Gsite.univ-provence.fr

Copyright 2008 by the American Psychological Association Native Language Influences on Word Recognition in a Second Language: Kristin Lemho¨fer, Ton Dijkstra, Herbert Schriefers, Centre National de la Recherche Scientifique Many studies have reported that word recognition in a second language (L2) is affected by the nativelanguage (L1). However, little is known about the role of the specific language combination of thebilinguals. To investigate this issue, the authors administered a word identification task (progressivedemasking) on 1,025 monosyllabic English (L2) words to native speakers of French, German, and Dutch.
A regression approach was adopted, including a large number of within- and between-language variablesas predictors. A substantial overlap of reaction time patterns was found across the groups of bilinguals,showing that word recognition results obtained for one group of bilinguals generalize to bilinguals withdifferent mother tongues. Moreover, among the set of significant predictors, only one between-languagevariable was present (cognate status); all others reflected characteristics of the target language. Thus,although influences across languages exist, word recognition in L2 by proficient bilinguals is primarilydetermined by within-language factors, whereas cross-language effects appear to be limited. An addi-tional comparison of the bilingual data with a native control group showed that there are subtle butsignificant differences between L1 and L2 processing.
Keywords: visual word recognition, bilingualism, progressive demasking task, cross-language influences,multiple regression Since the introduction of a common European currency, the of speaking a second language next to one’s native language. In the development of cultural, political, and economic integration within European Community as well as in many other parts of the world, the European Union has shifted into a higher gear. In the course of English has become the standard language of international com- this development, there is a growing awareness of the importance munication and one of the most frequently spoken second lan-guages. According to a survey by the European Union in winter2005, 56% of all Europeans were able to have a conversation in aforeign language (European Commission, 2006). For 38% of the Kristin Lemho¨fer, Ton Dijkstra, Herbert Schriefers, Nijmegen Institute E.U. population, this second language was English. Next in line of Cognition and Information, Radboud University Nijmegen, Nijmegen, were German and French (both 14%). In countries such as Lux- the Netherlands; R. Harald Baayen, Interfaculty Research Unit for Lan- embourg, Slovakia, Latvia, and the Netherlands, well over 90% of guage and Speech, Radboud University Nijmegen; Jonathan Grainger,Laboratoire de Psychologie Cognitive, Aix–Marseille University, and Cen- the citizens are bilingual or even multilingual.
tre National de la Recherche Scientifique, Marseille, France; Pienie Zwit- This growing awareness of the importance of multilingualism serlood, Psychological Institute II, University of Mu¨nster, Mu¨nster, Ger- has also stimulated research on bilingualism and multilingualism, including studies on how bilinguals recognize words in their first This research was part of Kristin Lemho¨fer’s PhD project at Radboud or second language. A long-standing debate in this area (Kolers, University Nijmegen within the framework of a MaGW International 1966; Macnamara & Kushnir, 1971) was concerned with the issue Comparative Research project, titled Lexical Competition in Bilinguals: A of whether the bilingual word-recognition process involved the Cross-Language Comparison, granted by the Netherlands Organization for initial activation of word representations from a target language Scientific Research to the Nijmegen Institute for Cognition and Informa-tion (400-60-000).
only (language-selective lexical access) or whether all words We would like to thank Mona Aicher, Mohamed El Halfaoui, Marije known to an individual, including those from a nontarget language, Michel, and Lonneke Bucken for their practical help in preparing the are considered as potential candidates for recognition (nonselective materials and running the experiments. We are also grateful to Marc access). A large number of studies have contributed to this dis- Brysbaert and Steve Lupker, who commented on earlier versions of this cussion, the majority of them demonstrating that the two languages do interact during word recognition. For instance, it has been Correspondence concerning this article should be addressed to Kristin shown that when bilinguals recognize words in one of their lan- Lemho¨fer, who is now at Max-Planck-Institute for Psycholinguistics, P.O.
Box guages, they process orthographically identical cognates (words that share form and meaning in two languages, e.g., art in French and English) and, under certain circumstances, also false friends members (stare in the place of stair) more often than graphemic (words that share their form but not their meaning, e.g., pain in controls (e.g., stars), whereas this effect was not present for French [meaning bread] and English) differently from words that Chinese speakers. The authors concluded that native speakers of exist in the target language only (Lemho¨fer & Dijkstra, 2004; a Korean (a language with a highly consistent grapheme-to- more detailed overview of this research appears below). Moreover, phoneme mapping) rely more on phonological than on ortho- van Heuven, Dijkstra, and Grainger (1998) demonstrated that graphic information, whereas the reverse is true for speakers of a bilingual word recognition in a given language is affected by logographic native language such as Chinese.
orthographically similar words, so-called orthographic word The lack of studies comparing the influence of different mother neighbors, from the other language.
languages within the same alphabet on word recognition in L2 These findings have led to the conclusion that bilingual lexical hints at a tacit agreement that at least for fairly related languages, access can most accurately be described as initially language the specific L1–L2 combination does not play a crucial role.
nonselective (for an overview, see Dijkstra & van Heuven, 2002; However, this assumption is not necessarily true and remains to be Kroll & Dijkstra, 2001). It is especially well established that word tested. Even though they share the same script, alphabetic lan- recognition in the second language (L2) can be affected by the guages differ in many respects, some of which have been claimed native language (L1), although effects in the other direction have to affect how monolingual speakers recognize words. One example also been demonstrated (e.g., Duyck, 2005; van Hell & Dijkstra, is the orthographic depth hypothesis (Katz & Frost, 1992), claim- 2002; van Wijnendaele & Brysbaert, 2002). However, hardly any ing that the regularity of the mapping between spelling and sound research has investigated the extent to which the size and types of in a language (orthographic depth) determines which route to the cross-linguistic effects in L2 depend on the characteristics of the mental lexicon is used. According to this hypothesis, word recog- respective native language and its relation to L2. The first aim of nition in a language where the sound of a word can be directly the present study was to determine the extent to which results derived from its spelling (i.e., in “shallow” languages, such as obtained from bilinguals with a particular L1–L2 combination Italian or Finnish) is possible without the involvement of the generalize to other bilingual populations with the same L2 but mental lexicon. In contrast, recognition of words in orthographi- different L1s. We addressed this issue by comparing three groups cally deep languages benefits from the use of all sorts of lexical of bilinguals (native speakers of French, German, and Dutch) with (i.e., frequency or semantic) information. In line with this reason- respect to their performance on a word recognition task in their ing, differences in orthographic depth have been found to modulate the effects of word frequency (Frost, Katz, & Bentin, 1987) or of The second aim of the study was to create an experimental semantic variables (Cuetos & Barbo´n, 2006) in word recognition setting containing a large and representative set of word stimuli not tasks, in particular in word naming but also in silent tasks such as specifically designed to generate cross-language influences. In lexical decision. Similarly, differences in phonological consistency contrast, virtually all previous studies on bilingual word recogni- between Dutch and English appear to cause differences in sublexi- tion have focused on whether cross-language interactions can be cal clustering during visual word recognition (Martensen, Maris, & observed at all, and have used factorial designs and restricted sets Dijkstra, 2000; Ziegler, Perry, Jacobs, & Braun, 2001), which of selected items (such as form-identical cognates) in an effort to might show up in varying effects of variables such as bigram make those interactions as visible as possible. However, it is as yet frequency. If different languages do indeed entail varying pro- unclear what role cross-language influences play within the spec- cesses and strategies during visual word recognition, these pro- trum of all possible influences of stimulus characteristics, includ- cesses and strategies might (partially) be transferred from the ing the influence of within-language variables. The present study native to a second language, just as what has been claimed for examines how these different variables might simultaneously af- native speakers of languages with different scripts.
fect word recognition in a second language (rather than holding Thus, even within the same alphabet, different language back- within-language variables constant, as in factorial studies).
grounds might give rise to different reading strategies in a common So far, only little systematic research has addressed the first aim, second language. In the present study, three groups of participants the impact of different native languages on visual word recognition were included for whom their mother language (French, German, in L2. The existing studies focus on the difference between L1–L2 or Dutch) and their second language (English) are western Euro- combinations in terms of whether the two languages use the same pean languages sharing the same alphabet (except for accents and or different writing scripts. The reasoning underlying these studies special letters) but with different degrees of orthographic depth: is that bilinguals might transfer reading strategies they acquired in German is the most shallow among these orthographies, with their native language to their second language. For example, almost a one-to-one mapping of orthography and phonology, Muljani, Koda, and Moates (1998) observed that native speakers whereas Dutch is somewhat deeper, followed by French, and of Indonesian (a language with an alphabetic writing system) English has the deepest orthography of these four languages (Sey- performed better in a lexical decision task in their second lan- mour, Aro, & Erskine, 2003). By choosing participants who all had guage, English, than equally proficient participants with Chinese the same L2, we were able to use the same (English) word as their first language. Akamatsu (2002) showed that bilingual materials for all three participant groups and thus to single out the speakers with Chinese, Japanese, or Persian as L1 and English as differential role of the mother language.
L2 displayed differential effects of word frequency but comparable This study was also designed to provide a further test of the effects of phonological regularity in English word naming. In a nonselective nature of bilingual lexical access. However, rather study using the semantic categorization task developed by van than orthogonally manipulating the factors in question and care- Orden (1987), Wang, Koda, and Perfetti (2003) found that native fully selecting a small number of stimuli, as in previous studies, we speakers of Korean misclassified homophones of English category used a multiple regression design, involving a large stimulus set (1,025 English words). Regression designs have a number of is a variant of the perceptual identification task in which a word important advantages over the more commonly used factorial that slowly emerges from a pattern mask must be identified as designs. For instance, as pointed out by Balota, Cortese, Sergent- quickly as possible and must be typed in after identification.
Marshall, Spieler, and Yap (2004), there are a number of problems Presumably, masking slows down the word recognition process, that arise from the a priori selection of restricted sets of stimuli in making the task especially sensitive to factors affecting the early factorial designs, such as difficulties with matching the stimuli on stages of word recognition. Of course, the fairly recent emergence all relevant dimensions, the possible occurrence of experimenter of this task as an alternative to other more standard word recog- biases during stimulus selection, and the sometimes disproportion- nition paradigms may complicate the comparison with the litera- ate use of words that take extreme values on the target dimensions.
ture. However, as Carreiras, Perea, and Grainger (1997) pointed In contrast, in regression designs, the variables of interest are not out, PDM and other perceptual identification tasks represent a used as selection criteria of the stimulus set, thereby avoiding these purer measure of orthographic word processing than tasks like difficulties. Furthermore, considering that most variables under lexical decision or word naming, because (unlike lexical decision) investigation are continuous (e.g., word frequency, orthographic they require the unambiguous identification of the word and are neighborhood), regression analyses avoid the loss of information not influenced by external factors like the nature of nonword foils that is associated with the categorization of continuous variables.
or articulatory factors. Furthermore, in a bilingual setting, the Finally, regression designs allow for establishing not only whether PDM paradigm seems to be more appropriate than the lexical a particular factor has an effect on the dependent variable but also decision task, because problems associated with selecting a set of how large the contribution of this factor is in accounting for the nonwords that are “neutral” with respect to the different mother languages involved are avoided. Moreover, note that the naming Because of these advantages, an increasing number of large- task is also not ideal for use in a nonnative language setting, scale psycholinguistic studies have recently used regression de- because participants’ L1 will likely determine how easily they can signs (e.g., Alario et al., 2004; Baayen, Feldman, & Schreuder, 2006; Balota et al., 2004; Bird, Franklin, & Howard, 2001; Cuetos This task was carried out on 1,025 monosyllabic English words & Barbo´n, 2006; Spieler & Balota, 1997). However, to our knowl- by three groups of 20 bilinguals each, with French, German, or edge, only one study of this kind has so far been conducted within Dutch as their native languages. Additionally, a control group of the domain of visual word recognition in bilinguals. De Groot, 20 native speakers of English performed the same experiment, to Borgwaldt, Bos, and van den Eijnden (2002) had Dutch–English enable evaluation of the obtained results with respect to native bilinguals carry out a lexical decision and a naming task in both speakers’ performance. We chose to include all variables that Dutch (L1) and English (L2). They examined the differences both current approaches of word recognition regard as important fac- between the two tasks and between the languages with respect to tors, insofar as they could be calculated using the lexical databases the effects of a number of semantic, orthographic, and phonolog- British National Corpus (BNC Consortium, 2001) for English; ical variables. Generally speaking, the results revealed consider- CELEX (Baayen, Piepenbrock, & Gulikers, 1995) for English, able differences between the tasks (explainable by the varying task Dutch, and German; and LEXIQUE (New, Pallier, Brysbaert, & demands) and smaller differences between the two languages, Ferrand, 2004) for French. In the following, we describe the especially for lexical decision. In that study, effects of the indi- chosen predictors and how they were expected to influence word vidual variables were reported in terms of absolute predictor– recognition performance in our populations of second language criterion correlations, which are difficult to interpret owing to the high intercorrelations between the predictors themselves. Leavingthis problem aside, the results of that study suggest that among the cross-language variables, cognate status had a facilitatory effect onboth lexical decision and naming reaction times (RTs), but only in L2, whereas (for short words) cross-language orthographic neigh-bors affected naming but not lexical decision RTs in both lan- The word frequency effect (more frequent words are recognized faster than words with a lower frequency) is one of the most robust In contrast to the study by de Groot et al. (2002), the present findings in the visual word recognition literature (e.g., Howes & study involved a methodological design that allows for the simul- Solomon, 1951; Schilling, Rayner, & Chumbley, 1998; Whaley, taneous assessment of the partial effects of a large number of 1978) and has been incorporated in virtually every monolingual predictor variables. If the claim of a profoundly nonselective model of word recognition. For instance, interactive activation nature of bilingual lexical access holds true, variables that have models of lexical access assume that frequency affects the resting previously been shown to carry interlingual influences, like cog- activation levels of word representations (McClelland & Rumel- nate and false friend status and between-language orthographic hart, 1981; for a more detailed discussion of frequency effects and neighborhood, should also show significant effects for the present their locus in visual word recognition, see Allen, Smith, Lien, comprehensive set of stimuli and predictor variables.
Grabbe, & Murphy, 2005; Hino, Lupker, Ogawac, & Sears, 2003).
The task chosen in the present study was the progressive de- Even though the size of the effect is task dependent (e.g., Balota et masking (PDM) task developed by Grainger and Segui (1990; see al., 2004), it has been reported for all standard tasks of visual word Dufau, Stevens, & Grainger, in press, for freely available software recognition, including the PDM task (Grainger & Segui, 1990; for this task). This task has been shown to produce similar patterns Perea, Carreiras, & Grainger, 2004). For the bilingual domain, of results as the more frequently used lexical decision task (Dijk- some evidence suggests that the frequency effect might even be stra, Grainger, & van Heuven, 1999; van Heuven et al., 1998). It larger in the second as compared with the first language (van Wijnendaele & Brysbaert, 2002). However, de Groot et al. (2002) times in reading; see New, Ferrand, Pallier, & Brysbaert, 2006, for observed different sizes of the frequency effect in L2 relative to L1 a more complex pattern in lexical decision). Of more interest, only for word naming and not for lexical decision. Furthermore, if Ziegler et al. (2001) have shown that word length effects were languages with varying orthographic depths indeed give rise to larger in German as opposed to English, possibly due to German differences in the frequency effect, as suggested by Frost et al.
having a more shallow orthography than English. Consequently, (1987), native speakers of Dutch, German, and French (the last owing to the possible transfer of reading strategies from L1 to L2, having the deepest orthography among the three) might display word length effects may also differ for bilingual readers varying in different frequency effects also in their L2, English, due to a their L1 when reading words in their common L2.
transfer of word recognition processes from L1 to L2. In that case,native speakers of French should show a larger frequency effect than those with more shallow L1 orthographies, such as Dutch andGerman.
Effects of orthographic neighborhood (i.e., of words that differ Recently, the idea has emerged that owing to sex differences in from the respective word in one letter only; Coltheart, Davelaar, verbal memory (Kimura, 1999), females process words differently Jonasson, & Besner, 1977) are thought to reflect the activation of from males. For example, Ullman et al. (2002) found that the role multiple word candidates during word recognition (e.g., Andrews, of word frequency during the production of inflected verb forms is 1997). The relative importance of various neighborhood measures modulated by sex. Here, we included the sex by frequency inter- for the different standard word recognition tasks has been dis- action as a predictor variable to investigate whether sex differences cussed extensively (see Andrews, 1997, and Perea & Rosa, 2000, in verbal memory have consequences for the frequency effect, for reviews). In PDM or similar perceptual identification tasks, the number of higher frequency neighbors has repeatedly been found Finally, Baayen et al. (2006) demonstrated that the relative to slow down recognition latencies for the target word, whereas the frequency in written compared with spoken English (quantified as total number of neighbors had no or only little effect on recogni- the ratio between the two) played an important role in both English tion performance (Carreiras et al., 1997; Grainger & Jacobs, 1996; monolingual lexical decision and word naming: The more frequent but see Snodgrass & Mintzer, 1993). Higher frequency neighbors a word was in spoken relative to written English, the faster it was are thought to delay the pass of the recognition threshold for a recognized. In the present study, we took a slightly different target word through lateral inhibition (Grainger & Jacobs, 1996).
approach and included both written and spoken frequency (astaken from the British National Corpus) as predictors in the re- gression analyses, to determine their separate roles both for non-native and for native speakers of English. Our populations of Average bigram frequency is a variable capturing the ortho- bilinguals, who are not immersed in an English-speaking environ- graphic typicality of a word in the context of its own language ment, might be exposed to proportionally more written than spo- (Rice & Robinson, 1975). Whether or not it is relevant to the word ken English relative to native speakers, possibly leading to a recognition process has been subject to debates from the first difference in the relative importance of written versus spoken word appearance of this variable in monolingual word identification frequency for first and second language speakers.
studies (Broadbent & Gregory, 1968; Gernsbacher, 1984; McClel-land & Johnston, 1977; Rumelhart & Siple, 1974). All of these studies used perceptual word identification tasks. Those that re-ported an effect of bigram frequency (which was primarily the case Recently, it has repeatedly been shown that the number of for low-frequency words) found that it was inhibitory, which is derivations and compounds in which a word occurs, the morpho- probably a consequence of the difficulty to discriminate “typical,” logical family size, facilitates response latencies in monolingual high-bigram-frequency words from similar words. More recently, and bilingual lexical decision (de Jong, Schreuder, & Baayen, Westbury and Buchanan (2002) demonstrated that it was minimal 2000; Dijkstra, Moscoso del Prado Martı´n, Schulpen, Schreuder, bigram frequency (i.e., the frequency of the least likely bigram in & Baayen, 2005; Schreuder & Baayen, 1997). Paradoxically, it a word), representing a “simple sublexical marker of how unusual seems to be the number of morphological family members that a word is” (p. 68), that, for high-frequency words, had an (inhib- influences recognition latencies, not their frequency. This argues itory) effect on lexical decision latencies. Assuming that effects of against a purely frequency-based account of the morphological bigram frequency indeed reflect sublexical orthographic process- family size effect. Schreuder and Baayen themselves suggest a ing, they may be sensitive to the different language backgrounds of semantic source of the effect. Whether the effect generalizes across our participant groups, again assuming the transfer of reading tasks and participant populations is still subject to research, espe- cially because Schreuder and Baayen failed to find the effect in thePDM task.
Research on the relationship between the form and semantic level of word representation has made extensive use of words that It is hardly surprising that for most word recognition tasks, have several meanings. However, there is considerable disagree- longer words take longer to recognize (e.g., McGinnies, Comer, & ment concerning whether and how this variable influences word Lacey, 1952, for perceptual word identification; Ziegler et al., recognition (e.g., Borowsky & Masson, 1996; Duffy, Morris, & 2001, for word naming; Just & Carpenter, 1980, for eye fixation Rayner, 1988; Gernsbacher, 1984; Hino, Pexman, & Lupker, 2006; Piercey & Joordens, 2000; Rodd, Gaskell, & Marslen-Wilson, 2004) and whether related word senses have to be discriminatedfrom unrelated word meanings (Klein & Murphy, 2001; Kle- Subjective word familiarity has been claimed to play an impor- pousniotou, 2002; Rodd, Gaskell, & Marslen-Wilson, 2002). To tant role in (monolingual) word recognition (Connine, Mullennix, our knowledge, the effect of number of meanings has not yet been Shernoff, & Yelen, 1990; Gilhooly & Logie, 1982; Kreuz, 1987; investigated for the PDM task, in either L1 or L2 processing. Thus, Williams & Morris, 2004). Gernsbacher (1984) even reported that the question is whether native and nonnative speakers are influ- effects of other variables (bigram frequency, concreteness, and enced by the number of word meanings during a word recognition number of meanings) on lexical decision latencies disappeared task that presumably involves relatively little semantic processing.
when familiarity was controlled for. Balota, Pilotti, and Cortese Considering that representations of L2 words have been regarded (2001) showed that familiarity ratings, as they are usually ob- as less “richly populated” (i.e., possessing fewer senses) than L1 tained, are different from subjective ratings of frequency and that words (Finkbeiner, Forster, Nicol, & Nakamura, 2004), it is pos- they are more related to a semantic variable (meaningfulness).
sible that the number of meanings affects word recognition in the Effects of familiarity (as rated by a bilingual population) have also first but not in the second language.
been reported for bilinguals, in both their L1 and their L2 (deGroot et al., 2002). In the present study, we included familiaritynot only to investigate its effects on PDM latencies for the differ- ent participant groups but also to test whether effects of seeminglyrelated variables, such as word frequency, bigram frequency, num- In contrast to semantic ambiguity, syntactic ambiguity of words ber of meanings, concreteness, or meaningfulness, would persist has as yet received relatively little attention in research on isolated words. In a study using event-related potentials, Elston-Gu¨ttler andFriederici (2005) demonstrated that initially, both meanings of syntactically ambiguous homonyms presented in sentence contextwere activated but that the ambiguity was resolved at a later point Even though meaningfulness, defined as the ease with which a in time. However, the disambiguation mechanism was shown to be word can be associated with other words, has been included in more effective in native than in nonnative speakers. The latter many lexical databases (Locascio & Ley, 1972; Paivio, Yuille, & finding would imply that relative to bilinguals, native speakers Madigan, 1968; Spreen & Schulz, 1966; Toglia & Battig, 1978; should show smaller or no effects of syntactic ambiguity, owing to Wilson, 1988), few studies have investigated its effect on word recognition. Johnson and Zara (1964) and Johnson, Frincke, andMartin (1961) observed that words with higher meaningfulnessratings had lower visual recognition thresholds than those with low ratings, at least for the low-frequency range. Recently, the variable For a subset of words (n ϭ 659) for which these values were was found to be an important predictor of subjective frequency available, we included the semantic variables concreteness, famil- ratings (Balota et al., 2001) and to facilitate RTs in lexical decision iarity, and meaningfulness from the MRC Psycholinguistic Data- for young adults (Balota et al., 2004). Presumably, the stronger base (Wilson, 1988). For all of these variables, the orthographic semantic connections in which meaningful words are embedded in depth hypothesis would predict smaller effects for orthographi- conceptual memory help their recognition, similar to what has cally shallow languages (e.g., German) relative to deeper lan- been claimed with respect to concrete words.
guages (French), and in the case of a transfer of reading strategiesfrom L1 to L2, this might also show up in the present participants’ L2, English. Note that English is the deepest among the fourlanguages; therefore, any semantic effects should be largest for the Orthographic Neighborhood Variables With Respect to L1 Besides the investigation of the effects of word characteristics with respect to English itself, we expanded our approach to cross- language influences and examined the role of the participants’ firstlanguage in the present experiment. As a first set of cross-language Concreteness has been considered one of the major semantic variables, we included orthographic neighborhood variables with variables that influence the recall and recognition performance of respect to L1. Any occurrence of a between-language neighbor- words. Within the monolingual word-recognition literature, some hood effect would indicate that during the recognition of a word in studies have shown that concrete words are recognized more easily a second language, word candidates of the (nontarget) native than abstract ones in various word recognition tasks (James, 1975; language become active as well and compete for recognition.
Juhasz & Rayner, 2003; Kroll & Merves, 1986; Richards, 1976), Thus, such an effect (as well as any other effect of a between- with the effect being more pronounced for words in the lower language variable) would provide strong support for a nonselective frequency range. These effects are usually ascribed to a greater view of bilingual lexical access. In one of the few studies on richness, associative embedding, or imagistic quality of the seman- between-language orthographic neighborhood effects, van Heuven tic representation for concrete compared with abstract words et al. (1998) found that Dutch–English bilinguals performing PDM (Samson & Pillon, 2004); this results in faster semantic processing, and lexical decision in English (L2) reacted more slowly when the which can in turn influence the orthographic processing of a word, number of Dutch (L1) orthographic neighbors increased. In con- trast, in a nonfactorial design, as the present one, de Groot et al.
(2002) found null effects for both English and Dutch neighbor- interlingual homographs, Dijkstra et al. (1999) observed that ho- hood density in an English lexical decision task, when analyz- mographs that shared their spelling but not their pronunciation in ing short words (of three to five letters) only. Thus, the evi- Dutch and English were recognized faster than English control dence on interlingual neighborhood effects is as yet both scarce words, both in an English lexical decision task (see also Lemho¨fer & Dijkstra, 2004) and in a PDM task.
The discrepancy between the reliable cognate effects and the Interlingual Cognates and Homographs fragile or variable homograph effects indicates that there arefundamental differences in the processing of these two kinds of Investigating how words that overlap in form across two lan- words. Thus, whether form-identical or similar interlingual word guages—that is, orthographically identical cognates (hereafter pairs also share their meaning is crucial. For instance, it is possible simply referred to as cognates) and interlingual homographs—are that cognates share one orthographic representation in the bilingual processed by bilinguals in different task and language contexts can lexicon whereas noncognate homographs do not (Gollan, Forster, provide valuable information on the structure of the bilingual & Frost, 1997; Sa´nchez Casas, Davis, & Garcı´a Albea, 1992).
language system. Different RTs and error rates for these words The three native languages of our participants differ with respect compared with control words that exist in only one language are to the number of cognates and false friends they share with usually interpreted as a consequence of the coactivation of both English: For instance, Dutch shares more cognates with English readings in the two languages, and thus as support for the nonselective than does French (see Table 4 later in article). Possibly, the nature of bilingual lexical access (e.g., Dijkstra et al., 1999).
bilingual language system exploits this sort of lexical similarity Most studies on the recognition of cognates by bilinguals have more effectively when it occurs more often. Thus, Dutch partici- demonstrated that cognates are recognized faster and/or with pants might show a larger cognate effect than native speakers of higher accuracy than control words. This has most frequently been French, because they might make more use of overlapping lan- shown for lexical decision in L2 (Caramazza & Brones, 1979; Cristoffanini, Kirsner, & Milech, 1986; de Groot et al., 2002;Dijkstra, van Jaarsveld, & ten Brinke, 1998) or L3 (Lemho¨fer, Dijkstra, & Michel, 2004). Furthermore, Dijkstra et al. (1999)obtained the facilitatory cognate effect not only for lexical decision In contrast to the consistent picture concerning cognate effects, Twenty-one native speakers of French, German, or Dutch with experimental evidence on interlingual homographs, or false friends English as a second language participated in the experiment, all in (without shared meaning in the two languages), is less conclusive.
their own country. The participants were mostly recruited among Both Dijkstra et al. (1998) and de Groot, Delmaar, and Lupker university students and staff. Participants were tested for their (2000) found that whether and in which way such words are English proficiency by means of a vocabulary-size test, described processed differently from control words depends on the task, the stimulus list composition, and the target language (L1 or L2). In The data from 1 participant in each country were excluded English (L2) lexical decision tasks with a purely English stimulus because of high error rates in the proficiency test, so that the list, both studies observed small, nonsignificant inhibitory effects remaining number of participants was 20 per group. The number of for those homographs that had a low frequency in English and a men and women among these remaining participants was, respec- high frequency in Dutch (L1). Similar but significant inhibitory tively, 2 and 18 in France, 6 and 14 in Germany, and 7 and 13 in homograph effects were found by von Studnitz and Green (2002).
the Netherlands. The participants’ experience with English, as In contrast to these studies showing null effects or inhibition of reported in a language questionnaire, is listed in Table 1.
Table 1Means (and Standard Deviations) in the Language Questionnaire for the Three ParticipantGroups How often do you read English in leisure time? Self-rated speaking experience in English Asterisk indicates dimension for which there were significant ( p Ͻ .05) differences between the three participant populations, as analyzed by one-way analyses of variance.
a Self-ratings were given on a scale from 1 (low) to 7 (high).
One-way analyses of variance were conducted to examine Word frequency and morphological family size. whether there were significant differences between the three par- spoken word form frequency were taken from the British National ticipant populations. Such a difference was found only for the Corpus (BNC Consortium, 2001) and normalized for corpus size.
frequency with which they read English at work. Paired t tests Word form frequency refers to the frequency of occurrence of a showed that both French and Dutch participants reported reading specific surface word form excluding its inflections (e.g., the plural more English for their work than did German participants: form). Frequency was one of our predictors that was characterized French Ͼ German, t(57) ϭ 4.46, p Ͻ .001; Dutch Ͼ German, by a highly skewed distribution. To avoid the possibility that a few t(57) ϭ 4.88, p Ͻ .001. None of the other comparisons were atypical data points with extreme values would exert undue lever- age in the regression, we removed most of this skew by means of For the control experiment, 20 native speakers of English cur- a logarithmic transformation. This practice is common in the word rently visiting Nijmegen, the Netherlands, who reported having recognition literature (e.g., Baayen et al., 2006; Baayen, Tweedie, little knowledge of Dutch and not speaking any foreign language & Schreuder, 2002; Balota et al., 2001). When a word had several regularly, performed the same experiment as the bilingual speak- entries in the database, as many words had (e.g., bite is both a verb ers. They were, on average, 21.9 years old (SD ϭ 2.6). Ten of them and a noun), the frequencies of the different entries were added were male, 10 female. They had grown up in the United Kingdom (6 participants), the United States (11), Canada (2), or Australia Morphological family size refers to the number of words and (1). Some reported having been in contact with foreign languages compounds that are derived from the word itself. For example, other than English (Spanish, French, German, Italian, or Dutch).
according to CELEX, the word bride has five family members The mean rating for their degree of experience with those foreign (with a frequency larger than 0): bridal, bridecake, bridegroom, languages was 2.7 on a scale from 1 (low) to 7 (high); the mean bridesmaid, and bride-to-be. Following the same motivation as for rating for their frequency of usage was 1.8.
frequency, we logarithmically transformed the family size counts.
English orthographic neighborhood. hood characteristics, the number and summed frequency (taken The experimental stimulus set consisted of English words se- from the CELEX corpus) of orthographic neighbors of a word lected from the CELEX database (Baayen et al., 1995). They met within English (i.e., words of the same length differing in exactly the following criteria: They were between three and five letters one letter) were counted. This measure did not include neighbors long; only content words were used (i.e., nouns, verbs, adjectives, with a frequency lower than 1 o.p.m. For testing whether higher and adverbs); they were monosyllabic; each word had only one and lower frequency neighbors would have differential effects on possible spelling and one pronunciation; the written lemma fre- word recognition, the number of orthographic neighbors with a quency of the words according to the CELEX database lay be- frequency higher or lower than the word itself was calculated for tween 10 and 10,000 occurrences per million (o.p.m.); and words each word. Again, all neighborhood variables were logarithmically with more than two different entries in CELEX (i.e., due to the number of syntactic categories the word can belong to) were not included. Exceptions to these restrictions were made for words determined, based on either word types or word tokens in CELEX.
(n ϭ 20) that had been part of the stimulus materials in our earlier For the type count, the number of word forms with the same studies, to allow us to compare the results with previous ones in the length, sharing a given bigram in the same position, was taken into account. For the token count, the word form frequencies of these To make sure that words would be known by the participants, words were summed for each bigram. On the basis of preliminary we gave a list of the 200 least frequent words to five students analyses (hierarchical regression analyses with different variable (native speakers of Dutch) at the University of Nijmegen, drawn orders), we selected the two most predictive counts: the token from the intended population of Dutch participants in the main counts of the mean bigram frequency (average frequency of all experiment. They were asked to indicate which of the words they bigrams in the word) and minimal bigram frequency (frequency of did not know. Words that were unknown to one or more persons the least frequent bigram in the word). Both variables were loga- (80 out of the 200 words) were excluded. The final set of stimuli Number of syntactic word categories and meanings. For the multiple regression analyses, a number of within- and ber of CELEX entries for each word was also entered as a variable.
between-language variables were calculated as predictors. For the In most cases, this variable reflected the number of syntactic native control group, obviously, only the within-language predic- categories a word can adopt. In only two cases (sake and rose), there were separate entries for two meanings of a word eventhough they belong to the same syntactic category, because they Description of Predictors of the Regression Analyses possess different etymological roots. Because there were few items The variables characterizing the English properties of the words with more than two entries (see the selection criteria above), the were drawn from the English part of the CELEX database, as well variable was dichotomized (one entry vs. more than one entry).
as from the British National Corpus (BNC Consortium, 2001).
For the number of meanings of a word, The Wordsmyth Dictio- nary (Wordsmyth, 2002) was used to determine the number of Length was coded in number of letters, ranging 1 The list of stimuli as well as the item means for the four participant from three to five. All words were monosyllabic in English.
groups can be obtained from Kristin Lemho¨fer.
Table 2Characteristics of the Word Materials With Respect to English Number of English lower frequency neighbors Number of English higher frequency neighbors English log summed neighborhood frequency Number of word meanings (Wordsmyth) Unless stated otherwise, variables are reported in absolute rather than logarithmic values. BNC ϭ British National Corpus.
a This variable was dichotomized (equal to 1 vs. greater than 1). b Available for only a subset of words (659items).
semantically unrelated meanings, similar to the method of Rodd et Dutch, German, and French were calculated. In addition, the al. (2002). The Wordsmyth Dictionary is suitable because it lists number of neighbors was split into high- and low-frequency neigh- the number of unrelated meanings regardless of syntactic category; bors. Here, an absolute frequency split was used, distinguishing thus, this variable is not confounded with the previous variable, the between L1 neighbors above and below a frequency of 50 o.p.m.3 number of CELEX entries.2 For example, the word host has three In analogy to the within-language variables, all neighborhood semantically distinct word meanings: (a) a person who entertains variables were logarithmically transformed. In Table 3, a summary guests, (b) a very large number of people or things, and (c) of the word characteristics is given with respect to the between- The characteristics of the 1,025 words with respect to the mentioned intralingual variables are shown in Table 2.
their form but not their meaning in two languages as (noncognate) interlingual homographs or false friends; words that have both the the MRC database (Wilson, 1988): concreteness, familiarity, and same spelling and the same meaning are referred to as (orthograph- meaningfulness. These values were available for only a subset ically identical) cognates. Owing to computational restrictions, (659) of the 1,025 words. The ratings for familiarity and concrete- only words that are orthographically identical with their reading in ness in the MRC database were merged values from three sets of the respective other language were counted as homographs or norms and transformed to values between 100 and 700. The cognates; for example, the English word bed was counted as a meaningfulness ratings listed in the database were taken from cognate with respect to Dutch (bed) but not with respect to German Toglia and Battig (1978) and were also multiplied so that they lay (Bett). Homograph and cognate status between English and the in a range from 100 to 700. The word characteristics with respect respective L1 were coded as dichotomous (0 or 1) variables. The to these semantic variables for the subset of 659 words are also numbers of interlingual homographs and form-identical cognates for each of the three language pairs are given in Table 4.
For the coding of between-language variables, the following Because of the large number of trials, the experiment was databases were used: the German and Dutch part of the CELEX conducted in three sessions, which were held on different days.
database for word forms, excluding words with a frequency lowerthan 1 o.p.m., and the French LEXIQUE, which also contains 2 Rodd et al. (2002) distinguished between the number of (unrelated) words with a frequency of 1 o.p.m. or more. Note that for a given meanings and the number of (related) senses. We did not include the latter word, whereas within-language variables take the same values for measure because it is correlated with the number of CELEX entries, as the all three participant groups, cross-language variables are by defi- different syntactic categories a word can adopt are counted as senses in The nition different for the French, German, and Dutch participants (e.g., a given English word might be a cognate with respect to 3 Because of the impossibility of directly comparing first- and second- French but not regarding German or Dutch).
language word frequencies for unbalanced bilinguals, an absolute split criterion was used rather than the relative one used in the case of within- number and summed frequency of orthographic neighbors in Table 3Characteristics of the Word Materials With Respect to French, German, and Dutch No. of orthographic neighbors with freq. Ͼ 50 No. of orthographic neighbors with freq. Ͻ 50 All variables are reported in absolute rather than logarithmic values. Freq. ϭ frequency.
Participants were tested individually. The experiment was con- time-out deadline would have been after 650 refresh cycles, but trolled by software developed in collaboration with the Technical this time limit was never reached). After the participant had Group of the Nijmegen Institute for Cognition and Information, responded, a window appeared on the screen with the text please running on Macintosh computers. The experiments were run in enter the word, and the participant was to type the word using the Nijmegen, the Netherlands; Aix-en-Provence, France; and Mu¨n- computer keyboard. One second after the participant had pressed ster, Germany. The same software was used in all three laborato- Enter to enter the word, the next trial was started by the participant ries. In each lab, the monitor was a 17-in. screen with a refresh rate pressing a button. No feedback was given to the participant con- of 66 MHz and a resolution of 640 ϫ 480 pixels, placed at a cerning whether the answer had been correct.
distance of about 60 cm from the participants.
Every experimental session took about 60 to 75 min. Partici- Each session began with a practice block (10 trials in the first pants received course credit or money for their participation after session, 5 in the second and third ones), consisting of words that were not used in the actual experiment. In the main experiment thatfollowed, 342 (first and second sessions) or 341 words (third session) were presented in seven blocks of 50 words (except forthe last block, which was shorter). Every participant within each At the beginning of the first session, before the experiment, the group received a differently randomized order of words; however, participants completed an English vocabulary-size test. The results the same 21 randomizations were used in all three labs. An English of this test were analyzed immediately to make sure that the person instruction was given to the participants explaining that they would met the proficiency criteria (no more than 20 errors out of the total see a word alternating with a mask, with the word gradually of 60 items) for the experiment. The test was a nonspeeded lexical becoming more visible. They were asked to press a button on a decision task on 40 low-frequency English words and 20 highly button box as soon as they had recognized the word and to type in wordlike nonwords and was derived from an unpublished the word after a prompt had appeared.
vocabulary-size test developed for high-proficiency populations by Words were presented in black, lowercase Courier letters (size P. Meara (Meara, 1996). This test version is described in more 18 point) in a white window, which was surrounded by a black detail in a previous publication (Lemho¨fer et al., 2004). The test background. At the beginning of a trial, a fixation cross appeared score was calculated using a percentage correct measure (i.e., the in the middle of the screen. When the participant pressed the unweighted mean percentage of correctly recognized words and button, one of two checkerboard masks appeared on the screen, correctly rejected nonwords). The results of the three groups of consisting of the same number of checkerboard blocks as the word bilingual participants, as well as those of the native speakers, are had letters. The mask remained on the screen for 25 refresh cycles (or 378 ms) and was replaced by the word presented on the same According to a one-way analysis of variance carried out on the location, which was visible for one cycle (or 15 ms). The word was data of the four participant groups, there were significant differ- followed by the second mask (the inverse pattern of the first), after ences with respect to the proficiency scores, F(3, 76) ϭ 17.38, p Ͻ which it was presented again, and so on, with the duration of theword presentation increasing by one cycle at every alternation andthat of the mask decreasing by the same time (15 ms). The process stopped when the participant pressed the button (otherwise, the Means, Standard Deviations, and Ranges of Scores on theProficiency Test for the Four Participant Groups Number of Interlingual Homographs and Cognates in the Scores were calculated as the unweighted mean of the percentage correct for words (n ϭ 40) and for nonwords (n ϭ 20), respectively.
.001. Planned comparisons showed that the native speakers of English had higher scores than the three bilingual groups, as Intercorrelations of Log RTs on Correct Trials (Based on Item evident in a significant difference between the native speakers and Means) Between the Four Participant Groups the best of the bilingual groups, the French participants, t(38) ϭ5.23, p Ͻ .001. Furthermore, French participants had higher scores than German participants, t(38) ϭ 2.59, p Ͻ .02, but there was no significant difference between French and Dutch or between Ger- man and Dutch participants (both ps Ͼ .06).
All correlations are significant at the p Ͻ .01 level (n ϭ 1,025).
For all statistical analyses, the RTs were logarithmically trans- formed in order to avoid the “long right tail” of the skewed RTdistributions overly influencing the results. The mean RTs on three bilingual groups, we carried out a second set of analyses correct trials, error rates, and standard deviations per participant involving the three nonnative groups only. The resulting R2 values (indicating the proportion of shared variance) are also reported inTable 8.
Similarity of Item Means Between Participant Groups As can be seen from Table 8, a large proportion (around 60%) of the RT variance in each nonnative group could be explained by The first aim of the present study was to assess the extent to the item means of the other two bilingual groups. In contrast, the which the results of the three bilingual participant groups were predictability of the native data by the bilingual item means was comparable. A first indication of the similarity of their data is considerably reduced (R2 ϭ .49). Similarly, the regression weights given by the between-group correlations. The correlations between show that the item means of the English group were not as good a the three bilingual participant groups, as well as those with the predictor of the nonnative data as those of the other nonnative native control group, calculated across the 1,025 item means, are groups were. This result gives a first indication that whereas the three nonnative participant groups largely processed the English As can be seen from Table 7, the three nonnative groups words in the same way, the native group was less similar to the correlate to a larger degree with each other than with the native bilinguals. To investigate the similarities and differences between group. Pairwise comparisons of the correlation coefficients using the participant groups in more detail, we included a number of Fisher Z transformations confirmed this observation (all ps Ͻ .02).
within- and between-language variables in multiple regression For a different view of the similarity between the data of the four groups, we conducted a number of linear multiple regressionanalyses with the item means of one group as dependent variableand those of the other three groups as predictors, respectively. The Regression Analyses Including Within- and Between- resulting proportion of explained variance (R2) indicates the de- Language Predictors (Bilingual Participants) gree to which the outcomes of one group could be predicted by Because the focus of the present article was on bilingual word those of the other three groups, or in other words, how much recognition, we analyzed the data of the nonnative participants first variance is shared between the groups. Additionally, the standard- and compared the results with those of the native group as a second ized regression weights give an indication of the relative contri- step (described below). For the regression analyses, trials with RTs bution of the three predictor groups and the direction of this shorter than 500 ms or greater than 5,000 ms were removed from contribution. The resulting regression weights and R2 values are the data set. A linear mixed-effects analysis of covariance The first set of analyses involved all four participant groups; to obtain measures of how much variance was shared between the Table 8Results of the Multiple Regression Analyses Regarding theMutual Predictability of Log RTs (for Correct Trials) Between the Four Participant Groups, Calculated Across Item Means Mean RTs (Absolute and Logarithmic) for Correct Trials (inms), Error Rates, and Standard Deviations in the Progressive Demasking Task for the Four Participant Groups All R2 values and regression weights are significant at p Ͻ .001.
(Baayen, in press; Bates & Sarkar, 2005; Faraway, 2005; Pinheiro Control variables in our study were session, trial number (within & Bates, 2000; Wood, 2006) with the logarithmically transformed a session), previous RT, error (vs. correct response),5 and partic- RT as dependent variable and with crossed random effects for ipant group. Table 9 shows that with each successive session, subject and item was fit to the data, using a stepwise variable participants responded faster. Participants also recognized words selection procedure. The following variables did not reach signif- faster as they progressed through the trials in a session. We icance as predictors (i.e., their regression weights were nonsignif- included the RT to the preceding trial as a covariate, as previous icant) and were therefore dropped from the model: mean bigram research has suggested that difficult preceding trials have a spill- frequency, number of low-frequency English neighbors, total num- over effect on the next trial (Taylor & Lupker, 2001; Wurm, ber and frequency of English neighbors, number of meanings, Aycock, & Baayen, 2007). In our data, a longer RT to the previous noncognate homograph status, total number and frequency of L1 trial indeed implied a longer RT for the target trial. Finally, trials neighbors, number of high-frequency L1 neighbors, and number of for which participants reported a different word from the word low-frequency L1 neighbors. Furthermore, the interactions of all actually shown tended to have shorter RTs, suggesting that on remaining variables with participant group were tested but were these trials, the button was pressed before the word had been not significant unless reported otherwise. All p values were sup- identified correctly. The accuracy of the response also interacted ported by Markov chain Monte Carlo confidence intervals sampled with written word frequency and minimal bigram frequency (fur- from the posterior distribution of the predictors.4 Inspection of the residuals revealed marked nonnormality due to Written word frequency emerged as a nonlinear predictor, with higher estimation errors for a number of data points with very long facilitatory effects on RTs. Nonlinearities for word frequency were RTs. After removing overly influential outliers (defined as having also reported by Gordon (1985), Balota et al. (2004), and Baayen absolute standardized residuals exceeding 2.5) from the data set, et al. (2006). Across these studies, the effect of frequency leveled we refitted the model. We report only the significant fixed effects off for the higher frequency words. Additionally, and in line with for this final model (see footnote 6 for the random effects). Note the monolingual findings by Baayen et al. (2006), spoken fre- that because the semantic variables (concreteness, familiarity, and quency also had an independent, facilitatory effect on RTs, but meaningfulness), as taken from the MRC database, were available only as a linear predictor (the quadratic predictor was nonsignifi- for only a subset of words, those predictors were tested in a cant and therefore excluded from the model). Apparently, the separate analysis. The first analysis was carried out on the com- familiarity with a word in speech (besides its use in written plete nonnative data set, but excluding the set of semantic predic- language) is important not only for native speakers of English but tors. The resulting regression weights, t values, and significance also for nonnative speakers. Surprisingly, and possibly as a con- sequence of females having a better verbal memory than males(Kimura, 1999), the frequency effect was stronger for women thanfor men (whereas there was no main effect of sex). Finally, thefacilitation of written frequency was attenuated for error trials.
Table 9Results of the Mixed-Effects Regression Model Including the Together with the facilitatory main effect of error, this interaction suggests that words of low frequency sometimes led to faster butpremature erroneous responses, probably due to the confusion of the correct word candidate with a similar higher frequency word.
The morphological family size measure revealed a facilitatory effect, as expected. Furthermore, words with more than one entry in the CELEX lexical database were processed faster than words with only one entry. Compared with native speakers of Dutch, this effect was significantly reduced for French but not for German Although extreme multicollinearity characterized our original full set of predictors, the condition number (following Belsley, Kuh, & Welsch, 1980) for the item variables in our final model was 32. A condition number this high indicates substantial collinearity. However, exclusion of number of letters from the set of predictors reduced the condition number to 15.
Furthermore, the coefficients of a model excluding number of letters were very similar to those in the model including this variable. (The correlation between the two sets of coefficients was .9999.) This allows us to conclude that multicollinearity is not distorting our regression model.
5 We opted to include error trials in the model (along with the error variable) in order to keep as many data as possible, and also because anypossible effects of the lexical variables on error trials might provide valuable insights into the processes at work when an error occurs. An Additional effect for the French compared with the Dutch group.
ditional effect for men compared with women.
analysis of only the correct trials showed very similar results, with the same error trials compared with correct trials.
As the only cross-language effect that emerged across the com- p values). These results are comparable to those reported by plete data set, words that are cognates in the participant’s native Baayen et al. (2006) for the familiarity ratings obtained by Balota language elicited shorter response latencies than noncognates.
et al. (2001). The ratings for meaningfulness can be predicted from Word length, measured in number of letters, was inhibitory as spoken and written frequency, concreteness, and the number of expected. Similarly, minimum bigram frequency had an inhibitory CELEX entries. The adjusted R2 for this model was .12.
effect on RTs, reflecting the difficulty of unambiguously identify- To avoid problems with increased multicollinearity, we included ing more typical (and thus easily confusable) English words. This the residuals of the models for familiarity and for meaningfulness effect confirms the findings of Westbury and Buchanan (2002) for as predictors in our mixed-effects model for the response latencies high-frequency words in monolingual lexical decision. However, in PDM. These residuals are thus corrected for the influence of all for error responses, the beta weight was reduced. Analogous to the variables correlated with the original familiarity and meaningful- interaction of error with word frequency, this interaction shows ness measures. Both residual familiarity and residual meaningful- that (owing to their high confusability) difficult words (i.e., words ness had significant facilitatory effects on RTs: familiarity (␤ ϭ with a high minimum bigram frequency) in some cases elicited –.0002), t(38800) ϭ –1.99, p Ͻ .05; meaningfulness (␤ ϭ –.0001), relatively fast but incorrect responses. Among the variables coding t(38800) ϭ –2.38, p Ͻ .02. Meaningfulness also interacted with for English orthographic neighborhood, only the number of higher participant group, with a stronger facilitatory effect of this predic- frequency English neighbors was significant: Words with many of tor (by –.0001), t(38800) ϭ –2.86, p Ͻ .01, for the French as these neighbors elicited longer RTs. This replicates what has been opposed to the Dutch group. In contrast, the effect of meaningful- observed for the PDM task in monolingual studies before and ness for German participants did not differ significantly from that confirms the notion of higher frequency neighbors being a better for Dutch speakers ( p Ͼ .20). All other effects that had evolved in predictor of word recognition performance than the total number the previous model on the complete data set (apart from the effect of the number of CELEX entries) were present for this subset of Even though (noncognate) homograph status was not significant the data as well. Note that these bilingual data did not provide as a dichotomous variable across the complete data set, we ran an evidence for any main effects (e.g., bigram frequency) disappear- additional analysis to examine whether evidence for a possibly ing once familiarity was controlled for (or included in the regres- hidden effect was nevertheless present in our data. For a subset of sion model), as claimed by Gernsbacher (1984).
items (see Table 4) including only those words that were noncog-nate homographs with the respective L1, we investigated whetherthe frequency of the L1 reading had an influence on RTs. Indeed, Control Group: Native Speakers of English in a regression model simplified owing to the small data set, with To investigate whether the monolingual English speakers devi- written and spoken English frequency, number of letters, and ated from the bilinguals in terms of the effects of the lexical number of English higher frequency neighbors as lexical predic- variables, we used the same model that had emerged as the one tors, the frequency of the homographs in L1 had a significant providing the best fit to the bilingual data (but, of course, exclud- inhibitory effect on RTs (␤ ϭ .013), t(3742) ϭ 2.20, p Ͻ .03. This ing cross-language variables and the variable participant group) for replicates the findings of previous studies reporting that effects of the group of native English speakers. Most predictors included in interlingual homographs depend on the relative frequencies of the the original model were found to be significant again, with the readings in the two languages (de Groot et al., 2000; Dijkstra et al., following exceptions: The facilitatory effect of morphological 1998; Kerkhofs, Dijkstra, Chwilla, & de Bruijn, 2006).
family size failed to reach significance ( p ϭ .30); there was nosignificant difference between words with only one and those with more than one entry in CELEX ( p ϭ .66); and the interactions oferror with minimal bigram frequency ( p ϭ .08) and with written or The influence of the semantic variables taken from the MRC spoken frequency (both p Ͼ .15) were no longer significant. Of database was investigated in a second analysis, on 659 out of the most interest, however, written frequency was no longer signifi- 1,025 words. The same variables that had emerged as significant cant for native speakers of English (linear: p ϭ .32, quadratic: p ϭ predictors in the previous analysis were entered in this analysis,with the addition of the three semantic variables concreteness,familiarity, and meaningfulness. However, in a subsequent step, 6 We complete the specification of the mixed-effects model with a report the interaction of participant group with number of CELEX entries of the random-effects structure. The model contained five random effects.
was dropped from the model, because it was not significant any A random effect concerns a random variable that follows a normal distri- more for this subset ( p Ͼ .30). Similarly, and in line with what we bution and that is characterized by a mean equal to zero and an unknown expected in the light of previous findings, concreteness did not standard deviation. For each of the five random effects in our model, we have a significant effect on RTs ( p Ͼ .80) and was excluded from therefore have a standard deviation that characterizes its spread around zero. The standard deviation for the random effect of word was 0.052—that Familiarity and meaningfulness are rating measures that them- is, the by-word adjustments to the intercepts are characterized by thisstandard deviation. Our model included three random effects involving the selves can be predicted from other variables in our data. More than participants: by-participant random intercepts (␴ ϭ 0.164) and by- half of the variance in the familiarity ratings (adjusted R2 ϭ .60) participant random slopes for written frequency (␴ ϭ 0.006) and number is explained by an ordinary least squares regression model with of letters (␴ ϭ 0.021). Centering of the data points showed that parameters familiarity as dependent variable and spoken frequency, number of for the correlations between random effects were not necessary ( p Ͼ .10).
higher frequency English neighbors, concreteness, and meaning- All random intercepts and slopes were supported by likelihood ratio tests fulness as (significant) predictors (see Appendix A for the ␤, t, and .43), and spoken frequency (linear and quadratic) remained as the Correlations and Mutual Predictability Between Groups only significant frequency predictor. Apparently, written fre-quency is an important factor only for nonnatives who are primar- All analyses that were carried out to investigate aspects of the ily exposed to written English, but not for native speakers. These first issue, the generalizability of L2 word recognition results insignificant effects were excluded from the model for this group across different bilingual groups, showed that the overlap between the results of the three nonnative participant groups was consid-erable. Descriptively, this is already evident from the high corre- All other effects that had been observed for the second language lations of the item means between these three participant groups speakers were present in the native data as well. Other variables (above .70; see Table 7). Furthermore, linear multiple regression that had been dropped from the bilingual model because of non- analyses (see Table 8) showed that far more than 50% (56%– 62%) significance also turned out to be nonsignificant for the native of the item variance in each group could be explained by the data speakers. Furthermore, the same random-effects structure was found to be present. The table of results of the final model for the This high degree of similarity between the three bilingual monolingual English speakers can be found in Appendix B.
groups is surprising, given that a number of differences between Similarly, the effects of the three semantic variables on the these groups had been expected, as specified in the introduction.
subset of the monolingual data for which these measures were One possible implication of this result is that during reading in a available very much resembled those observed for the bilinguals: second language, word recognition is mainly determined by factors Although the facilitatory effect of concreteness just failed to reach within that language itself, and that it should thus be similar for all significance ( p ϭ .054), residualized familiarity and meaningful- users of this language. However, in contrast to the large overlap of ness both had significant facilitatory effects on RTs: familiarity the nonnative groups, the group of native English speakers differed (␤ ϭ –.0002), t(12938) ϭ –2.71, p Ͻ .01; meaningfulness (␤ ϭ considerably from the bilinguals. Apparently, although the specific –.0002), t(12938) ϭ –3.54, p Ͻ .001. This time, two effects turned native language of a bilingual does not make much of a difference, nonsignificant in this subset analysis, probably as a consequence of second language processing seems to be different from native the reduced statistical power in the data subset: that of the qua- language processing. A more detailed investigation of the specific dratic term of spoken word frequency (␤ ϭ .009), t(12938) ϭ 1.72, differences between the participant groups with respect to English p ϭ .08, and the interaction of sex and spoken word frequency word recognition was provided by the regression analyses involv- (␤ ϭ .003), t(12938) ϭ 1.66, p ϭ .10.
ing a large number of word characteristics as predictors.
We now discuss some selected effects of individual variables In recent years, many studies have demonstrated that bilinguals that were novel or unexpected, or that differed across the partici- are influenced by knowledge of their native language during word pant groups, before we return to the question of the similarity of recognition in their second language. However, little is known the data across the groups and the question of the impact of about whether the findings concerning bilingual word recognition between-language influences during lexical access in L2.
are generalizable across bilinguals with different language combi-nations, especially for languages that share the same alphabet. Thefirst aim of the present study was to investigate this issue by comparing word recognition in English as a second language for When we compared the frequency effects for the three bilingual three groups of participants that had different western European groups, no significant differences emerged. All three groups were native languages. Additionally, this comparison included a control influenced by both written and spoken word frequency to compa- group of native speakers of English, to allow us to relate the results rable extents. We had hypothesized that differences in the fre- for the bilinguals to the “standard” case of language processing by quency effect across the groups might occur as a consequence of monolinguals and to directly compare L1 and L2 processing.
the different orthographic depths of the participants’ native lan- Furthermore, most previous studies have used factorial designs and guages and of the resulting differences in routes of lexical access special categories of stimulus words in order to create the most (Frost, 1994; Frost et al., 1987; Katz & Frost, 1992). The absence favorable conditions for cross-language effects to occur. Our sec- of such differences in our data indicates that the nonnative partic- ond goal was to examine whether these cross-language interactions ipants used very similar reading strategies in their common second still play a vital role in the context of a large, representative, and language, English. It is interesting to note, however, that the effect unbiased set of word materials, and with the simultaneous influ- of written frequency completely disappeared for the native group.
ence of factors within the target language itself. If bilingual word Only spoken frequency influenced the speed with which these recognition is indeed profoundly nonselective with respect to speakers recognized words in their native language. This differ- language, between-language variables that have previously been ence in the relative importance of written and spoken word fre- found to affect recognition performance should account for a quency between native and nonnative speakers reflects the way in considerable amount of variance in the present data.
which these language users are exposed to English: Compared We now discuss the results of the first set of analyses (correla- with native speakers, the nonnatives’ experience with English is to tions and mutual predictability between groups) and those of the a large extent based on reading and writing, with relatively little multiple regression analysis separately, before evaluating the find- exposure to spoken English. The emergence of spoken frequency ings more generally in the light of the two issues at hand.
as a predictor that is superior to written frequency for native speakers confirms and extends the findings by Baayen et al. (2006) Semantic and Syntactic Ambiguity Variables and questions the present common practice to use only written The results for the bilingual participants for the two ambiguity word frequency measures in psycholinguistic research.
variables included in the present analyses showed a facilitatory An additional analysis on the complete data set (including native effect of the number of word entries—that is, syntactically ambig- and nonnative speakers) with a new factor, nativeness, revealed uous words were recognized more quickly. Thus, when recogniz- that whereas the effect of spoken frequency did not differ signif- ing isolated words, the bilinguals seemed to benefit from multiple icantly between native and nonnative speakers (linear: p Ͼ .90, syntactic representations, possibly owing to the coactivation of quadratic: p Ͼ .45), the effect of written frequency was signifi- both readings, as demonstrated by Elston-Gu¨ttler and Friederici cantly larger for nonnatives (linear: p Ͻ .001, quadratic: p Ͻ .01).
(2005). Whereas the effect was smaller for the French compared Taken together, these results indicate that the total effect of word with the Dutch and German participant groups, it was completely frequency is larger for nonnatives than for natives. This difference absent for native speakers. This is in line with Elston-Gu¨ttler and in the size of the frequency effect for first and second language Friederici’s observation of a more effective disambiguation mech- speakers cannot be caused by the orthographic depth of the native anism in native as opposed to nonnative speakers. Apparently, the language: With English being the deepest among the four orthog- more proficient the speakers are, the less they show an effect of raphies (Seymour et al., 2003), the frequency effect should, if multiple syntactic representations. With the French group being anything, be largest for the group of English native speakers.
the most proficient among the three bilingual groups and the Another possibility is that frequency effects are generally larger natives obviously being even more proficient than the French, this for second as compared with first language speakers. Indeed, van would account for the observed effect patterns.
Wijnendaele and Brysbaert (2002) and de Groot et al. (2002) The second ambiguity variable, number of unrelated meanings, observed a larger frequency effect in standard word naming in L2 did not have a significant effect on the RTs for any of the when compared with L1. Such a difference can be accounted for in participant groups in this experiment. This is in contrast with a terms of the nonlinear form of the frequency effect, with an large body of literature that has demonstrated effects of the number increased frequency sensitivity in the lower frequency range; for of meanings on word recognition (e.g., Borowsky & Masson, unbalanced L2 speakers, the subjective frequency distribution of 1996; Kellas, Ferraro, & Simpson, 1988). The discrepancy might L2 words is likely to be shifted toward the most sensitive left be due to task differences, as most of the existing literature has region, when compared with monolinguals.
made use of the lexical decision paradigm. Note that it has beenshown before that semantic ambiguity effects are modulated by task requirements (Hino, Lupker, & Pexman, 2002). Furthermore,the present ambiguity count was different from that in the majority With respect to the quite recent finding that word recognition of previous studies; in fact, in many of them, the ambiguity rate is influenced by morphological family size, the three bilingual variable is a mixture of the number of unrelated word meanings, groups showed a similar pattern of results as in the lexical decision the number of related word senses, and the number of syntactic task in Schreuder and Baayen (1997) and in Dijkstra et al. (2005): roles a word can take. For instance, more than 40% of the “am- Morphological family size facilitated RTs above and beyond the biguous” stimuli of Experiment 2 in the frequently cited study by effect of word frequency. For the native control group, however, Borowsky and Masson (1996) have only one unrelated meaning, the effect was not significant. Similarly, Schreuder and Baayen according to The Wordsmyth Dictionary. The roles of the different also failed to find an effect of morphological family size in the forms of lexical ambiguity are at present unclear. For the present PDM task. Possibly, the effect of family size is magnified in a task and populations, though, we can conclude that the number of second language owing to an increased sensitivity to the number of strictly unrelated meanings did not influence word recognition in occurrences of a word, similarly to what we observed for effects of Number of Higher Frequency English Neighbors The results with respect to the three semantic variables show In line with previous results obtained with the PDM paradigm, that whereas concreteness had no significant effect on the RTs, we found that the number of higher frequency English neighbors meaningfulness and familiarity both facilitated them. These find- slowed down RTs for both native and nonnative participants, ings are noteworthy in several respects. First, the results indicate whereas other neighborhood measures (total number of neighbors that even a primarily perceptual task like PDM involves some or their cumulative frequency) did not significantly affect RTs.
degree of semantic processing; moreover, this is the case not only According to the multiple read-out model (Grainger & Jacobs, in L1 but also in L2, where semantic processing might be thought 1996), the PDM task is especially sensitive to lateral inhibition by to be weaker and/or slower than in L1 (Ardal, Donald, Meuter, strong competitors, because the target word has to be unambigu- Muldrew, & Luce, 1990; Kotz & Elston-Gu¨ttler, 2004). Second, ously identified (unlike in the lexical decision task, for example).
the semantic ratings from the MRC database are taken from first, This requires the target word representation to pass an activation not second, language speakers of English; nonetheless, meaning- threshold, which is delayed by the simultaneous activation of fulness and familiarity influenced both native and nonnative high-frequency competitors. Lower frequency neighbors cause speakers. Apparently, the variables captured word aspects that are relatively little lateral inhibition and therefore have little effect on universal across speakers with different language backgrounds.
the activation rate of the target word (unlike in lexical decision, Among the three bilingual groups, the results also showed that where global lexical activation is another factor determining RTs).
the French participants were more strongly affected by meaning- fulness than the Dutch and the German group. This time, this interaction is in line with the concept of orthographic depth, inparticular, with the claim that speakers of orthographically deep In summary, the detailed picture of the effects of a host of languages show enhanced semantic processing (Cuetos & Barbo´n, lexical variables provided a number of important insights on recent 2006; de Groot et al., 2002; Katz & Feldman, 1983): The bilingual and unsettled issues in the mono- and bilingual word-recognition group with the deepest L1 orthography, French, showed the largest literature—for example, the findings concerning the relative im- effect of a purely semantic variable, meaningfulness. Thus, the portance of spoken relative to written word frequency in L1 and data suggest that although semantic processing in English was very L2, the difference in the size of the frequency effect both between similar across speakers of different native languages, there was men and women and between native and nonnative speakers, the some modulation of the size of semantic effects that is in agree- lack of an effect of lexical ambiguity, and the greater relevance of ment with the orthographic depth hypothesis.
minimal as opposed to mean bigram frequency. Moreover, manyeffects that have previously been reported for (usually English-speaking) monolingual populations and for different experimental tasks have now for the first time been shown to generalize to L2speakers with different L1s and to the PDM task (e.g., effects of Previous studies have indicated that the two languages of a morphological family size, higher frequency orthographic neigh- bilingual interact during word recognition in a given language. In bors, and familiarity and meaningfulness).
these studies, two classes of variables have usually been used as In general, the multiple regression analysis confirmed the pic- indicators for cross-lingual interaction: cross-language ortho- ture obtained in the analyses before: The similarities across the graphic neighborhood variables and interlingual homographs and three bilingual groups were substantial. Among the 20 variables orthographically identical cognates. Both clusters of variables that were included in the regression, only 2 (number of CELEX were included in the present study as well.
entries and meaningfulness) affected the three groups in signifi- The data showed that, similar to the results of de Groot et al.
cantly different ways; in both cases, this interaction was caused by (2002), none of our L1 neighborhood measures (number of high- a difference in the size of the effect, not in the direction of it.
and low-frequency neighbors, total number and summed frequency Besides the large overlap between the groups, the multiple of neighbors) turned out to be a significant predictor of response regression analyses also showed that the degree of cross-language latencies. Thus, in the context of the many within-language influ- interaction in bilingual word recognition was limited. Among the ences already present in the model, there was no evidence of between-language variables, only cognate status had an overall cross-language neighbors from the participants’ native language measurable effect. Apparently, across a large and unbiased set of becoming active upon the presentation of the English target word.
words from the second language, bilinguals processed only words In contrast, a significant effect of cognate status was observed: that overlap in all respects between the two languages differently Bilinguals recognized cognates faster than noncognates, with sim- from other words. If it is assumed that cognates share the same ilar effect sizes across all three bilingual groups. However, no such orthographic and/or semantic representation in the bilingual lexi- effect was found for noncognate homographs or false friends. This con (Gollan et al., 1997; Sa´nchez Casas et al., 1992), the cognate pattern of results is consistent with the picture in the literature, as effect might be explained even without the online activation of the summarized in the introduction: Cognate effects are robust, native language: The effect would then simply be a consequence of whereas effects of interlingual homographs are more variable and the higher cumulative frequency of the joint representation.
sometimes even reverse (see also Dijkstra et al., 1999, for effects However, whereas all other between-language effects on the of cognates and homographs in a PDM task). However, note that whole data set were nonsignificant, a closer look at a subset of the number of noncognate homographs among the stimuli was words revealed that the native language must have been activated.
relatively small (about 60 out of 1,025 words, depending on the For noncognate homographs or false friends, the frequency of their respective L1; see Table 4). In contrast to the balanced designs reading in the native language had an inhibitory effect on RTs.
used in the described factorial studies, the imbalance between This is in line with previous studies and supports the view of homographs and nonhomographs might in this case complicate the bilingual lexical access being principally nonselective with respect comparison between these two categories. Therefore, we con- to language. However, the fact that homograph status had no effect ducted an additional analysis, including only those words that are when analyzed across the complete data set shows that globally, noncognate homographs in the participants’ L1, and investigated whether an English word had the same spelling as a word from the whether the word’s frequency in the L1 had an effect on the speed participants’ L1 did not play an important role.
with which it was recognized. Indeed, for this small set of words, The regression analyses also provided some important and sur- there was an inhibitory effect of L1 frequency on RTs. This result prising insights concerning the comparison of L1 and L2 speakers.
confirms the general pattern observed in previous factorial studies: On first sight, the pattern of effects found for natives in the The recognition of noncognate interlingual homographs in L2 regression analysis was fairly similar to that of the nonnatives, depends on their frequency in L1, with, in particular, a high L1 with the majority of significant predictors overlapping between the frequency making recognition more difficult. Thus, although a two groups. However, a closer investigation revealed a number of microanalysis was able to confirm the results of previous studies differences as well: Whereas both written and spoken frequency with factorial designs, it also became apparent that in the grand independently affected word recognition times in nonnative speak- picture of a large representative sample of English word stimuli, ers, only spoken frequency played a role for native speakers.
interlingual (noncognate) homograph effects do not play an im- Moreover, the total effect of word frequency was larger for non- portant role during word recognition.
native than for native speakers. Furthermore, the effects of mor- phological family size and the number of syntactic representations participants with an orthographically deep mother tongue. This disappeared for native speakers. Generally, this pattern of differ- might be a consequence of speakers of orthographically deep ences shows that compared with native speakers, nonnative speak- languages being accustomed to basing their responses on the full ers are more sensitive to the number of occurrences of a word (i.e., (including the semantic) word representation rather than on the how often the word itself occurs or in how many derivations and orthographic code alone. Of course, whether other language com- syntactic classes it occurs). Presumably, this increased sensitivity binations with larger contrasts in orthographic depth (e.g., includ- is a consequence of the lower subjective frequency of words in an ing non-European languages) or a different target language than L2: Small changes in the number of times or ways a word is English might give rise to increased differences in lexical process- encountered have the largest consequences when the degree of ing, thus lending more support to the orthographic depth hypoth- experience with the given word is low, which tends to be the case esis, remains to be tested. At this point, our data show that for for L2 words more than for words from one’s mother tongue.
native speakers of different western European languages having Thus, the results show that although the overlap between native English as an L2 (just like, as mentioned in the introduction, 38% and nonnative speakers of English is still large, the specific way in of the people within the European Union), L2 processing is very which within-language word variables that are related to frequency similar and largely unaffected by L1.
and ways of occurrence affect performance is different for L1 and One reason for the large degree of overlap of the different L2 word recognition, suggesting that the organization and function bilingual participant groups might be that proficient bilinguals are of the L1 and L2 language systems is not completely identical.
able to adapt to the requirements of a second language, eventhough these might differ from those of their native language. In this vein, de Groot et al. (2002) claimed that Dutch–Englishbilinguals performed Dutch and English word-naming tasks in Three major conclusions can be drawn from the present results.
different ways, owing to the difference in orthographic depth First, bilingual speakers of different native languages process L2 between the two languages. Similarly, Meschyan and Hernandez words in largely the same way. Second, and in line with the first (2006) reported that in English–Spanish bilinguals, different brain conclusion, the extent of cross-language influences in L2 word regions were involved while they were reading orthographically recognition, when investigated across a large, unbiased set of transparent Spanish words as compared with less transparent En- words, is small. Finally, the data indicate that word recognition in glish words. Such an adaptation of bilinguals to the most appro- L1 and L2 differ primarily with respect to the sensitivity to priate lexical access strategies of the target language is likely to frequency-related variables (such as written vs. spoken frequency, develop with increasing L2 proficiency. Further research is needed morphological family size, and number of syntactic categories).
in order to clarify whether less proficient participants with varying Thus, even though there are not many L1-specific effects on L2 native languages would display larger differences in L2 word word recognition, L2 speakers differ from monolinguals in terms recognition than the present groups did.
of frequency-related aspects of the organization of their language As a second important main finding, the limited role of between- language predictors evident in the present data qualifies and ex- The high degree of similarity between the three bilingual groups tends previous findings regarding the selectivity of lexical access is in contrast with what had been expected in terms of an extension in bilinguals. Our results confirm the existence of some cross- of the orthographic depth hypothesis. Even though the set of language interactions as shown by factorial studies before, and languages that we tested admittedly does not include the full thus indicate that the bilingual word-recognition system is not variation in orthographic depth (not including extremely deep principally divided by language and that its architecture allows for languages like Arabic or Chinese, for example), it still covers a coactivation of both languages. However, the present analyses also representative range of orthographies when considering European show that for a word recognition task in a single target language alphabetic languages, as the present study intended (given that involving a proficient bilingual population, a substantial part of the these are still the most intensively studied languages in experi- data can already be accounted for by characteristics of the target mental psycholinguistic research). Such variations are thought to language itself, with only a small or barely present additional cause differences in lexical processing, which become visible, for contribution of cross-language influences. This led to only mini- instance, in the effects of word frequency, word length, bigram mal processing differences between participants with different frequency, or semantic variables (de Groot et al., 2002; Frost et al., 1987; Ziegler et al., 2001). Although this has primarily been The largest group difference we found in the present data was argued for monolingual speakers, the question arises how bilingual the one between first and second language speakers, which indi- speakers with differently deep native languages deal with a second cates that L2 processing (regardless of specific cross-language language. For instance, Wang et al. (2003) demonstrated differ- influences) is fundamentally different from word processing in L1.
ences in English word recognition between native speakers of This is in accordance with Grosjean’s (1989) warning, “Neurolin- Korean (with a shallow orthography) and Chinese (with a deep guists, beware! The bilingual is not two monolinguals in one writing system). In our data, the only indication for orthographic depth of L1 affecting word recognition in English was the effect ofmeaningfulness, which differed for the native speakers of French compared with the other bilingual groups in a way that is in linewith the concept of orthographic depth. This finding suggests that Akamatsu, N. (2002). A similarity in word-recognition procedures among even though there was an effect of semantic variables for all second language readers with different first language backgrounds.
participants, the degree of their involvement was stronger for Applied Psycholinguistics, 23, 117–133.
Alario, F.-X., Ferrand, L., Laganaro, M., New, B., Frauenfelder, U. H., & Cuetos, F., & Barbo´n, A. (2006). Word naming in Spanish. European Segui, J. (2004). Predictors of picture naming speed. Behavior Research Journal of Cognitive Psychology, 18, 415.
Methods, Instruments, & Computers, 36, 140 –155.
de Groot, A. M. B., Borgwaldt, S., Bos, M., & van den Eijnden, E. (2002).
Allen, P. A., Smith, A. F., Lien, M.-C., Grabbe, J., & Murphy, M. D.
Lexical decision and word naming in bilinguals: Language effects and (2005). Evidence for an activation locus of the word-frequency effect in task effects. Journal of Memory and Language, 47, 91–124.
lexical decision. Journal of Experimental Psychology: Human Percep- de Groot, A. M. B., Delmaar, P., & Lupker, S. J. (2000). The processing of tion and Performance, 31, 713–721.
interlexical homographs in translation recognition and lexical decision: Andrews, S. (1997). The effect of orthographic similarity on lexical re- Support for nonselective access to bilingual memory. Quarterly Journal trieval: Resolving neighborhood conflicts. Psychonomic Bulletin & Re- of Experimental Psychology, 53A, 397– 428.
de Jong, N. H., Schreuder, R., & Baayen, R. H. (2000). The morphological Ardal, S., Donald, M. W., Meuter, R., Muldrew, S., & Luce, M. (1990).
family size effect and morphology. Language and Cognitive Processes, Brain responses to semantic incongruity in bilinguals. Brain and Lan- Dijkstra, T., Grainger, J., & van Heuven, W. J. B. (1999). Recognition of Baayen, R. H. (in press). Analyzing linguistic data: A practical introduc- cognates and interlingual homographs: The neglected role of phonology.
tion to statistics. Cambridge, England: Cambridge University Press.
Journal of Memory and Language, 41, 496 –518.
Baayen, R. H., Feldman, L., & Schreuder, R. (2006). Morphological Dijkstra, T., Moscoso del Prado Martı´n, F., Schulpen, B., Schreuder, R., & influences on the recognition of monosyllabic monomorphemic words.
Baayen, R. H. (2005). A roommate in cream: Morphological family size Journal of Memory and Language, 55, 290 –313.
effects on interlingual homograph recognition. Language and Cognitive

Source: http://gsite.univ-provence.fr/gsite/Local/lpc/dir/grainger/lemhofer.pdf