Shimizu and Green 2002). Teachers introduced kanji to learners one by one together with their reading and meaning without breaking them into smaller units, and told the learners to practise reading and writing the kanji numerous times. Such learning conditions do not provide a good opportunity for learners to understand the structure and function within and between characters. At the same time, the use of contextual information was emphasized due to the belief that the reading and meaning of kanji words are highly context dependent (Shimizu and Green 2002), presumably under the influence of top-down theories of reading. However, Mori (2003) found that context often provided syntactic information but not semantic information about the words. Since the early 1990s, the use of mnemonic devices has been introduced to kanji instruction due to the influence of research findings in cognitive psychology (Shimizu and Green 2002). Mnemonic strategies often involve the analysis of smaller components of characters, that is, breaking a graphically complex kanji into manageable smaller units and making a short story by attaching meanings to the smaller units. However, while this method may assist readers in remembering the structure of individual kanji, it cannot give the readers semantic and phonological information across kanji. Explicit instruction on the use of the functional components has received attention in recent years. On-line and off-line teaching materials focusing on © 2011 Blackwell Publishing Ltd


䉬

403

teaching the functional components of kanji have become available. Evidence demonstrates that learners retain new kanji characters better using componential analysis (Kubota and Toyoda 2001). Although knowledge of functional components is essential, the interviews with the skilled readers (L1 readers) revealed that they do not know all the existing functional components and that they make use of only reliable functional components. In other words, when L1 readers have failed to access the meaning of kanji, they examine the form of a character by checking bushu as well as other parts. However, if the information conveyed by a bushu has no, or weak, relationship with the meaning of a kanji, or if the kanji that share a bushu have considerably different meanings, skilled readers would probably not use that bushu in character recognition. In fact, too much emphasis on components may encourage a microscopic approach which results in recognition errors. In order to bring L2 readers up to the most refined phase of character recognition development, repeated training for strengthening reliable links between units may be necessary. For example, when new words are introduced, L2 readers should be encouraged to pay attention to individual constituent characters, and to examine how the characters are, or are not, related to the word. For each of the constituent kanji, if it is a compound character, L2 readers should be assisted in analysing how the semantic information in the bushu is, or is not, related to the meaning of the character. Finally, exposure to an extensive range of kanji vocabulary, and some training in structuring and restructuring the vocabulary by categorizing the words under several abstract concepts may be critical for further development in character recognition.

Acknowledgements We are grateful for Dr Carsten Roever of The University of Melbourne for his generous and invaluable advice on the statistical analysis of data in this paper.

Notes 1. Firstly, all the kanji that were used in kanji recognition tasks (to be described in the following section) were marked as 1,2,3 or 4 according to their levels (1 is the most difficult level and 4 is the easiest). The ones marked as levels four and three were combined, as the number of kanji in these levels was relatively small. The total number of kanji selected was 315, with 27 characters in the combined group of levels four and three, 152 characters in level two, and 136 characters in level one. To make the number of items in each of the three levels more balanced, 108 characters from levels four and three were randomly drawn from the database, and formed a group of 135 kanji for the combined level. © 2011 Blackwell Publishing Ltd

404

䉬


2. 40 kanji consisted of 5 from each of the following groups Type

Low

Type

High

LMMM

Low complexity and the rest are medium Low character frequency and the rest are medium Low radical frequency and the rest are medium Low homophone size and the rest are medium

HMMM

High complexity and the rest are medium High character frequency and the rest are medium High radical frequency and the rest are medium High homophone size and the rest are medium

MLMM MMLM MMML

MHMM MMHM MMMH

Note: L = low, H = high and M = medium

References Akamatsu, N. (2003) The effects of first language orthographic features on second language reading in text. Language Learning 53.2: 207–31. Chikamatsu, N. (2006) Developmental word recognition: a study of L1 English readers of L2 Japanese. The Modern Language Journal 90.1: 67–85. Flores d’Arcais, G. B., H. Saito, and M. Kawakami (1995) Phonological and semantic activation in reading kanji characters. Journal of Experimental Psychology: Learning, Memory, and Cognition 21.1: 34–42. Grabe, W. (2004) Research on teaching reading. Annual Review of Applied Linguistics 24.1: 44–69. Gough, P. B. (1991) The first stages of word recognition. In L. Rieben and C. A. Perfetti (eds.), Learning to read: basic research and its implications. Hillsdale, NJ: Lawrence Erlbaum Associates. 47–56. Harris, M., and M. Coltheart (1986) Language processing in children and adults: an introduction. London: Routledge and Kegan Paul. Hatta, T., A. Kawakami, and K. Tamaoka (1998). Writing errors in Japanese kanji: a study with Japanese students and foreign learners of Japanese. In C. K. Leong and K. Tamaoka (eds.), Cognitive processing of the Chinese and the Japanese languages (Vol. 14). Dordrecht: Kluwer Academic Publishers. 303–16. Haynes, M., and T. H. Carr (1990) Writing system background and second language reading: a component skills analysis of English reading by native speaker-readers of Chinese. In T. H. Carr and B. A. Levy (eds.), Reading and its development: component skills approaches: New York, Academic Press. 375–421. Ho, C. S. H., and P. Bryant (1997) Learning to read Chinese beyond the logographic phase. Reading Research Quarterly 32.3: 276–89. Ke, C. (1998) Effects of language background on the learning of Chinese characters among foreign language students. Foreign Language Annals 31.1: 91–100. Koda, K. (1992) The effects of lower-level processing skills on FL reading performance: implications of instruction. The Modern Language Journal 76.4: 502–12. Koda, K. (1995) Cognitive consequences of L1 and L2 orthographies. In I. Taylor, and D. R. Olson (eds.), Scripts and literacy: reading and learning to read alphabets, syllabaries and characters. Dordrecht: Kluwer Academic Publishers. 311–26. © 2011 Blackwell Publishing Ltd


䉬

405

Koda, K. (1999) Development of L2 intraword orthographic sensitivity and decoding skills. The Modern Language Journal 83.1: 51–64. Koda, K. (2002) Writing systems and learning to read in a second language. In W. Li, J. S. Gaffney and J. L. Packard (eds.), Chinese children’s reading acquisition. Boston, MA: Kluwer Academic Publishers. 225–48. Kubota, M., and E. Toyoda (2001) Learning strategies employed for learning words written in kanji versus kana. Australian Review of Applied Linguistics, 24.2: 1– 16. Leck, K. J., B. S. Weekes, and M. J. Chen (1995) Visual and phonological pathways to the lexicon: evidence from Chinese readers. Memory and Cognition 23.4: 468–76. Linacre, J. M. (2009) WINSTEPS. Rasch measurement computer program. Beaverton, OR: Winsteps.com Mochida, A. and M. Harrington (2006) The yes/no test as a measure of receptive vocabulary knowledge. Language Testing 23.1: 73–98. Mori, Y. (2003) The roles of context and word morphology in learning new kanji words. The Modern Language Journal 82.3: 404–20. Nation, P. (1993) Using dictionaries to estimate vocabulary size: essential, but rarely followed procedures. Language Testing 10.1: 27–40. Nation, K., and M. J. Snowling (1999) Developmental differences in sensitivity to semantic relations among good and poor comprehenders: evidence from semantic priming. Cognition 70.1: B1–B13. Noguchi, M. S. (1995). Component analysis of kanji for learners from non-kanji using countries. The Language Teacher 19.10: 11–14. Nomura, M. (1984) Kanji no tokusei wo hakaru [Measuring characteristics of kanji]. In H. Kaiho (ed.), Kanji wo kagaku suru [Sciencing kanji]. Tokyo: Yuhikaku. 1–33 Read, J. (1993) The development of a new measure of L2 vocabulary knowledge. Language Testing 10.3: 356–71. Sayeg, Y. (1996) The role of sound in reading kanji and kana. Australian Review of Applied Linguistics 19.2: 139–51. Selinker, L. (1972) Interlanguage. International Review of Applied Linguistics in Language Teaching 10.3: 209–31. Shillaw, J. (1996) The application of Rasch modelling to yes/no vocabulary tests. Unpublished paper, Centre for Applied Language Studies (CALS), University of Reading. Available at http://www.lognostics.co.uk/vlibrary/index.htm. Retrieved 17 May 2010. Shillaw, J. (1999) The application of the Rasch model to yes/no vocabulary tests. PhD Thesis. University of Wales, Swansea. Shimizu, H., and K. E. Green (2002) Japanese language educators’ strategies for and attitudes toward teaching kanji. The Modern Language Journal 86.2: 227–41. Shu, H. (2003) Chinese writing system and learning to read. International Journal of Psychology 38.5: 274–85. Tamaoka, K., K. Kirsner, Y. Yanase, Y. Miyaoka and M. Kawakami (2002) A webaccessible database of characteristics of the 1945 basic Japanese kanji. Behavior Research Methods, Instruments and Computers 34.2: 260–75. Toyoda, E. (2009) Learning to read non-alphabetic script. Saarbrücken: VDM. Wilks, C., and P. Meara (2002) Untangling word webs: graph theory and the notion of density in second language word association networks, Second Language Research 18.4: 303–24. © 2011 Blackwell Publishing Ltd

406

䉬


Yamada, J., and H. Takashima (2001) The semantic effect on retrieval of radicals in logographic characters. Reading and Writing: An Interdisciplinary Journal, 14.1– 2:179–94. emails: [email protected] [email protected]

© 2011 Blackwell Publishing Ltd