orthographic neighborhood effects in bilingual word

Orthographic Neighborhood Effects in Bilingual Word Recognition

Walter J. B. van Heuven and Ton Dijkstra

Nijmegen Institute for Cognition and Information (NICI), University of Nijmegen, Nijmegen, The Netherlands

and

Jonathan Grainger

CNRS and Universite´ de Provence, Aix-en-Provence, France

A series of progressive demasking and lexical decision experiments investigated how therecognition of target words exclusively belonging to one language is affected by the existence oforthographic neighbors from the same or the other language of bilingual participants. Increasingthe number of orthographic neighbors in Dutch systematically slowed response times to Englishtarget words in Dutch/English bilinguals, while an increase in target language neighborsconsistently produced inhibitory effects for Dutch and facilitatory effects for English targetwords. Monolingual English speakers also showed facilitation due to English neighbors, but noeffect of Dutch neighbors. The experiments provide evidence for parallel activation of words inan integrated Dutch/English lexicon. An implemented version of such a model making theseassumptions, the Bilingual Interactive Activation (BIA) model, is shown to account for theoverall pattern of results. © 1998 Academic Press

One of the striking features of bilingual lan-guage performance is the apparent ease withwhich the bilingual manages to keep interfer-ence from the non-target language at a minimallevel. The fact remains, however, that interfer-ence from one language to the other does occurand is observable with respect to both languagestructure and linguistic processing. For exam-ple, in language production, interference fromthe first language can be noticed both at thephonological level (foreign accents) and at the

sentence level (borrowed syntax), as well as inintrusions of words from the other language(accidental lexical borrowings).

In the area of language comprehension re-search, there has been a debate for several de-cades about how exactly cross-language inter-ference effects relate to the way words fromdifferent languages are stored and processed.With respect to lexical storage, a distinction hasbeen made between the hypotheses of languageindependent and language dependent lexica.The first hypothesis proposes that bilingualspossess one integrated lexicon for both theirlanguages, while the second assumes two sepa-rate lexica, one for each language. With respectto on-line processing, an analogous contrast hasbeen made between language selective and non-selective access views. Given that all contem-porary models of visual word recognition as-sume a one-to-many mapping from inputrepresentation to lexical representations inmemory (e.g., Forster, 1976; McClelland &Rumelhart, 1981; Paap et al., 1982), the issue iswhether words of both languages are activated(or considered as candidates) during word rec-

The writing of this paper was supported by a grant fromthe NFS-Fund for Dutch–French Collaboration (Nether-lands Organization for Scientific Research, NWO, TheNetherlands) to the Nijmegen Institute of Cognition andInformation (NICI). The authors thank Herbert Schriefersand three anonymous reviewers for their constructive com-ments on an earlier version of this paper. They also thankPeter van de Pol and Sander Martens for their help inrunning Experiments 3 and 4. Finally, they express theirgratitude to the MRC Applied Psychology Unit, Cambridge,UK, for putting their experimental facilities at the authors’disposal.

Address correspondence and reprint requests to WalterJ. B. van Heuven, NICI, University of Nijmegen, P.O. Box9104, 6500 HE Nijmegen, The Netherlands. E-mail:[email protected].

4580749-596X/98 $25.00Copyright © 1998 by Academic PressAll rights of reproduction in any form reserved.

JOURNAL OF MEMORY AND LANGUAGE 39, 458–483 (1998)ARTICLE NO. ML982584

ognition, or only words belonging to the tar-geted language.

Restricting ourselves to those stages of bilin-gual word recognition involved in word-formaccess, four basic hypotheses can be distin-guished with respect to lexical organization andon-line processing. These are represented inFigure 1.

Early work in this field provided support forhypothesis A in Figure 1, which combines theindependent-lexica hypothesis with that of lan-guage selective access determined by languagemode information (Macnamara & Kushnir,1971; Scarborough, Gerard, & Cortese, 1984;Soares & Grosjean, 1984). According to thisviewpoint, a selection mechanism, called an‘input switch‘, guides all incoming auditory orvisual information to the English lexical systemof the bilingual while (s)he is performing anEnglish monolingual task. The high selectivityof the system implies that the linguistic inputinitially (i.e., at the orthographic or phonologi-cal level) only contacts representations in onelanguage. If the lexical representation corre-sponding to the input is not found in the activelexicon, contact is established with the otherlexical system. A typical description of this typeof model is in terms of self-terminating search

procedures examining first the representationsin one lexicon followed by those in the other.The second hypothesis, represented in Figure1B, that of selective access to an integratedlexicon, is functionally equivalent to hypothesisA. Since words of the non-selected language arenever activated, they can have no influence onhow words in the selected language are pro-cessed. This amounts to having independentlexica.

According to hypotheses C and D, wordsfrom both languages that are partially compati-ble with the stimulus are activated in parallel.These two hypotheses are distinguished in theextent to which words from different languagesinteract once activated. Within the frameworkof search models of lexical access (e.g., Forster,1976), this is determined by whether or notsearch is organized by language. Hypothesis Cstates that search is organized by language, withwords of one language being checked beforethose of the other language. Hypothesis D statesthat search is language independent, words fromboth languages being examined as a function oftheir relative frequency. In an interactive acti-vation framework (McClelland & Rumelhart,1981), the integrated lexicon hypothesis postu-lates the existence of inhibitory connections be-tween words from different languages. The sep-arate lexica hypothesis, on the other hand, limitsinhibitory connectivity to within languages.

The discussion in the literature has focusedon testing between the selective-access, inde-pendent-lexica and the non-selective-access, in-tegrated-lexicon views (hypotheses A and D inFigure 1). Early evidence for a selective accessposition came from research comparing pure-language and mixed-language sentence compre-hension in bilinguals. Macnamara and Kushnir(1971) asked bilinguals to judge the truth ofwritten and spoken English and French mixed-language sentences. They observed that switch-ing languages within a sentence takes time com-pared to monolingual passages. More recently,Soares and Grosjean (1984), using a phoneme-triggered lexical decision task, demonstratedthat bilinguals responded slower than monolin-guals to nonwords in monolingual and bilingualspeech. Furthermore, they found that bilinguals

FIG. 1. Four theoretical viewpoints on bilingual wordrecognition models: (A) language-selective access, indepen-dent lexica; (B) selective access, integrated lexicon; (C)language non-selective access, independent lexica; (D) non-selective access, integrated lexicon. Black filled circles in-dicate inhibitory connections.

459NEIGHBORHOOD EFFECTS IN BILINGUALS

responded slower to code-switched words inbilingual speech than to the same words inmonolingual speech. They concluded that,when operating in a bilingual speech mode,participants search the base-language lexiconfirst, which induces longer recognition times forcode-switched words. A similar conclusion wasreached by Scarborough, Gerard, and Cortese(1984), who demonstrated that English/Spanishbilinguals responded to non-target (Spanish)words in an English lexical decision task as ifthey were nonwords (i.e., at approximately thesame speed as nonwords and without exhibitingfrequency effects). The bilingual participantsthus seemed to be capable of operating in anentirely language-specific manner without anyinfluence from their knowledge of the otherlanguage. In a similar vein, Gerard and Scar-borough (1989) found that the reaction times(RTs) of English/Spanish bilinguals in an En-glish lexical decision task were not affected bytheir knowledge of form-identical Spanishwords, suggesting that the bilinguals were infact functioning as English monolinguals. Thisresult, as well as the finding that word latencieswere primarily determined by frequency of us-age in the target language, was considered to beconsistent with the predictions of a language-specific access and separate-lexica model.

Evidence for non-selective access, on theother hand, has been collected in research usinginterference paradigms such as the Stroopcolor–word task (Dyer, 1971), the picture–wordinterference task (Ehri & Bouchard-Ryan,1980), or the flanker task (Guttentag, Haith, &Goodman, 1984), all of which provide evidencethat between-language interference can be justas great as within-language effects. Between-language interference has been repeatedly ob-served in the time to reject nonwords in a lexicaldecision task. Altenberg and Cairns (1983)found that the nonword rejection latencies ofbilinguals performing an English lexical deci-sion task were affected by the legality of thenonword in German (e.g.,pflok) just as much asits legality in English (e.g.,twoul). Furthermore,Nas (1983) observed that Dutch/English bilin-guals took longer to reject nonwords that wereorthographically but not phonologically identi-

cal to real Dutch words than nonwords withoutresemblance to Dutch words. Examples of thefirst kind of nonwords arebeeld(image, statue)or mank (crippled). In the second experiment,similar results were obtained for nonwords thatwere phonologically but not orthographicallysimilar to real Dutch words, such assnay,whichaccording to English spelling-to-sound rules ispronounced like the Dutch wordsnee(cut), orspailer, pronounced approximately like theDutch wordspeler(player).

The studies by Altenberg and Cairns (1983)and Nas (1983) provide evidence that non-tar-get language representations are activated dur-ing word recognition in bilinguals. Neverthe-less, their results only concerned correctnegative responses to nonword stimuli. Only afew studies have investigated cross-languageinterference effects in positive responses toword stimuli in non-primed recognition para-digms such as lexical decision (Grainger & Dijk-stra, 1992; Beauvillain, 1992). In the presentseries of experiments we will add to the lattertype of studies.

The main aim of the current set of experi-ments was to test both processing and organi-zational accounts of the bilingual mental lexi-con by manipulating both within- and across-language neighborhood density. Target wordsbelonged only to one language (i.e., there wereno cross-language homographs, homophones,or cognates). Cross-language interference ontarget word recognition was examined by vary-ing the number of orthographic neighbors of thetarget word in the non-target language (cf.Grainger & Dijkstra, 1992; Beauvillain, 1992).An orthographic neighbor is any word differingby a single letter from the target word, respect-ing length and letter position (Coltheart et al.,1977). Our study builds upon the empiricalfinding that among the set of lexical candidatesthat are activated during visual word recogni-tion, orthographic neighbors take a prominentposition. Target word identification and targetword naming have been shown to be sensitive tothe number of neighbors (neighborhood den-sity) and the frequency of such orthographicallysimilar words (e.g., Andrews, 1989, 1992; Car-reiras, Perea, & Grainger, 1997; Grainger et al.,

460 VAN HEUVEN, DIJKSTRA, AND GRAINGER

1989; Grainger & Jacobs, 1996; Grainger &Segui, 1990; Snodgrass & Mintzer, 1993;Ziegler, Rey, & Jacobs, in press). The presentexperiments use effects of number of ortho-graphic neighbors as an index of the relativeinfluence of non-target language words on tar-get word recognition in different experimentaltasks and conditions.

Returning to the bilingual situation, lan-guage-selective and non-selective access mod-els differ in their predictions about neighbor-hood effects across languages. As describedabove, in a selective access model, only wordsfrom the target language are activated (or con-sidered) on a given trial. If, in an experiment,target word presentation is blocked by languageand participants are informed of this blocking,we expect the input switch of the selective ac-cess model to always be set on the appropriatelexicon. In this case, the selective access modelpredicts that recognition of a target word isdetermined by the neighborhood characteristicsof the target language only. In a non-selectiveaccess model, on the other hand, sensory inputactivates words from both languages simulta-neously, and it therefore predicts neighborhoodeffects of both languages during the word rec-ognition process. It has recently been arguedthat such a language non-selective explanationis more flexible than a selective access hypoth-esis because it allows for different degrees ofactivation in the two languages, dependent onexperimental circumstances (Grosjean, 1997;Li, 1996). This view suggests that a neighbor-hood density manipulation could yield evidencefor non-selective access but that the degree ofeffect observed might be task-dependent.

With respect to the organization of the bilin-gual lexicon, non-selective access could occurwith either separate or integrated lexica (hy-potheses C and D in Figure 1). Because neigh-borhood density effects are assumed to ariseduring word identification, cross-language ma-nipulation of neighbors allows us to examinethis structural issue as well. According to anintegrated lexicon hypothesis, recognition of atarget word will be affected by the presence ofboth target and non-target language neighborsin situations where both languages are active

(bilingual language mode, Grosjean, 1997). Ac-cording to an independent lexicon hypothesis,recognition of the target word should not beaffected by interlexical neighborhood density,since there are no direct interaction effects be-tween the two lexica. In fact, manipulation ofneighborhood density may be one of the veryfew experimental means of testing the inte-grated vs independent lexica hypotheses.

Thus, only a non-selective-access integrated-lexicon model predicts that target word recog-nition will be influenced by orthographic neigh-bors from both the target and the non-targetlanguages. All other types of models describedhere predict no effects of non-target languageneighborhood density on target word recogni-tion. Of course, we cannot exclude that morecomplex model variants could be formulatedthat would make different predictions. Finally,we may consider whether orthographic neigh-bors from the same or different language as thetarget word will exert a facilitatory or an inhib-itory effect on target recognition. Because theexamination of neighborhood effects in bilin-guals is uncharted territory, the answer to thisquestion is unknown. However, whether anyneighborhood density effects observed will befacilitatory (as observed in the monolingual En-glish lexical decision task, by Andrews, 1989,1992) or inhibitory (as in the fragmentation taskof Snodgrass & Mintzer, 1993) is not directlyrelevant for the present purposes of setting con-straints on possible bilingual model types interms of lexical access and structure. Most rel-evant is the general agreement that effects ofnumber of orthographic neighbors in a givenlanguage, whether they be inhibitory or facili-tatory, reflect the influence of simultaneouslyactivated word representations from that lan-guage.

Four experiments were carried out using thesame stimulus materials in different tasks. Thisprovided us with a means of assessing cross-experimental transitions in RT-patterns result-ing from task and instruction differences. Thefirst two experiments involved a word identifi-cation paradigm, that of progressive demasking.While in the first experiment Dutch/English bi-lingual participants were presented with blocks


of English or Dutch target words, the secondpresentation involved a mixed presentation ofEnglish and Dutch words. Next, two visual lex-ical decision experiments were run. In the thirdexperiment, bilingual participants performed ageneralized lexical decision (say ‘yes‘ to anystring that is either a Dutch or an English word).In the fourth experiment, two groups of partic-ipants (Dutch/English bilinguals and Englishmonolinguals) made a lexical decision in re-sponse to English words and English-like non-words only. With respect to these last two ex-periments, not only the cross-language ortho-graphic similarity of the word stimuli wasconsidered, but that of the nonword stimuli wasas well.

STIMULUS SELECTION

Since all experiments to be reported use thesame set of stimulus words, it is convenient tofirst describe how the stimuli were selected.First, a list of English and Dutch four-letterwords was extracted from the CELEX database(Baayen, Piepenbrock, & van Rijn, 1993). Onlynouns, adjectives, verbs, and adverbs with aprinted frequency of at least 2 occurrences permillion (o.p.m.) were selected. This resulted ina list of 1,323 English words and a list of 982Dutch words. For each target word, the numberof neighbors in each language and their fre-quency were calculated (following Coltheart etal., 1977). Four conditions of items from eachlanguage were defined by orthogonally varyingthe number of neighbors in English (English N)and Dutch (Dutch N): (1) words with manyneighbors in the target language and many inthe non-target language (large English N andlarge Dutch N), (2) words with many neighborsin the target language (large English N or DutchN, depending on target language) and few in thenon-target language (small Dutch N or EnglishN, respectively), (3) words with few neighborsin the target language (small English N or DutchN) and many neighbors in the non-target lan-guage (large Dutch N or English N, respec-tively), and (4) with few neighbors in bothlanguages (small English N and small Dutch N).Each condition consisted of 20 words, mostlynouns and adjectives. All selected items were

purely monolingual (no cross-language homo-graphs or cognates were selected). Within eachlanguage, the target items in the four conditionswere matched for word frequency.

In addition, for both English and Dutch targetitems, the word frequency of the neighbors wasmatched as closely as possible over conditions.To realize this, the neighbors of the target wordswere divided into three word frequency catego-ries: low frequency (LF), neighbors with a fre-quency of less than 20 o.p.m.; medium fre-quency (MF), neighbors with a frequencybetween 20 and 50 o.p.m.; and high frequency(HF), neighbors with a frequency of at least50 o.p.m. The number of neighbors of eachfrequency category was approximately equalfor each test condition.1

We selected only English and Dutch wordsexpected to be known to the participant popu-lation that we intended to use for experimenta-tion. To check the list of potential English testwords, 11 students from a Dutch participantpopulation with low proficiency in English wereasked to examine the list and mark unknownitems. This resulted in a set of test stimuli inwhich the English words were about three timesas frequent as the Dutch words, reflecting thefact that the Dutch participants in Experiments1–4 were not perfectly balanced bilinguals. Thesubjective frequency of English words for theseDutch participants is expected to be much lowerthan for English native speakers, on which thefrequency statistics in the written CELEX-cor-pus are based. The descriptive statistics for theword stimuli in each condition are summarizedin Table 1, and the complete list of stimuli canbe found in the Appendix.

1 The mean number ofhigherfrequency neighbors acrossconditions was not explicitly matched. However, the num-ber of higher frequency neighbors correlated very highlywith the total number of neighbors. For Dutch targets acorrelation of 0.94 (N 5 8) was obtained, and for Englishtargets 0.97 (N 5 8). A further attempt was made to matchthe test conditions within and across languages with respectto bigram frequency, but, given the other matching criteria(target frequency, frequency, and number of neighbors) anddue to the correlation between neighborhood density andbigram frequency, this match was not perfect.


EXPERIMENT 1: PROGRESSIVEDEMASKING (BLOCKED

PRESENTATION)

Prior work using progressive demasking(PDM) and related techniques in the monolin-gual domain (Grainger & Segui, 1990; Grainger& Jacobs, 1996; Snodgrass & Mintzer, 1993;Carreiras et al., 1997) has demonstrated thesensitivity of the paradigm to the influence oforthographic neighbors on target word recogni-tion. In the PDM task, the presentation of thetarget word is alternated with that of a mask.During this process of alternation, the targetpresentation time slowly increases while that ofthe mask decreases. The participant’s task is topush a button as soon as the target word isidentified. Compared to other paradigms (suchas lexical decision), PDM reduces the rate ofpresenting the sensory information to the par-ticipant, thus effectively slowing the targetidentification process.

In the present experiment, two masks wereused which covered the entire word matrix. Onemask consisted of black and white blocks(checkerboard pattern), and the other was theinverse of it (black became white and whitebecame black). The two masks were presentedin turn. The mask presented on the first cycle ofeach trial was changed for each participant.Extensive pre-testing with various mask typesindicated that this alternating checkerboard pat-

tern was preferable to the usual row of hashmarks (####), since the former type of maskresulted in fewer identification biases for par-ticular letters and features (Jacobs & Grainger,1991).

Method

Participants.Forty-two Dutch students withnormal or corrected-to-normal vision partici-pated. Two groups of participants were selected,differing in their proficiency in English. Stu-dents in the High Proficiency (HP) group weremainly students of English or students who hadvisited an English speaking country for a shortperiod (6–12 months), while those in the LowProficiency (LP) group consisted of students ofDutch or other disciplines. The allocation ofstudents to these two groups was checked by theresults of a questionnaire that examined theirproficiency in more detail. All students werenative speakers of Dutch.

Stimuli and design.Participants saw twoblocks of items, one for each language. Stimu-lus selection was described above and can besummarized as follows. Each language blockconsisted of 80 items, 20 for each of the fourconditions defined by the factorial combinationof neighborhood density in Dutch and English(see Table 1). One group of participants waspresented with the Dutch block first, followedby the English block, and the other group saw

TABLE 1

Descriptive Statistics for the Stimuli Used in the Experiments

Neighbors

Word frequency

Number of neighbors

Dutch N English N Dutch English Total

Dutcha large large 18.6 3.50 3.50 7.00large small 18.7 3.50 1.00 4.50small large 18.6 1.00 3.50 4.50small small 18.6 1.00 1.00 2.00

Englisha large large 51.8 3.50 3.60 7.10large small 54.1 3.50 1.15 4.65small large 55.2 0.95 3.50 4.50small small 55.5 1.00 1.20 2.20

Note.Word frequency per million.a n 5 20 for each condition.


the English block first and then the Dutch block.Each block was preceded by 25 practice trials,and five dummy trials were placed at the begin-ning of each test block. Each participant saw adifferent randomized order of test items withina block.

Procedure.Presentation of the visual stimuliand recording of the RTs was controlled by anApple Macintosh IIcx microcomputer con-nected to an Apple 140 Trinitron monitor. RTswere measured at a 1 msaccuracy by a buttonbox connected to the computer. This box andthe experimentation software were developed incollaboration with the Technical Group of theNijmegen Institute for Cognition and Informa-tion. The words consisted of black Courier cap-ital letters (18 points or approximately 8 mm onscreen) presented at the center of the computerscreen on a white background. The monitor wasplaced 60 cm from the participant, in order toprovide projection within the fovea of the eye.

Participants were tested individually. Theywere told that they had to identify four-lettertarget words that would gradually appear on thecomputer screen out of a background of visualnoise. They were further informed that the ex-periment consisted of two blocks, first one withDutch words and then one with English words(or vice versa). Before each block started, theparticipants were given written instructions inthe language of the subsequent trials. They wereinstructed to react as soon as they identified atarget word but without making errors.

At the beginning of each trial the words‘‘NEXT WORD’’ (in the English block) or‘‘VOLGEND WOORD’’ (in the Dutch block)were presented. After the participant pressed abutton two small lines appeared, 6.6 mm (15pixels) above and below the center of thescreen. After 1500 ms, the screen was clearedand one of the two checkerboard masks waspresented at the center of the screen. Next, thetarget word appeared at the same position, fol-lowed by the other mask, and so on. In the firstcycle the mask appeared for 300 ms and wasfollowed by the target word, which was pre-sented for 15 ms. Then the other mask waspresented, but now for 285 ms, followed againby the target word for 30 ms, and so on. The

time that the mask was visible decreased, whilethe time that the target word was visible in-creased until the mask presentation time waszero. The PDM-cycling process continued untilthe participant pushed the response button or atime-out period of 6 seconds was reached.

Immediately after the participants hadpressed the button to indicate that they hadidentified the target word, this word was re-placed by a checkerboard backward mask. Atthe same time, a dialog box appeared with thewords ‘‘Enter the word’’ (in the English block)or ‘‘Tik het woord in’’ (in the Dutch block).After the participants entered the word that theyhad identified, the next trial started. Before thesecond block began the participants read theinstructions again, now written in the other lan-guage. After the experiment, the participantfilled out a questionnaire concerning his or herknowledge about and experience with the En-glish language, in order to check the allocationof participants to the two proficiency groups.The session lasted about thirty minutes.

Results

Mean RTs were computed for each partici-pant and for all test conditions in each languageseparately. Erroneous responses (2.3%) andRTs that were outside the range of two standarddeviations from the participant and item mean(0.8%) were omitted from the latency analysis.The results are presented separately for eachlanguage in Tables 2 and 3.

An analysis of variance (ANOVA) was car-ried out for each language separately. The num-ber of neighbors in Dutch (Dutch N, large orsmall) and English (English N, large or small)were treated as within-participant factors, whileProficiency of the participant (high or low) andOrder of Presentation (Dutch followed by En-glish or English followed by Dutch) were be-tween-participant factors.

Reaction time data.For Dutch items, a sig-nificant effect of Dutch N was obtained[F1(1,38)5 38.64,p , .001;F2(1,76)5 7.71,p , .01]. Dutch words with many Dutch neigh-bors were responded to 59 ms slower thanwords with few Dutch neighbors (see Table 7for summary data from this experiment and the


following experiments). The effect of English Non Dutch items was also inhibitory (41 ms), butsignificant only in the participant analysis[F1(1,38)5 7.07,p , .05;F2(1,76)5 1.67,p 5.20]. The interaction between Dutch N and En-glish N was not significant [bothFs , 1]. Nomain effects were found for Proficiency[F1(1,38), 1; F2(1,76)5 19.69,p , .001] orOrder of Presentation [F1(1,38) , 1; F2(1,76)5 21.34,p , .001]. However, in combinationthese two factors interacted with English N[F1(1,38)5 4.67,p , .05;F2(1,76)5 5.06,p ,.05]. No other interactions were significant inboth participant and item analyses.

Further analysis of the interaction betweenProficiency, Order of Presentation, and EnglishN revealed that high proficiency participantsshowed a marginally significant interaction be-tween Order of Presentation and English N[F1(1,19)5 3.11,p 5 .09;F2(1,78)5 6.48,p ,.05], while low proficiency participants did not[F1(1,19)5 1.49,p 5 .24; F2(1,78), 1]. Thelatencies of high proficiency participants were

affected by English N only in the first block [66ms inhibition effect;F1(1,10)5 6.62,p , .05;F2(1,78)5 4.42,p , .05] and not in the secondblock [1 ms inhibition; bothFs , 1].

The latencies of English words showed asignificant inhibitory effect of Dutch N[F1(1,38)5 23.86,p , .001;F2(1,76)5 4.84,p , .05]. In contrast with the Dutch items,however, the effect of English N on Englishwords was facilitatory, but this effect was onlysignificant by participant [F1(1,38) 5 13.70,p , .01; F2(1,76)5 1.45,p 5 .23]. The inter-action between Dutch N and English N wasonly marginally significant [F1(1,38) 5 23.23,p , .001;F2(1,76)5 2.92,p 5 .09]. Similar tothe Dutch items, no significant effects werefound for Proficiency [F1(1,38), 1; F2(1,76)57.52, p , .01] or Order of Presentation[F1(1,38) 5 2.08, p 5 .16; F2(1,76) 5 88.86,p , .001]. Furthermore, no significant interac-tions were obtained.

Error data. An analysis of variance run onthe error data for each language separately

TABLE 2

Mean Identification Latencies of Experiment 1 (in Milliseconds) of Participants Presented with Dutch Words in theFirst Block or the Second Block (Standard Deviation and Percentage of Errors Are Presented in Parentheses.)

Low proficiency High proficiency

Dutch N English N Block 1 Block 2 Block 1 Block 2 Mean

Large large 1681 (310, 2.7) 1659 (182, 3.5) 1739 (166, 2.7) 1674 (219, 3.5) 1689Large small 1684 (329, 1.8) 1634 (170, 3.0) 1675 (229, 3.6) 1672 (180, 5.5) 1667Small large 1614 (275, 1.8) 1600 (179, 3.0) 1707 (163, 1.4) 1595 (186, 4.5) 1630Small small 1586 (258, 4.5) 1529 (135, 1.5) 1639 (217, 3.2) 1599 (217, 2.5) 1589

TABLE 3

Mean Identification Latencies of Experiment 1 (in Milliseconds) of Participants Presented with English Words in theFirst Block or the Second Block (Standard Deviation and Percentage of Errors Are Presented in Parentheses.)

Dutch N English N

Low proficiency High proficiency

Block 1 Block 2 Block 1 Block 2 Mean

Large large 1784 (185, 2.0) 1616 (309, 3.2) 1669 (152, 2.5) 1604 (174, 2.7) 1666Large small 1751 (212, 2.5) 1588 (285, 3.2) 1670 (158, 3.0) 1609 (183, 3.6) 1652Small large 1630 (150, 4.0) 1513 (259, 5.5) 1581 (183, 2.5) 1517 (206, 3.6) 1558Small small 1689 (170, 5.0) 1598 (314, 3.6) 1648 (200, 3.0) 1631 (202, 1.4) 1640


showed no significant main effects and onlyone significant interaction for Dutch itemsbetween Order of Presentation, Dutch N, andEnglish N [F1(1,38)5 6.59,p , .05;F2(1,76)5 5.92, p , .05].

Discussion

The most important result of Experiment 1 isthat target word recognition was significantlyinfluenced by the number of orthographicneighbors in the non-target language. For En-glish (L2), a significant inhibitory effect ofDutch N was found in both the participant anditem analysis, while for Dutch (L1), the inhibi-tory effect of English N was significant only inthe participant analysis. These findings repre-sent a further demonstration of effects of ortho-graphic neighborhood across languages in bilin-gual participants, adding to the work ofBeauvillain (1992), Grainger and Dijkstra(1992), and Bijeljac-Babic, Biardeau, andGrainger (1997). As previously argued byGrainger and Dijkstra (1992) and Grainger(1993), such results cannot be accommodatedby a serial search model of bilingual lexicalaccess in which language context determinesthe order in which L1 and L2 representationsare examined. Only a model that allows simul-taneous activation of word representations inboth languages (i.e., a non-selective accessmodel) can account for such results. Further-more, the inhibitory nature of these between-language neighborhood effects suggests thatword units in both languages are interconnectedvia inhibitory links (see also Bijeljac-Babic, Bi-ardeau, & Grainger, 1997). Thus, these resultssupport the hypothesis of non-selective accessto an integrated lexicon.

The within-language neighborhood effectsshowed a more complex pattern of neighbor-hood effects. Dutch words showed clear inhib-itory effects of increasing number of Dutchneighbors (59 ms). English words showed a 34ms facilitation effect when the number of En-glish neighbors increased, but this effect wasonly significant in the participant analysis. Be-cause perceptual identification paradigms typi-cally show inhibitory effects of neighborhooddensity (Carreiras et al., 1997; Grainger & Ja-

cobs, 1996; Ziegler et al., in press; Snodgrass &Mintzer, 1993), it is the facilitatory pattern ob-tained with English target words that requires anexplanation here. We will return to this point inthe General Discussion.

Next, we must consider the manipulation ofsecond-language proficiency in the experiment.The two groups of participants were assumed todiffer with respect to their second-language pro-ficiency. However, the results suggest that theproficiency manipulation was not very effec-tive, since Proficiency neither yielded a signif-icant main effect nor interacted with the neigh-borhood density factors. However, Proficiencydid interact with the combination of presenta-tion order of the Dutch items and number ofEnglish neighbors. Latencies of high profi-ciency participants to Dutch target wordsshowed more interference from English neigh-bors when the Dutch block was presented be-fore the English block than vice versa. Thisinteresting result suggests that our high profi-ciency participants in some way controlled theeffects of non-target language neighbors de-pending on their expectations. Their Englishlexicon may have been more active when theyexpected the English language to be relevant (atthe beginning of the experiment), and less sowhen it was considered less relevant (in theDutch block after the English block had oc-curred). Note that the participants were awarethat their knowledge of English was potentiallyrelevant because they were explicitly asked toperform in a bilingual experiment. This sugges-tion is in agreement with the language mode (orrelative language activation) hypothesis pro-posed by Grosjean (1997).

We tried to obtain extra support for this viewby carrying out an additional ANOVA over theDutch response times of the high proficiencyparticipants in which each Dutch block, preced-ing the English block or following it, was di-vided into two parts. Apart from Condition andOrder of Presentation (the Dutch block fol-lowed by the English block or vice versa), thisanalysis included as an extra within-participantfactor Part of Block (first or second part). Dueto the post hoc nature of this analysis, the itemsin each condition were not counterbalanced,


since items were randomized for each partici-pant. The analysis showed a significant effect ofOrder of Presentation [F2(1,156)5 13.54,p ,.001] and a significant interaction between En-glish N and Order of Presentation [F2(1,156)56.63,p , .05]. The interaction reflects a gradualchange over the experiment in the response timedifference between large and small English Nconditions from inhibition towards facilitation(see Dijkstra & Van Heuven, 1998). The inter-action between English N, Order of Presenta-tion, and Part of Block was not significant[F2(1,156) , 1]. To conclude, the hypothesisthat high proficiency participants are able toexert a certain degree of control over the rela-tive activity of their lexica clearly deserves fur-ther investigation (also see Dijkstra, Van Jaars-veld & Ten Brinke, 1998).

Experiment 2 provided a further examinationof within- and between-language neighborhoodeffects in the progressive demasking task, thistime using a mixed language presentation. Inthis situation, we expected an even strongereffect from non-target neighbors on target iden-tification because both languages must be keptactive during the experiment. In addition, weattempted to replicate the facilitatory effect ofnumber of English neighbors on English wordidentification.

EXPERIMENT 2: PROGRESSIVEDEMASKING (MIXED PRESENTATION)

Method

Participants. Forty Dutch students withDutch as their native language and with normalor corrected-to-normal vision participated inthis experiment. Participants were paid for theirparticipation.

Stimuli and design.The same materials andtask were used as in Experiment 1. The onlydifference was that Experiment 2 consisted ofonly one block of items in which both Englishand Dutch words were presented in a randomorder. The language of instruction and feedbackin the experiment was counterbalanced. Onegroup of participants received instruction andfeedback in Dutch, while a second group re-ceived them in English.

Procedure.The participants started with apractice session of 25 trials. Next, all 160 testitems were presented in a random order in oneblock, starting with five dummy items. The pro-cedure, presentation, and recording were thesame as in Experiment 1. The experiment tookabout 20 minutes.

Results

For items of both languages, erroneous re-sponses (3.0%) and RTs which were outside therange of two standard deviations from the par-ticipant and item mean (1.4%) were excludedfrom the latency analyses. The mean RTs anderror rates for items of each language are pre-sented in Table 4. Separate ANOVAs were car-ried out on the latency data of each language inwhich the number of neighbors in Dutch (DutchN) or English (English N) were treated as with-in-participant factors.

Reaction time data.For Dutch words, theanalysis showed, as in Experiment 1, a signifi-cant inhibitory effect of Dutch N [F1(1,39) 549.51,p , .001;F2(1,76)5 9.23,p , .01]. RTsto Dutch words with a larger number of Dutchneighbors were 70 ms slower than to Dutchwords with only a few Dutch neighbors. Theeffect of English N on Dutch items was nowsignificant in both participant and item analyses[F1(1,39)5 16.52,p , .001;F2(1,76)5 4.24,p , .05], indicating that Dutch words withmany English neighbors were responded tomore slowly than words with few Englishneighbors. No interaction was observed be-tween Dutch N and English N [bothFs , 1].

The analysis of the RTs to English wordsrevealed a pattern similar to Experiment 1.

TABLE 4

Mean Identification Latencies of Experiment 2 (in Milli-seconds) (Standard Deviation and Percentage of Errors ArePresented in Parentheses.)

Dutch N English N English Dutch

Large large 1777 (247, 4.8) 1756 (249, 7.0)Large small 1761 (245, 5.5) 1710 (205, 5.1)Small large 1660 (233, 3.1) 1689 (229, 4.4)Small small 1739 (240, 3.0) 1637 (189, 4.0)


There was a significant inhibitory effect ofDutch N [F1(1,39)5 49.90,p , .001;F2(1,76)5 6.58, p , .05]. The facilitatory effect ofEnglish N again was only significant in theparticipant analysis [F1(1,39)5 7.13,p , .01;F2(1,76) 5 1.01,p 5 .32]. The interaction be-tween Dutch N and English N was not signifi-cant [F1(1,39) 5 21.15,p , .001; F2(1,76) 52.48,p 5 .12].

Error data.An ANOVA on the error data forDutch items showed a marginally significantinhibitory effect of Dutch N [F1(1,39) 5 7.11,p , .01;F2(1,76)5 3.14,p 5 .08] but no effectof English N [F1(1,39) 5 3.36, p 5 .07;F2(1,76) 5 1.13, p 5 .29], and no interactionbetween Dutch N and English N [F1(1,39) 51.49,p 5 .23;F2(1,76), 1]. For English items,there was a significant inhibitory effect of DutchN [F1(1,39)5 6.33,p , .05; F2(1,76)5 4.46,p , .05], but no effect of English N [bothFs ,1], and no interaction between the two factors[both Fs , 1].

Discussion

The results of Experiment 2 show a signifi-cant inhibitory effect of increasing the numberof non-target language neighbors for both En-glish and Dutch target items. This effect con-firms that non-target language neighbors influ-ence the identification of targets in theprogressive demasking task. Again, the hypoth-esis of non-selective access to an integratedlexicon is supported.

Comparing the results of Experiments 1 and2, there appears to be an increase in the effect ofnon-target language neighbors when going froma blocked (Experiment 1) to a mixed (Experi-ment 2) experimental design. This can be illus-trated by the inhibition effect of large English Non the latencies to Dutch words. This effect isnonsignificant (by items) in Experiment 1, butbecomes significant in Experiment 2. Thoughthe difference in size of effect across the exper-iments is admittedly only 8 ms, it might be thatthe effect of the non-target language on targetidentification is strengthened when the non-tar-get language also becomes relevant for taskperformance and/or target items from another

language occur in the same list as the targetitem.

The pattern of within-language neighborhoodeffects was the same as in Experiment 1. In bothexperiments, a facilitatory effect was obtainedto the English target words (only significant inthe participant analysis). Dutch words contin-ued to show an significant inhibition effect.

EXPERIMENT 3: GENERALIZEDLEXICAL DECISION

The main aim of Experiment 3 was to testwhether the results obtained with the progres-sive demasking technique generalize to anotherexperimental technique, widely used in studiesof visual word recognition, the lexical decisiontask. Exactly the same word conditions as inExperiment 2 (i.e., mixed language presenta-tion) were tested in the lexical decision task.

In addition, four nonword conditions, alsovarying in terms of number of orthographicneighbors in Dutch and English, were added tothe experiment. In lexical decision experiments,not only the positive reaction to words has beenshown to be sensitive to neighborhood charac-teristics but the negative reaction towards non-words has as well (Andrews, 1989, 1992; Car-reiras et al., 1997; Coltheart et al., 1977; Sears,Hino, & Lupker, 1995). We therefore investi-gated the effect of both English and Dutchneighborhood density on responses to word andnonword stimuli in the same experiment.

Method

Participants. Forty-eight Dutch studentsfrom the same population as Experiment 2 par-ticipated in this experiment. None of them hadparticipated in the previous experiments. Allstudents were native speakers of Dutch.

Stimuli and design.All 160 words (80 Dutchand 80 English words) of the previous ex-periments were included in this experimentalong with 160 nonwords. Each nonword wasconstructed by changing one letter in a Dutchor English word (not a target word in the ex-periments) in such a way that it formedan orthographically and phonologically legal


letter string in both languages (bigram analysisensured that only legal combinations wereincluded).

Analogous to the word conditions, four ex-perimental nonword conditions were definedwith respect to the number of neighbors inDutch and in English, each consisting of 40nonwords: (1) nonwords with many neighborsin both English and Dutch (large English N,mean 3.5, and large Dutch N, mean 3.5), (2)nonwords with many neighbors in English(large English N, mean 3.5) and few in Dutch(small Dutch N, mean 1.0), (3) nonwords withfew neighbors in English (small English N,mean 1.0) and many in Dutch (large Dutch N,mean 3.5), and (4) nonwords with few wordneighbors in either language (small English N,mean 1.0, and small Dutch N, mean 1.0). Thefour nonword conditions were matched asclosely as possible with respect to the number ofhigh frequency (larger than 50 o.p.m.) neigh-bors and positional bigram frequency. The latterwas computed as the summed frequency ofwords containing the bigram of the nonword atthe specified position, averaged over the num-ber of bigrams in the nonword.

Procedure.Presentation of the visual stimuliwas controlled by an Apple Macintosh Quadraconnected to a 67-Hz Trinitron monitor and tothe same type of button box as used in earlierexperiments. Before the experiment, partici-pants were informed about their task by meansof written instructions. They were told that aseries of letter strings would appear on thescreen, one after the other, and that they had todecide as quickly and as accurately as possiblewhether each of the presented items was a word(Dutch or English) or not. Responses weremade by pressing one of two buttons on theresponse box. Half of the participants wereasked to respond with their left finger when thestimulus was a Dutch or English word and withtheir right finger when it was a nonword. Theother half of participants received the reversedinstruction. Language of instruction and feed-back (Dutch or English) was counterbalancedby a similar, further subdivision of participants.

When the participants had read the instruc-tions, the experiment began with the 24 practice

trials. Next, the 320 experimental stimuli werepresented to each participant in a pseudo-ran-dom order. Participants received no more thanfour words or four nonwords in sequence. Theparticipant took a short break after 164 trials.Four dummy trials were inserted at the begin-ning of the experiment and 2 dummy items wereinserted after the break. Prior to the presentationof each stimulus word, two horizontal fixationlines were presented at the center of the screenwhere the stimulus word was to appear. Thesefixation lines disappeared after 800 ms and wereimmediately followed by the stimulus. Thestimulus remained on the screen until a deadlineof 1500 ms was reached, or until the participanthad pressed one of the response buttons. Theinterval between two successive trials was 500ms. The experiment lasted about thirty minutes.

Results

Erroneous responses (6.6%) and RTs whichwere outside the range of two standard devia-tions from the participant and item mean (1.7%)were excluded from the latency analyses (total8.5%). The mean RTs and error rates are pre-sented in Table 5. Separate ANOVAs were car-ried out on the latency and error data of wordsfrom each language and for the nonwords, inwhich the number of neighbors in Dutch (DutchN) or English (English N) were treated as with-in-participant factors.

Reaction time data.The analysis of Dutchword latencies revealed that responses wereagain slower for words with many Dutchneighbors than for words with few Dutchneighbors, but this effect was now only signif-icant by participants [F1(1,47) 5 20.81, p ,.001; F2(1,76) 5 1.87, p 5 .18]. Also, theinhibitory effect of English N was again onlysignificant in the participant analysis [F1(1,47)5 4.85,p , .05; F2(1,76), 1]. As in Experi-ments 1 and 2, no interaction was found be-tween these factors [F1(1,47)5 3.58,p 5 .07;F2(1,76), 1].

The analysis for English words revealed asignificant inhibition effect (27 ms, see Table 7)of Dutch N [F1(1,47) 5 32.13, p , .001;F2(1,76)5 7.07,p , .05]. Again, we obtained


a facilitation effect of English N, which wassignificant in both item and participant analyses[F1(1,47) 5 29.49, p , .001; F2(1,76) 55.46,p , .05]. The interaction between DutchN and English N was not significant [bothFs , 1].

The analysis of the nonword latency datarevealed a significant inhibition effect of DutchN [F1(1,47) 5 78.40,p , .001; F2(1,156) 511.55,p , .01]. There was no effect of EnglishN [both Fs , 1], and no interaction betweenDutch N and English N [F1(1,47)5 1.01,p 5.32; F2(1,156), 1].

Error data.An ANOVA on the error data forDutch items showed no effect of Dutch N[F1(1,47) 5 1.61, p 5 .21; F2(1,76) , 1],English N [both Fs , 1], or Dutch N andEnglish N together [bothFs , 1]. Similar toExperiment 2, errors on English items showed asignificant effect of Dutch N [F1(1,47)5 24.42,p , .001;F2(1,76)5 5.70,p , .05]. Again, noeffect was found of English N [bothFs , 1],and the errors showed no interaction of Dutch Nand English N [F1(1,47) 5 2.69, p 5 .11;F2(1,76), 1].

For nonword items, a significant inhibitoryeffect of Dutch N was found [F1(1,47)5 20.66,p , .001; F2(1,156)5 6.59,p , .001]. Therewas no effect of English N [bothFs, 1] and nointeraction of Dutch N and English N [bothFs , 1].

Discussion

The results of Experiment 3 extend thoseobtained in the two previous experiments with anew group of bilingual participants and a dif-

ferent experimental task. Once again, signifi-cant inhibition from Dutch neighbors on En-glish target items was observed. And onceagain, the effect of within-language neighborswas facilitatory with English stimuli tested inthe same experiment. However, in contrast withthe previous experiments, the effects of Dutchand English neighbors on Dutch target itemsfailed to be significant in both participant anditem analyses. As seen in Table 7, the lexicaldecision task of Experiment 3 generally led tosmaller neighborhood density effects than theprogressive demasking task of Experiments 1and 2.

In this generalized lexical decision experi-ment, the latencies for the nonwords were sig-nificantly affected by the neighborhood size oftheir Dutch neighbors, and not by their Englishneighbors.2 It would appear that in this experi-mental situation, although both English andDutch were relevant target languages, the dom-inant language exerted the major influence onperformance to nonword stimuli. Being themother tongue (L1) of the bilingual participants,Dutch exerted the strongest effect on rejectiontimes to nonwords. This should not, however,be the case when only English (L2) words andnonwords are presented to Dutch/English bilin-guals in a ‘monolingual‘ lexical decision task.

2 Recently, we replicated the neighborhood density ef-fects for nonwords in a generalized lexical decision taskinvolving Dutch/English cross-language homographs (Ex-periment 3 in Dijkstra, Van Jaarsveld, & Ten Brinke, 1998).Nonwords derived from Dutch or English words revealed(inhibitory) effects of about 11 ms for English N and 37 msfor Dutch N.

TABLE 5

Mean Identification Latencies of Experiment 3 (in Milliseconds) (Standard Deviationand Percentage of Errors Are Presented in Parentheses.)

Dutch N English N English Dutch Nonwords

Large large 621 (69, 13.3) 581 (59, 7.6) 675 (75, 9.7)Large small 642 (85, 12.6) 566 (60, 7.5) 672 (78, 9.7)Small large 590 (62, 6.6) 561 (58, 6.6) 647 (79, 6.7)Small small 619 (73, 8.7) 560 (56, 6.4) 649 (73, 6.4)


In this situation, responses to nonword stimulishould be sensitive to their neighborhood size inEnglish. Testing this hypothesis was one of theaims of Experiment 4.

Experiment 4 also tested the English wordsused in the previous experiments with a newgroup of Dutch/English bilingual participantsand an English monolingual control group. Weexpected effects of Dutch neighbors to appearonly in the bilingual participants’ performance,while effects of English neighbors should bepresent in both groups of participants.

EXPERIMENT 4: ENGLISHLEXICAL DECISION

Method

Participants.Twenty English and twenty-oneDutch right-handed participants with normal orcorrected-to-normal vision participated in thisexperiment. The English participants were allmonolingual native speakers of English, whowere tested in Cambridge (UK). The Dutchparticipants were all native speakers of Dutch.They were tested in Nijmegen (Netherlands).None of the participants participated in earlierexperiments.

Stimuli and design.The English set of 80stimulus words described above was used in theexperiment, together with a matched selectionof 80 of the 160 nonwords used in Experiment3. Both words and nonwords consisted of 20items for each of the four conditions describedearlier (small or large neighborhoods in Englishor Dutch).

Procedure.Presentation of the stimuli wascontrolled by an Apple Power Macintosh com-puter (Cambridge, PowerMac 8500; Nijmegen,PowerMac 7200). Latencies were recorded bymeans of a button box (the same as in previousexperiments) connected to the computer. Wordsand nonwords consisted of black Courier capitalletters (28 points) presented at the center of aApple monitor (Cambridge, 200 Trinitron; Nij-megen, 150 Multiscan) on a white background.The monitor was placed at a distance of approx-imately 60 cm from the participant. Each trialstarted with a fixation marker,z , which waspresented for 500 ms before the screen was

cleared. After 500 ms, the target item was pre-sented until the participant responded. The nexttrial started 800 ms after the participant re-sponded or after 2000 ms if the participant didnot respond.

Before the experimental block began, the par-ticipant started with a practice session of 24items. At the beginning of the experimentalblock and after the break, four dummy itemswere inserted. Each participant saw a differentrandomized order of the test items.

Participants read written instructions in En-glish explaining that strings of four letterswould appear at the middle of the computerscreen. They were instructed to give a speededresponse by pressing the right response buttonof the two-button response box with their rightindex finger if the string was a English word,and by pressing the left button with their leftindex finger if the string was not a Englishword.

After the experiment, each participant filled outa questionnaire concerning his or her knowledgeabout and experience with other languages, inorder to check to which extent the participantcould be considered as a pure English monolin-gual or Dutch/English bilingual participant. Thesession lasted about 20 minutes.

Results

First, the data were removed for five words(BIAS, BUTT, ATOM, MYTH, JERK) that hada mean RT of more than two standard devia-tions above the mean of their test condition inthe English monolingual group. Also, the datafor four nonwords (GONK, KNAT, FRIG,BOUL) were removed since some of the En-glish participants considered them to be low-frequency English words (e.g., a GONK wassaid to be an English toy).3 Overall, monolin-gual and bilingual participants made 4.7% erro-neous responses. The errors and the RTs whichwere outside the range of two standard devia-tions from the participant and item mean (1.8%)

3 We reanalyzed the data of our bilingual Experiments1–3 with these five word and four nonword items removed.This procedure did not change the pattern of results in anysubstantial way, nor did it affect the significance of effectsto any relevant extent.


were excluded from the latency analyses. Theresulting mean RTs and error rates are presentedin Table 6.

Separate ANOVAs were carried out on thelatency and error data of the English words andnonwords in which the number of neighbors inDutch (Dutch N, large vs small) and English(English N, large vs small) were treated aswithin-participant factors, and the LanguageStatus of the participants (monolingual or bilin-gual) as a between-participant factor.

Reaction time data.The analysis of the wordlatencies showed again a facilitatory effect ofEnglish N which was significant in the par-ticipant analysis but not in the item analysis[F1(1,39)5 10.13,p , .01; F2(1,71)5 2.09,p 5 .15]. The interaction between Dutch Nand English N was not significant [F1(1,39)51.07, p 5 .31; F2(1,71) , 1]. Not surpris-ingly, monolingual English participants weresignificantly faster than Dutch/English bilin-gual participants [F1(1,39)5 14.15,p , .001;F2(1,71)5 213.61,p , .001]. There was alsoa marginally significant inhibition effect ofDutch N [F1(1,39) 5 9.41,p , .01; F2(1,71)5 3.00, p 5 .09].

More importantly, the monolingual partici-pants showed no effect of Dutch N while thebilinguals did. This interaction between Lan-guage Status of the participant (monolingual,bilingual) and Dutch N was significant [F1(1,39)5 4.70, p , .05; F2(1,71) 5 4.10, p , .05].Language Status did not interact with English N[F1(1,39)5 5.02,p , .05;F2(1,71)5 2.06,p 5.16], and also showed no interaction with both

Dutch N and English N [F1(1,39)5 1.31,p 5.26; F2(1,71), 1].

The analysis of the nonword latenciesshowed a significant inhibitory effect of EnglishN [F1(1,39) 5 26.49, p , .001; F2(1,72) 55.73,p , .05]. The effect of Dutch N was notsignificant [bothFs , 1]. There was no inter-action between Dutch N and English N [bothFs , 1]. Again, monolingual participants weresignificantly faster than bilingual participants[F1(1,39)5 8.03,p , .01; F2(1,72)5 170.33,p , .001]. The Language Status of the partici-pant did not interact with Dutch N [F1(1,39)52.06, p 5 .16; F2(1,72) 5 1.23, p 5 .27],English N [both Fs , 1], or Dutch N andEnglish N [bothFs , 1].

Error data. For English items, error scoresshowed no effect of Dutch N [F1(1,39)5 13.45,p , .01; F2(1,71)5 2.60,p 5 .11], English N[F1(1,39)5 4.44,p , .05;F2(1,71)5 1.44,p 5.24], and no interaction between Dutch N andEnglish N [F1(1,39)5 2.89,p 5 .10; F2(1,71), 1]. Bilinguals produced more errors thanmonolinguals [F1(1,39) 5 28.48, p , .001;F2(1,76) 5 14.84,p , .001]. Language Statusshowed a marginally significant interaction withDutch N [F1(1,39)5 9.93,p , .01;F2(1,76)52.77,p 5 .10], but not with English N [F1(1,39), 1; F2(1,76), 1], and not with both EnglishN and Dutch N [F1(1,39) 5 1.00, p 5 .32;F2(1,76), .1].

The error scores for nonwords showed nosignificant main effect of Dutch N [F1(1,39)5 1.46, p 5 .24; F2(1,76) , 1]. In contrastwith the error scores of the English words,

TABLE 6

Mean Identification Latencies of Experiment 4 (in Milliseconds) of Monolingual and Bilingual Participants(Standard Deviation and Percentage of Errors Are Presented in Parentheses.)

Dutch N English N

Monolinguals Bilinguals

Words Nonwords Words Nonwords

Large large 488 (58, 4.0) 575 (86, 8.8) 585 (69, 12.1) 651 (94, 9.5)Large small 507 (71, 5.0) 558 (74, 4.5) 583 (74, 12.1) 635 (94, 4.0)Small large 484 (57, 3.0) 581 (87, 5.8) 561 (70, 4.8) 642 (99, 8.1)Small small 503 (53, 5.3) 560 (65, 6.1) 564 (73, 9.5) 626 (93, 3.5)


those for nonwords were affected by EnglishN [F1(1,39) 5 15.48,p , .001; F2(1,76) 57.46, p , .01]. No interaction was foundbetween Dutch N and English N [F1(1,39) 53.43, p 5 .07; F2(1,76) 5 1.14, p 5 .29].Furthermore, monolinguals and bilinguals didnot differ in their error scores [bothFs , 1],and this factor did not interact with Dutch N[both Fs , 1], or English N [F1(1,39)5 2.85,p 5 .10; F2(1,76) 5 2.75,p 5 .10], nor withboth English N and Dutch N [F1(1,39) 51.57, p 5 .22; F2(1,76) 5 1.05, p 5 .31].

Discussion

The results of Experiment 4 show that re-sponses of English monolingual participants toEnglish target words were not influenced by thenumber of Dutch word neighbors of these tar-gets. However, this factor did significantly in-fluence the responses of the Dutch/English bi-linguals. This is particularly interesting becausein this experiment (in contrast to the earlierones) the bilinguals were not using their Dutchlanguage at any time during the experiment.The observed pattern of results strongly sug-gests that it is knowledge of the non-targetlanguage (in this case the mother tongue) thatproduces the between-language neighborhoodeffects in the bilingual participants, not someuncontrolled variable. As in the previous exper-iments, increasing the number of English neigh-bors resulted in facilitation.

With respect to the nonwords, the monolin-gual and bilingual participant groups showedthe same type of effect, slower RTs when thenumber of English neighbors of the nonwordtargets increased. It is interesting to note thatin the generalized lexical decision task ofExperiment 3, the RT pattern for the non-words was especially affected by the numberof Dutch neighbors. Apparently, the presencein that experiment of target items from themother tongue of the bilingual participantshad a pervasive effect on nonword responses.This pattern of neighborhood effects on re-sponses to nonword stimuli will be discussedmore fully below.

GENERAL DISCUSSION

In the experiments presented in this paper wetested whether bilingual word recognition in-volves language selective or non-selective lex-ical access and whether this access is to lan-guage-dependent (separate) or integrated lexicalsystems. It was argued that target word identi-fication in bilingual participants will be influ-enced by orthographic neighbors from thenon-target language only if bilingual word rec-ognition involves non-selective access to an in-tegrated lexicon. In accordance with this hy-pothesis, both within- and between-languagemanipulations of orthographic neighborhooddensity influenced performance in bilingual par-ticipants on words and nonwords presented intwo experimental paradigms, progressive de-masking and lexical decision. The between-lan-guage effect (from Dutch neighbors on Englishtarget words) disappeared in an English mono-lingual control group tested in Experiment 4(see Table 7). These data provide clear evidencethat stimulus items automatically activate ortho-graphically similar words in both the target lan-guage and the other language of the bilingualparticipant. The effects of between-languageneighborhood were further obtained in condi-tions where stimuli were blocked by languageor presented in mixed language lists. Numeri-cally larger effects appeared in the latter condi-tion.

The observed effects of number of neighborsfrom the non-target language stand in clear con-tradiction with the predictions of models assum-ing selective access and/or independent lexica.Only models assuming non-selective access andintegrated lexica may provide a viable accountfor all results reported here. However, the effectof non-target language neighbors depended tosome extent on the experimental paradigm andstimulus list context (cf. Table 7). For instance,effects were generally smaller in lexical deci-sion than in progressive demasking. Further-more, in the blocked progressive demaskingtask of Experiment 1, response times of highproficiency participants to Dutch items weresignificantly affected by English neighborhooddensity when the Dutch words were presented


before the English words. In order to accountfor the pattern of data across experiments, alanguage non-selective integrated-lexica modelis required which allows some control over therelative activation of the two languages. Belowwe will examine an implemented model of thiskind, the BIA model. However, we will firstdiscuss how the task-dependence of the ob-served cross-language interference effects mayhave come about.

Task-Dependence of Cross-LanguageInterference Effects

Examination of Table 7 suggests that theremay be differences in cross-language neighbor-hood density effects depending on task require-ments. For instance, it may be observed thatdensity effects are generally smaller in the lex-ical decision task than in the progressive de-masking task (especially for Dutch stimuli).Such task dependence has been observed in themonolingual domain by Carreiras, Perea, andGrainger (1997), who found that the effect ofneighborhood density changed from an inhibi-tory trend in progressive demasking to non-significant facilitation in lexical decision andsignificant facilitation in word naming. If thesize and perhaps the direction of the bilingualneighborhood density effects also depend ontask demands, any bilingual processing model(such as the Bilingual Interactive Activation

[BIA] model) must be extended with a ‘‘meta-account’’ that specifies how a particular task isperformed (Dijkstra et al., 1998).

For instance, it could be assumed, followingGrainger and Jacobs (1996), that identificationlatencies in progressive demasking depend on acriterion set on the activation levels of individ-ual word representations, while participants in alexical decision task may apply several criteriato generate their response. First, as in the pro-gressive demasking task, they may base theirpositive responses on a criterion set on individ-ual word activation. This type of criterion couldbe extended to the bilingual domain by applyingthe criterion to all words in an integrated lexi-con. A second criterion proposed by Graingerand Jacobs is one set on summed lexical activ-ity. Words with more orthographic neighborsgenerate higher levels of summed lexical activ-ity, thus leading to faster positive lexical deci-sion responses. This type of criterion would notapply to progressive demasking, in which par-ticipants need to uniquely identify one lexicalcandidate. For the bilingual variants of lexicaldecision, it is highly relevant whether summedlexical activity is calculated across both lexicaor across each lexicon separately, because theresults would be very different for the two op-tions.

The third criterion proposed by Grainger andJacobs (1996) applies specifically to nonword

TABLE 7

Neighborhood Density Effects (in Milliseconds) of Number of English (English N) and Dutch Neighbors (English N) forExperiments 1–4 (Effect Size is Calculated by Subtracting RTs in Conditions with a Large Number of Neighbors fromThose in Conditions with a Small Number of Neighbors. A Minus Sign Indicates a Facilitatory, and a Plus Sign an InhibitoryEffect.)

Exp 1 Exp 2 Exp 3 Exp 4

Progressivedemasking

Progressivedemasking

Generalizedlexical decision

Englishlexical decision

Blocked Mixed Monolinguals Bilinguals

English English N 234 232 225 219 23Dutch N 57 70 27 2 22

Dutch English N 41 49 8Dutch N 59 70 13

Nonwords English N 1 19 19Dutch N 26 25 6


performance. Following Coltheart et al. (1977),they assumed that correct responses to nonwordstimuli are governed by a criterion set on timefrom stimulus onset. The value of this criterionalso varies as a function of summed lexicalactivity. The greater the summed activity, thehigher the criterion. Nonwords with many wordneighbors generate higher levels of summedlexical activity and therefore a higher negativeresponse criterion and longer RTs (this effect ofneighborhood density on correct reject RTs tononword stimuli has been reported many timesin the literature, e.g., Coltheart et al., 1977). Theresults of Experiments 3 and 4 suggest thatsummed lexical activity in the most activatedlexicon, rather than the sum of activity acrosslanguages, is the critical factor determining RTsto nonword stimuli. In generalized lexical deci-sion (Experiment 3) only L1 (Dutch) neighborsaffected RTs to nonwords. In the English lexicaldecision task of Experiment 4, on the otherhand, only L2 (English) word neighbors af-fected RTs to nonword stimuli. Such an accountof the nonword data is not incompatible with theintegrated lexicon hypothesis, since we assumethat summed lexical activity can be calculatedlanguage-specifically. This language-specificsummation of lexical activity is carried out bythe ‘‘language nodes’’ of the BIA model to bedescribed below.

The Bilingual Interactive Activation(BIA) Model

The BIA model (Dijkstra & Van Heuven,1998; Grainger & Dijkstra, 1992) is an algorith-mic model of bilingual word recognition thatimplements non-selective bottom-up processing(letters activate words from both languages inan integrated lexicon) and language-specifictop-down processing (language nodes selec-tively inhibit activity in words of the other lan-guage). By using this language-specific top-down control mechanism, the BIA model canhandle both selective and non-selective results(similar mechanisms have been proposed forbilingual language production by Green, 1986,and Monsell, Matthews, & Miller, 1992). Toexamine to what extent the current pattern of

results can be simulated by the model, we firstneed to discuss the model in more detail.

The BIA model, illustrated in Figure 2, sharesthe basic architecture and parameter settings ofthe monolingual Interactive Activation model(McClelland & Rumelhart, 1981, 1988). It hastwo major extensions: (1) an integrated Dutchand English lexicon, and (2) an extra represen-tational layer containing two ‘language nodes‘that are connected to all the word nodes in bothlexica. A verbal (and slightly different) versionof this model has been described elsewhere(Grainger & Dijkstra, 1992; Grainger, 1993).

When a string of letters is presented to theBIA model, this visual input affects particularfeatures at each letter position, which subse-

FIG. 2. The Bilingual Interactive Activation (BIA)model. Arrowheads indicate excitatory connections; blackfilled circles indicate inhibitory connections. Though onlyfour positions for features and letters are represented, themodel is not limited to the recognition of four-letter words.


quently excite letters that contain these featuresand at the same time inhibit letters for which thefeatures are absent. The activated letters nextexcite words in both languages for which theactivated letter occurs at the position in ques-tion, while all other words are inhibited. At theword level, all words inhibit each other, irre-spective of the language to which they belong.Activated word nodes send excitatory feedbackto their constituent letters. Activated wordnodes from the same language also send activa-tion on to the corresponding language node,while activated language nodes send inhibitoryfeedback to all word nodes in the other lan-guage. The main function of the language nodesis to collect activation from words in the lan-guage they represent and inhibit active words ofthe other language. The activation of the lan-guage nodes reflects the amount of activity ineach lexicon. They also represent a simplemeans of storing knowledge about the languageto which a particular word form belongs(Grainger & Dijkstra, 1992). A more extensivedescription of the implemented model and ofsimulation work involving the language nodesis given in Dijkstra and Van Heuven (1998).

Figure 3 presents the simulation results of theBIA model with a reduced English frequencyrange and limited asymmetric top-down inhibi-tion (from the Dutch language node on all En-glish words only), along with the results ofExperiment 2. Since the patterns of results forExperiments 1 and 2 were similar, the modelcaptures the general pattern of results acrossboth progressive demasking experiments.4 Ascan be seen, the BIA model captures the within-

language neighborhood effects in both Dutch(inhibition) and English (facilitation). How isthe model able to simulate these effects thatdiffer in direction? The inhibitory effect of largeDutch N on Dutch target words can be ac-counted for by the presence of lateral inhibitionbetween lexical candidates. The facilitation ef-fect of large English N on English target wordsrequires a more complex explanation.

In the BIA model, an English target wordalso activates Dutch neighbors. These Dutchneighbors activate the Dutch language node,which in turn inhibits all English words. Also,due to the reduced English frequencies, Dutchwords will be activated earlier in the recognitionprocess than English words. Thus, top-downinhibition and reduced English frequencies leadto an extra inhibitory effect from the Dutch tothe English lexicon. This effect is stronger forEnglish words that have only a small English Nthan for those with a large English N. Thus, arelative facilitation effect arises of large versussmall English N conditions. Simulation studiesshow that the capability of the model to simu-late the facilitatory effect of English neighbor-hood density on English target words dependson the setting of the top-down inhibition param-eter. When this parameter is set to zero, thefacilitatory effect of English neighborhood den-sity disappears while the effect of Dutch neigh-borhood remains inhibitory (see Dijkstra, VanHeuven, & Grainger, in press). In sum, simula-tions suggest that, at least for the progressivedemasking experiments with bilinguals, the fa-cilitatory effect of English N can be explainedby the relative activation of the two languagesdependent on word frequency in combinationwith asymetric top-down inhibition.5

With respect to the monolingual participantsin Experiment 4, the facilitatory effect of largeEnglish N observed in lexical decision may also

4 The quantitative fit of the model to the empirical data canbe computed by means of a chi-square procedure (see Footnote2 of Baayen, Dijkstra, & Schreuder, 1997). Since in the mixedstimulus presentation in Experiment 2 the overall activationstate of English and Dutch should be more comparable than inthe blocked Experiment 1, the BIA model was fitted to the dataof Experiment 2. After linearly transforming the cycle timesproduced by the model into milliseconds, a chi-square test wasperformed on the eight pairs of empirical and predicted RTs ofExperiment 2. For six degrees of freedom (eight minus twofree parameters, EFR and ATD), the resulting chi-square of12.15 was not significant (p . .05), indicating that the modeldata do not differ substantially from the empirical data in thisexperiment.

5 The relevance of the top-down feedback parameter alsobecomes clear when it is used to simulate the transitioneffect of English neighbors on Dutch target words that wasobserved over the experimental parts for high-proficiencyparticipants in Experiment 1. Manipulation of this parame-ter alone results in a simulation pattern that correlates .99with the eight empirical data means represented in Figure6.3 of Dijkstra and Van Heuven (1998).


be simulated with the BIA model, but in adifferent way. The BIA model with only onelexicon (simulating monolingual word recogni-tion) is equivalent to the IA model. As de-scribed earlier, Grainger and Jacobs (1996)have extended the IA model with multiple read-

out criteria to account for monolingual lexicaldecision. They show that a model equipped withthese criteria can in principle simulate monolin-gual facilitation effects of large versus smallEnglish neighborhoods. The current BIA modelis not yet extended with similar read-out crite-

FIG. 3. Mean identification latencies of Experiment 2 (in milliseconds) and simulation results of the BIAmodel (in cycles) for Dutch N and English N effects on Dutch and English items. In the simulation a reducedupper limit was used for the resting-level activations of English items (20.30), and the standard IA-model upperlimit was used for resting-level activations of Dutch items (0.00). The top-down inhibition parameter from theDutch language node to all English words was set to 0.03.


ria, because it is not clear, for example, howsummed lexical activity should be calculated inthe bilingual case, across both lexica or for eachlexicon separately. As a consequence, the modelis not yet able to simulate the results of thegeneralized lexical decision task in Experiment3, or the English lexical decision results of theDutch/English bilinguals in Experiment 4.

In the next two sections we will discuss thetheoretical consequences of the between-lan-guage interference effects and possible interpre-tations of the within-language facilitation ef-fects obtained with English stimuli in relation tothe monolingual literature.

Cross-Language Interference Effectsin Bilingual Word Recognition

The present experiments have demonstratedthat non-target language neighbors influenceidentification of target words. The simulationresults from the BIA model presented in Figure3 show that the model does indeed capture theseinhibitory effects of between-language neigh-bors. The model also simulates our finding thatbetween-language neighborhood effects werealways stronger in L2 target words (English inthe present study) than L1 target words. In themodel this is a consequence of the lower aver-age resting level activations of L2 words com-pared to L1 words. The reduced resting level ofL2 word activation in the BIA model is de-signed to reflect the reduced exposure of bilin-gual readers to L2 words relative to L1 words.

Cross-language interference effects have pre-viously been demonstrated in many differentsituations, and it would appear that these earlierobservations also find a natural explanationwithin the BIA model. For example, the nega-tive effects of changing languages on wordrecognition performance in bilinguals (e.g.,Grainger & Beauvillain, 1987; Soares & Grosjean,1986) can be explained in terms of the relativeinhibitory state of the target language lexiconafter processing a non-target language word.The significant diminution of this languagechange effect observed by Grainger and Beau-villain when target words had very language-specific spellings (and therefore had no neigh-bors in the non-target language) can be

explained by the rapid recovery of the targetlanguage lexicon in such cases. According tothe BIA model, the relative number of ortho-graphic neighbors in the target and non-targetlanguage determines the speed with which ac-tivation in the target language node will domi-nate non-target language node activation.

Bijeljac-Babic et al. (1997) have recentlyprovided further evidence for between-languageorthographic inhibition in bilinguals. Using themasked prime paradigm with very short primeexposures, Bijeljac-Babic et al. compared per-formance to target words preceded by ortho-graphically similar or dissimilar prime wordsthat were or were not from the same language asthe target word. Longer RTs were observed inthe orthographically related prime condition,thus replicating the monolingual results ofSegui and Grainger (1990). More interesting isthat this orthographic inhibition effect was ob-served both within and between languages. Thispattern of results was correctly predicted by theBIA model. Presentation of the prime wordcauses the corresponding word unit to rise inactivation. When an orthographically relatedtarget word follows, the representation corre-sponding to the prime word continues to receivepartial bottom-up support from the stimulus in-put (via shared letters). Thus, the prime wordrepresentation remains strongly activated dur-ing the initial processing of the target word andgenerates strong inhibition on the target wordrepresentation. That such inhibitory ortho-graphic priming effects are observed across lan-guages in bilingual participants is further evi-dence in favor of the non-selective access,integrated lexicon approach of the BIA model.

Cross-Language Differences in OrthographicNeighborhood Effects

We must now consider an unexpected aspectof the present results, namely, the different pat-tern of within-language neighborhood effectsobserved for Dutch and English words. Estab-lishing the presence of cross-language neigh-borhood effects was most relevant to the bilin-gual goals of the current paper, but thedifferences found with respect to the directionof within-language neighborhood effects for


English and Dutch may have consequences formodels of monolingual word recognition.

In both the lexical decision and progressivedemasking experiments, we systematically ob-tained facilitatory effects of neighborhood den-sity to English words. This confirms the alreadywell-established picture for the lexical decisiontask (Andrews, 1989; 1992; Forster & Shen,1996; Sears et al., 1995) but runs counter toprior observations with the progressive demask-ing paradigm and other perceptual identificationtasks that require unique word identification(Carreiras et al., 1997; Grainger & Jacobs,1996; Snodgrass & Mintzer, 1993; Ziegler etal., in press). With Dutch (native language)stimuli, on the other hand, within-languageneighborhood effects were systematically inhib-itory in both the lexical decision and progres-sive demasking tasks.

Explanations for this differential effect ofEnglish and Dutch neighborhood density can besought in at least three possible directions: interms of differences between languages, partic-ipants, and stimulus material (cf. Grosjean,1997). As a first possibility, one might considerthat the opposing effects reflect differences inthe properties of languages. In her overview ofmonolingual neighborhood size and frequencystudies, Andrews (1997) points out that a greatmajority of the orthographic neighbors of En-glish 4-letter words share all but the first letterwith the target word. Since the final three lettersof these words often form what is referred to asthe word body (the orthographic equivalent ofthe phonological rhyme), this implies that suchunits may play a critical role in determining thesize and direction of neighborhood density ef-fects. Thus, according to Andrews (1997), onepossible source of cross-language differences ineffects of neighborhood density is the relativeimportance of word bodies as functional units inthe word recognition process. Variability acrosslanguages in the statistical regularities of spell-ing-to-sound mappings at different grain sizes ishypothesized to be a determining factor.

Compared to English, spelling-to-sound cor-respondences in Dutch are relatively regular(Van Heuven, 1980), and the role of the body inword naming seems to be less important in

Dutch than in English (Martensen, Maris, &Dijkstra, in preparation). An analysis of allDutch and English words from the CELEX da-tabase showed that on average English wordshave more rhyme neighbors (mean 12.3) thanDutch words (mean 8.5), while the mean num-ber of orthographic neighbors of a word is aboutthe same for the two languages (English, 5.2;Dutch, 5.5)6. Analysis of our stimulus materialshowed a similar pattern. Further research isneeded to find out how differences in orthogra-phy and phonology across languages affect or-thographic neighborhood density effects in vi-sual word recognition.

Differences in the properties of the Dutch andEnglish lexicons might also interact with char-acteristics of the bilingual participants and thestimulus material used in the experiments. Forinstance, since our bilinguals were unbalanced,the English and Dutch word stimuli could not bematched with respect to subjective frequency. Ifthe direction of neighborhood density effectsdepends on target word frequency, an explana-tion could be sought in this direction. However,there is no conclusive evidence in the monolin-guistic literature that this is the case (Andrews,1997).

Furthermore, within-language facilitation ef-fects for English may have a different origin inmonolinguals than in bilinguals. Grainger andJacobs (1996) propose that one of the criteriamonolinguals use to make a lexical decision is aresponse criterion set on summed lexical activ-ity. It is not clear how such summed lexicalactivity is calculated by bilinguals. The non-word results of Experiments 3 and 4 suggestthat the bilinguals used summed lexical activityin the most activated lexicon rather than the sumof activation across languages. Also, in Exper-iment 1, bilinguals with a high L2 proficiencyseemed to be able to affect the influence of theirsecond language during the experiment. Bothobservations can be taken as evidence that the

6 Our analysis showed that removing words from theEnglish lexicon with more than one pronunciation did notchange the average number of rhymes in English much(rhyme neighbors 12.3 vs 11.9; orthographic neighbors 5.4vs 5.2). The Dutch lexicon contained only words with onepronunciation.


bilingual status of a participant plays a role inthe different within-language neighborhooddensity effects for English and Dutch.

One further possible explanation for the fa-cilitation effect in English relates to stimuluscharacteristics. Our Dutch and English stimuliwere matched with respect to the number ofneighbors they had in English and Dutch (seeTable 1). However, it cannot be excluded thatthe actual number of English neighbors knownby our bilingual participants was lower than thatestimated on the basis of the CELEX databasecorpus. Not all test conditions would be affectedequally by overestimating the number of neigh-bors actually known. The effect would be stron-gest for English words having many within-language neighbors. However, the presence offewer English neighbors would (in the currentview) imply less inhibition relative to the twoother conditions, rather than more facilitation.

Further empirical research is obviously nec-essary to determine which (combination) ofthese sources (language, bilingual, stimulus ma-terial) affect(s) the direction of bilingual neigh-borhood density effects. As described above,the BIA model takes a stand here because itsuggests that, at least for our bilingual partici-pants, the observed within-language facilitatoryeffect of English N is due to the bilingual natureof our participants, implying reduced Englishfrequencies (relative to Dutch) and asymmetrictop-down inhibition.

Conclusions

Without doubt, there are alternative frame-works for developing a model of bilingualword recognition that is sensitive to factorsshown to be critical in the present experi-ments (cf. Grosjean, 1997). We invite re-searchers with different orientations to de-velop more complex variants of selective-access or independent-lexica models that doso. In the meantime, we have shown in thepresent paper how the effects of non-targetlanguage neighbors on target word recogni-tion in bilingual participants can be accom-modated by a mechanism of mutual inhibitionwithin an integrated orthographic lexicon.

Future work on cross-language neighbor-hood effects should provide further con-straints on our modeling efforts within thearea of bilingual word recognition. For in-stance, the possible influence of activatedphonological representations on the presentneighborhood density results must be exam-ined. Given that the Dutch spelling system ismore shallow than the English one, it cannotbe excluded that our Dutch and English re-sults have been differently affected by thepotentially relevant information sources oflexical and sublexical phonology. Further-more, simulation work on, for instance, bilin-gual homograph recognition will clearly beless than optimal until phonological codeshave been implemented in the BIA model.The activation of phonological representa-tions may be essential in allowing target lan-guage node activation to dominate over thenon-target language node in the case of cross-language heterophonic homographs (such asDutch/English GLAD, which is pronounceddifferently in Dutch and English). On thispoint, it will be critical to know whether onlytarget language phonology or both phonolog-ical codes are automatically generated on pre-sentation of such cross-language homographsand whether these codes activate languagenodes as well. Work is currently in progresson this particular question (Dijkstra, Grainger,& Van Heuven, submitted).

The evidence presented in this paper sug-gests that the right question to ask may noteven be whether we opt for a selective accessor a non-selective access model of bilingualword recognition, but which mechanisms inthe human processing system allow for justthe right amount of interference or facilitationunder particular experimental situations. Al-though it has been shown that the theoreti-cally motivated mechanisms currently incor-porated in the BIA model have sufficientflexibility to allow qualitatively correct sim-ulation of some of the result patterns in thisstudy, precise quantitative fits await furthermodel development.


APPENDIX

Dutch Words Used in Experiments 1–3

Large Dutch N, Large English N: bons,borg, bril, dolk, hiel, klam, knie, oord, plek,rund, sein, spar, takt, tolk, vork, wolk, worp,woud, wrak, zalfLarge Dutch N, Small En-glish N: berg, beul, bouw, deun, dief, eter,fuik, kelk, kies, knal, kous, rede, snik, teug,touw, twee, unie, vals, verf, viesSmall DutchN, Large English N: brug, bult, draf, drie,fris, galg, hemd, heup, lach, meid, melk,munt, nota, pret, prik, smid, stug, vete, welp,wilg Small Dutch N, Small English N: akte,ambt, blad, erwt, ezel, gesp, gids, gips, inkt,joch, muts, ober, pech, pion, rots, snor, stro,toga, trui, veld

English Words Used in Experiments 1–4

Large Dutch N, Large English N: aunt,blue, farm, hawk, knit, left, loan, loud, maid,monk, moon, path, quit, shoe, suit, tool, verb,weak, wrap, zeroLarge Dutch N, Small En-glish N: army, atom, bias, bird, diet, edge,germ, huge, butt, jerk, keen, knee, liar, lion,myth, noon, nude, obey, poem, poorSmallDutch N, Large English N: bath, bomb, busy,clue, coin, desk, dial, dirt, dish, firm, grey, hurt,iron, joke, lamb, limb, loss, milk, prey, rudeSmall Dutch N, Small English N: deny, duty,earl, envy, evil, folk, frog, guts, idol, kiss, okay,oral, oval, soup, true, twin, ugly, used, vein,view

Nonwords Used in Experiment 3

Large Dutch N, Large English N: aril,aunk, blag, boul, boup, braf, bret, dris, duef,elap, fram, frip, furk, gonk, heud, jeef, knat,knub, koup, loem, meem, merd, mots, oram,peit, pern, piot, pral, pred, rama, sluf, sluk, snus,sols, stui, tess, trum, tult, vene, zorkLargeDutch N, Small English N: alof, besp, bito,bouf, daus, drot, epoe, etel, feik, goep, grul,heut, irok, jees, jeul, jund, jurf, kalp, kelf, kerd,keun, loga, morp, muig, mups, nazz, noge, nont,noto, obel, oune, pris, puif, reug, reun, slen,smir, viem, woup, zulsSmall Dutch N, LargeEnglish N: aute, bele, bulf, ceot, chah, cham,clet, dolo, drid, dulp, feul, foug, fran, genk, girs,

jant, jero, jert, liry, lurd, lurp, lusp, naul, nirk,nudo, orim, pani, prad, prog, puet, raut, reud,rion, ruze, seto, snam, tirk, tran, vich, vornSmall Dutch N, Small English N: aler, anas,arns, aurd, baun, cafa, chof, deim, dilm, drio,durs, enip, fenk, feup, frig, frus, giep, heif, hilp,jalp, jofe, kach, kiot, knaf, luet, maup, moug,nige, omil, paby, ridi, siom, taur, torp, tuni,twol, unar, vota, zous, zuke

Nonwords Used in Experiment 4

Large Dutch N, Large English N: aunk,blag, boul, boup, bret, dris, duef, elap, fram,frip, furk, gonk, jeef, knat, knub, koup, loem,mots, rama, slukLarge Dutch N, Small En-glish N: bito, grul, jees, jeul, kalp, keun, morp,mups, nazz, nont, noto, oune, pris, puif, reug,reun, slen, viem, woup, zulsSmall Dutch N,Large English N: jant, lurp, lusp, naul, nirk,nudo, orim, pani, prad, prog, puet, raut, reud,rion, seto, snam, tirk, tran, vich, vornSmallDutch N, Small English N: drio, frig, jofe,kach, kiot, knaf, luet, maup, moug, nige, omil,paby, ridi, siom, taur, torp, tuni, unar, zous,zuke

REFERENCES

Altenberg, E. P., & Cairns, H. S. (1983). The effects ofphonotactic constraints on lexical processing in bilin-gual and monolingual subjects.Journal of VerbalLearning and Verbal Behavior,22, 174–188.

Andrews, S. (1989). Frequency and neighborhood effects onlexical access: Activation or search?Journal of Exper-imental Psychology: Learning, Memory and Cogni-tion, 15, 802–814.

Andrews, S. (1992). Frequency and neighborhood effects onlexical access: Lexical similarity or orthographic re-dundancy? Journal of Experimental Psychology:Learning, Memory and Cognition,18, 234–254.

Andrews, S. (1997). The role of orthographic similarity inlexical retrieval: resolving neighborhood conflicts.Psychonomic Bulletin and Review,4, 439–461.

Baayen, H., Piepenbrock, R., & van Rijn, H. (1993).TheCELEX lexical database(CD-ROM). Philadelphia:University of Pennsylvania, Linguistic Data Consor-tium.

Baayen, H., Dijkstra, A., & Schreuder, R. (1997). Singularsand plurals in Dutch: Evidence for a parallel dual routemodel.Journal of Memory and Language,37,94–117.

Beauvillain, C. (1992). Orthographic and lexical constraintsin bilingual word recognition. In R. J. Harris (Ed.),Cognitive Processing in Bilinguals(pp. 221–235).Amsterdam: Elsevier Science Publishers B.V.


Beauvillain, C., & Grainger, J. (1987). Accessing interlexi-cal homographs: Some limitations of a language-selec-tive access.Journal of Memory and Language,26,658–672.

Bijeljac-Babic, R., Biardeau, A., & Grainger, J. (1997).Masked orthographic priming in bilingual word recog-nition. Memory & Cognition,25, 447–457.

Carreiras, M., Perea, M., & Grainger, J. (1997). Effects oforthographic neighborhood in visual word recognition:Cross-task comparisons.Journal of Experimental Psy-chology: Learning, Memory and Cognition,23, 857–871.

Coltheart, M., Davelaar, E., Jonasson, J. T., & Besner, D.(1977). Access to the internal lexicon. In S. Dornic(Ed.), Attention and Performance VI(pp. 535–555).New York: Academic Press.

Dijkstra, A., Grainger, J., & Van Heuven, W. J. B.Recog-nition of cognates and interlingual homographs: Theneglected role of phonology.Manuscript submitted forpublication.

Dijkstra, A., & Van Heuven, W. J. B., (1998). The BIAmodel and bilingual word recognition. In J. Grainger &A. Jacobs (Eds.),Localist Connectionist Approaches toHuman Cognition(pp. 189–225). Hillsdale, NJ: Law-rence Erlbaum Associates.

Dijkstra, A, Van Heuven, W. J. B., & Grainger, J. (in press).Simulating cross-language competition with the BilingualInteractive Activation Model.Psychologica Belgica.

Dijkstra, A., Van Jaarsveld, H., & Ten Brinke, S. (1998).Interlingual homograph recognition: Effects of taskdemands and language intermixing.Bilingualism, 1,51–66.

Durgunoglu, A. Y., & Roediger, H. L., III (1987). Testdifferences in accessing bilingual memory.Journal ofMemory and Language,26, 377–391.

Dyer, F. N. (1971). Color naming interference in monolin-gual and bilinguals.Journal of Verbal Learning andVerbal Behavior,10, 297–302.

Ehri, L. C., & Bouchard-Ryan, E. (1980). Performance ofbilinguals in a picture–word interference task.Journalof Psycholinguistic Research,9, 285–303.

Gerard, L. D., & Scarborough, D. L. (1989). Language-specific lexical access of homographs by bilinguals.Journal of Experimental Psychology: Learning, Mem-ory and Cognition,15, 305–313.

Grainger, J. (1990). Word frequency and neighborhoodfrequency effects in lexical decision and naming.Jour-nal of Memory and Language,29, 228–244.

Grainger, J. (1993). Visual word recognition in bilinguals.In R. Schreuder & B. Weltens (Eds.),The BilingualLexicon (pp. 11–25). Amsterdam/Philadelphia: JohnBenjamins Publishing Co.

Grainger, J., & Beauvillain, C. (1987). Language blockingand lexical access in bilinguals.Quarterly Journal ofExperimental Psychology,39A, 295–319.

Grainger, J., & Dijkstra, A. (1992). On the representation anduse of language information in bilinguals. In R. J. Haris

(Ed.),Cognitive Processing in Bilinguals(pp. 207–220).Amsterdam: Elsevier Science Publishers B.V.

Grainger, J., & Jacobs, A. M. (1994). A dual read-out modelof word context effects in letter perception: Furtherinvestigations of the word superiority effect.Journal ofExperimental Psychology: Human Perception andPerformance,20, 1158–1176.

Grainger, J., & Jacobs, A. M. (1996). Orthographic process-ing in visual word recognition: A multiple read-outmodel.Psychological Review,103,518–565.

Grainger, J., & O’Regan, J. K., (1992). A psychophysicalinvestigation of language priming effects in two En-glish–French bilinguals.European Journal of Cogni-tive Psychology,4, 323–339.

Grainger, J., O’Regan, J. K., Jacobs, A. M., & Segui, J.(1989). On the role of competing word units in visualword recognition: The neighborhood frequency effect.Perception and Psychophysics,45, 189–195.

Grainger, J., & Segui, J. (1990). Neighborhood frequencyeffects in visual word recognition: A comparison oflexical decision and masked identification latencies.Perception and Psychophysics,47, 191–198.

Green, D. W. (1986). Control, activation, and resource: Aframework and a model for the control of speech inbilinguals.Brain & Language,27, 210–223.

Grosjean, F. (1998). Processing mixed language: Issues,findings and models. In A. de Groot & J. Kroll (Eds.),Tutorials in Bilingualism: Psycholinguistic Perspec-tives(pp. 225–254). Hillsdale, NJ: Lawrence ErlbaumAssociates.

Guttentag, R. E., Haith, M. M., & Goodman, G. S. (1984).Semantic processing of unattended words by bilin-guals: A test of the input switch mechanism.Journal ofVerbal Learning and Verbal Behavior,23, 178–188.

Jacobs, A. M., & Grainger, J. (1991). Automatic letterpriming in an alphabetic decision task.Perception &Psychophysics,49, 43–52.

Jacobs, A. M., & Grainger, J. (1992). Testing a semisto-chastic variant of the interactive activation model indifferent word recognition experiments.Journal of Ex-perimental Psychology: Human Perception and Per-formance,18, 1174–1188.

Johnson, N. F., & Pugh, K. R. (1994). A cohort model ofvisual word recognition.Cognitive Psychology,26,240–346.

Macnamara, J., & Kushnir, S. (1971). Linguistic indepen-dence of bilinguals: The Input Switch.Journal of Ver-bal Learning and Verbal Behavior,10, 480–487.

Martensen, H., Maris, E., & Dijkstra, A. (in preparation).The body’s role in naming Dutch words: Measurementand effect of phonological consistency.

McClelland, J. L., & Rumelhart, D. E. (1981). An interac-tive activation model of context effects in letter per-ception, Part 1: An account of basic findings.Psycho-logical Review,88, 375–405.

McClelland, J. L., & Rumelhart, D. E. (1988).Parallel dis-tributed processing: Explorations in the microstructure ofcognition(Vol. 3). Cambridge: Bradford Books.


Monsell, S., Matthews, G. H., & Miller, D. C. (1992).Repetition of lexicalization across languages: A furthertest of the locus of priming.Quarterly Journal ofExperimental Psychology,44, 4, 763–783.

Nas, G. (1983). Visual word recognition in bilinguals: Ev-idence for a cooperation between visual and soundbased codes during access to a common lexical store.Journal of Verbal Learning and Verbal Behavior,22,526–534.

Paap, K. R., Newsome, S. L., McDonald, J. E., & Schvan-eveldt, R. W. (1982). An activation-verification modelfor letter and word recognition: The word-superiorityeffect.Psychological Review,8, 573–594.

Ransdell, S. E., & Fischler, I. (1987). Memory in a mono-lingual mode: When are bilinguals at a disadvantage?Journal of Memory and Language,26, 392–405.

Scarborough, D. L., Gerard, L., & Cortese, C. (1984).Independence of lexical access in bilingual word rec-ognition. Journal of Verbal Learning and Verbal Be-havior, 23, 84–99.

Sears, C. R., Hino, Y., & Lupker, S. J. (1995). Neighborhoodfrequency and neighborhood size effects in visual word

recognition.Journal of Experimental Psychology: Hu-man Perception and Performance,21, 876–900.

Smith, M. C. (1997). How do bilinguals access lexical infor-mation? In A. de Groot & J. Kroll (Eds.),Tutorials inbilingualism: Psycholinguistic perspectives(pp. 145–168). Hillsdale, NJ: Lawrence Erlbaum Associates.

Snodgrass, J. G., & Mintzer, M. (1993). Neighborhoodeffects in visual word recognition: Facilitatory or in-hibitory? Memory and Cognition,21, 247–266.

Soares, C., & Grosjean, F. (1984). Bilinguals in a monolin-gual and bilingual speech mode: The effect on lexicalaccess.Memory and Cognition,12, 380–386.

Van Heuven, V. J. (1980). Aspects of Dutch orthographyand reading. In S. F. Kavanagh & R. L. Venezky(Eds.), Orthography, reading, and dyslexia(pp. 57–73). Baltimore: University Park Press.

Ziegler, J. C., Rey, A., & Jacobs, A. M. (in press). Simu-lating individual word identification thresholds anderrors in the fragmentation task.Memory & Cognition.

(Received January 26, 1998)(Revision received April 21, 1998)


orthographic neighborhood effects in bilingual word

Documents