Revisiting the Neighborhood: How L2 Proficiency and Neighborhood Manipulation Affect Bilingual Processing

Mulder, Kimberley; van Heuven, Walter J. B.; Dijkstra, Ton

doi:10.3389/fpsyg.2018.01860

ORIGINAL RESEARCH article

Front. Psychol., 04 October 2018

Sec. Psychology of Language

Volume 9 - 2018 | https://doi.org/10.3389/fpsyg.2018.01860

Revisiting the Neighborhood: How L2 Proficiency and Neighborhood Manipulation Affect Bilingual Processing

$\r\nKimberley Mulder*$ Kimberley Mulder^1*

Walter J. B. van Heuven²

Ton Dijkstra^1,3

¹Centre for Language Studies, for Language Studies, Radboud University, Nijmegen, Netherlands
²School of Psychology, University of Nottingham, Nottingham, United Kingdom
³Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, Netherlands

We conducted three neighborhood experiments with Dutch–English bilinguals to test effects of L2 proficiency and neighborhood characteristics within and between languages. In the past 20 years, the English (L2) proficiency of this population has considerably increased. To consider the impact of this development on neighborhood effects, we conducted a strict replication of the English lexical decision (ELD) task by van Heuven et al. (1998, Experiment 4). In line with our prediction, English characteristics (neighborhood size, word and bigram frequency) dominated the word and non-word responses, while the non-words also revealed an interaction of English and Dutch neighborhood size. The prominence of English was tested again in two experiments introducing a stronger neighborhood manipulation. In ELD and progressive demasking, English items with no orthographic neighbors at all were contrasted with items having neighbors in English or Dutch (‘hermits’) only, or in both languages. In both tasks, target processing was affected strongly by the presence of English neighbors, but only weakly by Dutch neighbors. Effects are interpreted in terms of two underlying processing mechanisms: language-specific global lexical activation and lexical competition.

Introduction

A frequently used metaphor in monolingual and bilingual research is that of lexical activation. Upon the presentation of an input letter string, word candidates in the mental lexicon are assumed to become active depending on their overlap with the input and their frequency of usage. Most researchers nowadays hold that in bilinguals, word retrieval is initially determined by the formal overlap in letters rather than the language to which the word belongs. According to this ‘language non-selective lexical access’ view, in bilingual word reading, word candidates from both languages that are similar to the input are activated in parallel (for an overview of studies, see, e.g., Dijkstra, 2007; Dijkstra et al., 2010).

Most similar within and across languages are words that differ only in a single letter position. These are called ‘orthographic neighbors’ (Coltheart et al., 1977). Words can be neighbors within one language (e.g., light and night in English) or across languages (e.g., night in English and nicht, meaning ‘niece,’ in Dutch). Following a language non-selective access account, upon reading a word, orthographic neighbors from both target and non-target language are activated and influence target word processing. For example, reading the English word wood will activate, besides English form-similar words like good or word, Dutch neighbors like rood (meaning ‘red’) and wond (meaning ‘wound’). Thus, co-activation of lexical representations from different languages occurs if these share enough formal characteristics with the input letter string.

Furthermore, words of a higher frequency may become active more quickly than words of a lower frequency (in terms of an activation metaphor, the former would have a negative ‘resting level activation’ closer to zero, allowing them to become active more quickly). Because native language (L1) words are assumed to have been used more frequently than words from a second language (L2), on average this also holds for L1 vs. L2 words. As a consequence, L1 words are more competitive when they are activated than L2 words, providing an explanation for asymmetric effects in the word retrieval of unbalanced bilinguals, i.e., L1 exerting stronger effects on L2 than vice versa. The same reasoning can be applied to more proficient vs. less proficient bilinguals: The subjective frequency of L2 word usage is higher in the first group, resulting in stronger L2 effects.

Given that each word in a neighborhood (set of neighbors) has its own subjective frequency and associated representation strength, the summed activation of all neighbors in a particular language reflects the strength of the language the words belong to. This makes the manipulation of neighborhoods within and across languages well suited for assessing the relative strength of two languages of the bilingual.

The present neighborhood study had several aims. First, in Experiment 1 we investigated to what extent L2 proficiency differences can affect the occurrence of within- and between-neighborhood effects in L1 and L2. By fully replicating an earlier experiment by van Heuven et al. (1998), we tested the hypothesis that an increase in English (L2) proficiency in Dutch-English bilinguals shifts the relative contribution of their two languages toward English in English lexical decision (ELD) (see Introduction Experiment 1).

Second, in Experiment 2 we tested the effect of stimulus properties on neighborhood effects in a very similar ELD task by means of a stronger type of neighborhood size manipulation. Specifically, we introduced a manipulation in terms of hermit words that have no neighbors at all in one of the languages (see Introduction Experiment 2).

Third, in Experiment 3 we tested if the effects observed in ELD (Experiment 2) could be generalized across tasks by including the same materials in English progressive demasking (EPDM). This would also suggest that in our relatively English-proficient bilinguals, the native language Dutch (L1) does not exert strong effects in tasks in which only English (L2) words occur (see Introduction Experiment 3).

Fourth, by contrasting neighborhood effects for target items with few (Experiment 1) and with no (Experiments 2 and 3) neighbors, we wished to clarify both theoretical and empirical issues with respect to neighborhood studies. From a theoretical perspective, the comparison across item and task types allows us to analyze the processing mechanisms underlying performance in more detail. From an empirical perspective, such an analysis will help to clarify the puzzling finding of some fragile neighborhood effects in studies such as van Heuven et al. (1998) and Dirix et al. (2016).

To set the stage for a more detailed description of our experiments later, we will summarize the limited set of available bilingual neighborhood studies here. Studies on bilingual neighbors have been scarce and, as far as we know, restricted to van Heuven et al. (1998), Midgley et al. (2008), Grossi et al. (2012), Van Kesteren et al. (2012), Oganian et al. (2015), Dirix et al. (2016), and Oganian et al. (2016).

van Heuven et al. (1998) presented the first large study that examined effects of within- and between-language neighborhood size on bilingual word recognition by manipulating the number of English (L2) and Dutch (L1) orthographic neighbors in progressive demasking, generalized lexical decision, and language-specific lexical decision tasks as performed by Dutch–English bilinguals. When English target words had more orthographic neighbors in Dutch, this systematically resulted in slower response times, while a larger number of English neighbors produced facilitatory effects for English target words in progressive demasking and Dutch–English generalized lexical decision (in which a positive response is required for both Dutch and English words). Remarkably, this was not the case for language-specific ELD, for which a puzzling significant English neighborhood size effect of only 3 ms was reported (in the participant analysis only).

In fact, across the study as a whole, the observed effects were relatively large for the non-target language, which was the native language Dutch (see Tables 2, 3, and 8 below). The between-language (i.e., Dutch) neighborhood size effects disappeared in monolingual English speakers processing the same materials, but for them facilitation arose for within-language (English) neighborhood size.

More recently, Midgley et al. (2008) provided electrophysiological evidence for cross-linguistic neighborhood effects. They specifically focussed on the N400, an EEG-component that is sensitive to semantic aspects of word processing (e.g., Kutas and Hillyard, 1980). N400 amplitude is assumed to reflect how easily a word can be semantically integrated into the context, be it a single word, a sentence, or a discourse (Kutas and Federmeier, 2000, p. 464). In addition, the amplitude of the N400 is found to be larger when target words have more semantic associates (Kounios and Holcomb, 1992). In a monolingual study on neighborhood effects, Holcomb et al. (2002). observed that words with a larger number of orthographic neighbors resulted in greater semantic activation and, as a consequence generate larger N400s (cf. Mulder et al., 2013). Following Holcomb and Grainger (2007), who argued that the N400 reflects the mapping of whole-word form representation onto semantics, Midgley et al. (2008) hypothesized that “larger N400s for words with many orthographic neighbors would reflect inhibition across activated lexical representations that leads to increased difficulty in settling on a unique form-meaning association.” This mechanism was hypothesized to hold for both within-language and between-language neighborhood effects.

ERP-recordings of highly proficient French-English bilinguals reading in French or English revealed that words with many between-language neighbors generated a more negative-going ERP waveform in the N400 region than words with few between-language neighbors. Moreover, the between-language neighborhood size effects in the N400 ERP-component arose earlier and were more widely distributed for L2 (English) target words than L1 (French) target words. The authors concluded that “words with more cross-language neighbors suffer from the co-activation of the lexical representations of these neighbors, as reflected in the typically longer RTs found to these stimuli in behavioral studies […]”.

Grossi et al. (2012) partially replicated these results in an ERP-study with English-Welsh bilinguals who performed a semantic categorization task on English and Welsh words. In late bilinguals, words with many between-language neighbors elicited more negative ERP amplitudes than words with few of them between 175 and 500 ms after word onset. In the 300–500 ms window, this effect interacted with language (English or Welsh). Early bilinguals showed a more complex pattern of early effects and no N400 effects. To explain their findings, the authors suggest that activation of between-language orthographic neighbors is sensitive to how bilinguals learn and use their languages.

Van Kesteren et al. (2012) studied Norwegian–English bilinguals in a mixed ELD task and a mixed Norwegian lexical decision task using English and Norwegian word stimuli that included language-specific letters (“smør,” “hawk”) and bigrams (“dusj,” “veal”). The number of neighbors in English and Norwegian in these tasks was systematically manipulated. This manipulation led to null-results of neighborhood size, possibly because other sub-lexical markers of language membership (i.e., language-specific letters and bigrams) were more prominently used by the bilinguals under consideration.

Dirix et al. (2016) conducted a generalized Dutch-English lexical decision experiment as well as a large-scale eye-tracking study in which Dutch–English participants read the Dutch (L1) or English (L2) version of a novel by Agatha Christie. The generalized lexical decision experiment was comparable in stimulus materials and several other respects (but not analysis) to van Heuven et al. (1998) (Experiment 3). In line with van Heuven et al. (1998) a mixed-effect model analysis yielded an inhibitory effect of Dutch neighborhood density on English RTs for words with low bigram frequency, and a higher error rate on English words with more cross-linguistic neighbors. Unexpectedly, this finding was not paralleled by a main effect of Dutch neighborhood density in Dutch lexical decision RTs, nor by any significant effect of English. The authors ascribed this discrepancy for Dutch (L1) words relative to the monolingual literature as due to the use of a generalized lexical decision task, “which creates a bilingual context different from a normal unilingual lexical decision task.” Although this suggests there may be interactions between the effects of L1 and L2 neighbors under particular task conditions, these were not further considered. Remarkably, neither van Heuven et al. (1998) in ELD, nor Dirix et al. in generalized Dutch-English lexical decision reported a straightforward English (L2) neighborhood effect on the RTs. The results of the study became even more puzzling in Experiment 2, because in natural English reading in a one-language context, the presence of between-language neighborhood effects was confirmed, but the effects were largely facilitatory (rather than inhibitory) in nature.

Finally, Oganian et al. (2015, 2016) observed between-language neighborhood size effects in naming and language decision of language-specific and language-ambiguous pseudo-words. They found that neutral pseudo-words were preferentially categorized to the language that was predominant in their orthographic neighborhood. In addition, they observed that the processing of L1-marked pseudowords but not L2-marked pseudowords were affected by the number of orthographic neighbors from the two languages. This suggests that perception of L2 markers was sufficient to trigger language decisions, whereas the activation of lexical neighbors seemed to influence the decision process for L1 marked pseudowords. Thus, the authors suggest that between-language activation of the L1 may be restricted to cases of sublexical ambiguity, whereas activation of lexical representations may concern especially the presented L2 when the associated orthographic patterns are illegal in L1.

In sum, with one exception, factorial studies on between-language neighborhood size indicate that bilingual word recognition in a non-native language is indeed sensitive to the numbers of words (neighbors) similar to the target word in both their languages. This validates the manipulation of neighborhood density as a marker of the relative contribution of two languages to bilingual word recognition. At the same time, the puzzling results for within- and between-language effects of L1 neighbors, and the potential sensitivity of effects to task demands (e.g., generalized lexical decision vs. language-specific lexical decision), call for further research. In the present paper, we first replicate the ELD task (Experiment 4) by van Heuven et al. (1998) with the present generation of the same bilingual participant population. Next, we report on an ELD task (Experiment 2) and a progressive demasking task (Experiment 3) with a stronger neighborhood manipulation in terms of hermits.

Experiment 1: English Lexical Decision With Neighbors

To the best of our knowledge, no published study has yet replicated van Heuven’s et al. (1998) findings of bilingual neighborhood effects in ELD. We conducted an exact replication of the experiment by van Heuven et al. (1998, Experiment 4) 20 years later. This experiment involved a contrast between large and small neighborhoods for target items in both L1 and L2. In our study, a new generation of the Radboud University psychology student population was tested in Nijmegen. Because we had full access to the study of 1998, we were able to replicate the original experiment in the greatest detail, using exactly the same procedure and even identical stimulus lists.

In principle, two different predictions can be formulated with respect to the outcome of the replication. First, one might expect exactly the same result pattern as 20 years ago. However, models such as BIA/BIA+ propose that the activation of words depends on their frequency of usage. When participants possess a stronger proficiency in English (L2), the relative subjective frequency distribution of English and Dutch words shifts. Subsequent lexical activation differences might result in faster response times in an English task, and more prominent English and less prominent Dutch effects of neighborhood, bigram and word frequency. We are in favor of this second account, because there is abundant evidence to suggest that current Dutch students are more proficient in their L2 English than those 20 years ago. In 2001, a large survey by the European Commission, entitled “Europeans and their languages” (Special Eurobarometer 147, 2001, p. 16), reported that 52.1% of the Dutch claimed to have a ‘good’ level of English and 20.1% claimed a ‘very good’ level. The same question in Special Eurobarometer 386 (2012, p. T67) elicited claims of 58% ‘good’ and 32% ‘very good.’ According to this last survey (2012, p. 171), young people in Europe also judge themselves better on all dimensions of multilingual communication in a second language than older people (e.g., 41% vs. 20% of the two groups indicate they can follow English news reports via radio and television). In sum, there is a strong cross-generational difference in L2 proficiency.

In sum, we predict that in our present generation of Dutch psychology students, the relative strength of English to Dutch has increased (even when they can still be considered as late and unbalanced bilinguals). This should result in relatively strong effects of English in our Experiment 1, a replication of van Heuven et al. (1998), but also, even more clearly, in our later hermit experiments (including a stronger manipulation of neighborhood size).

Method

Participants

Thirty-two Dutch L2 speakers of English (mean age 23.7 years old, SD = 3.34), mostly undergraduates at the University of Nijmegen, were paid or received course credits to take part in this experiment. All were highly proficient in English, having learned English from the age of 11 onward. All had normal or corrected-to-normal eyesight. Care was taken to select participants with the same characteristics as those in van Heuven et al. (1998).

Materials and Procedure

All stimulus materials were identical to those in van Heuven et al. (1998, Experiment 4). Stimulus characteristics are summarized in Table 1. In van Heuven et al. (1998) the numbers of neighbors in each language were calculated following Coltheart et al. (1977). The item set consisted of 20 word and 40 non-word items in each condition. Stimulus lists, including stimulus order, in the present experiment were identical to those in the earlier study. The procedure followed was also identical, except that as a background survey the present experiment involved the Lextale task (Lemhöfer and Broersma, 2012) to assess English proficiency, and two language background questionnaires (one of which was identical to that used in van Heuven et al., 1998).

TABLE 1

TABLE 1. Stimulus characteristics in the neighbor manipulation (van Heuven et al., 1998, Experiment 4; Experiment 1) and the hermit manipulation (Experiments 2 and 3).

Participants performed an English visual lexical decision task, which was programmed in Psychopy and run on an HP Compaq Intel Core 2 computer with LCD monitor and a refresh rate of 120 Hz. The experimental set-up and stimulus presentation (font size and type of stimuli, background color, instructions, trial structure, etc.) were identical to those in van Heuven et al. (1998).

Results

The mean participant and item accuracy was 92.37, and 93.81%, respectively. One participant (47.8% correct) and one item (61.29% correct) that had an error rate above 30% were removed (e.g., keen). Finally, errors and RTs outside the range of 2.5 SD from the item and participant mean were removed. Tables 2, 3 present the mean RTs, standard deviations, error rates, and neighborhood effects for different word and non-word types. For a comparison of neighborhood effects based on different neighborhood size contrasts, the mean RTs of the lexical decision data of Experiment 4 of van Heuven et al. (1998) are also presented¹.

TABLE 2

TABLE 2. Mean RTs (in ms), standard deviations, error rates, and neighborhood effects for English word stimuli of English Lexical Decision in van Heuven et al. (1998, Experiment 4), our replication study (Experiment 1), and Experiment 2 with hermits.

TABLE 3

TABLE 3. Mean RTs (in ms), standard deviations, error rates, and neighborhood effects for English non-word stimuli of English Lexical Decision in van Heuven et al. (1998, Experiment 4) and our replication study (Experiment 1), and Experiment 2 with hermits.

Inspection of the distribution of the response latencies revealed non-normality. A comparison of a log-transform and an inverse transform (-1000/RT) revealed that the inverse RT was most successful in reducing the non-normality. The word and non-word data were then analyzed with linear mixed effects models with subject and item as crossed random effects. Similar to van Heuven et al. (1998) the following factorial predictors were considered in the analyses: English Neighbors (Large or Small) and Dutch Neighbors (Large or Small). Further, we added the following continuous predictors to our model in a step-wise inclusion procedure: English Frequency (log-transformed subtitle frequency, SBTLWF; Brysbaert and New, 2009), English Bigram Frequency and Dutch Bigram Frequency (both log-transformed; Duyck et al., 2004), Trial (the rank of the item in the stimulus list), and Previous RT (the log-transformed response latency at the previous trial).

We included English and Dutch bigram frequency as factors in our analyses, because bilinguals can use sublexical statistical information such as bigram frequency to identify language membership (Oganian et al., 2016). In addition, in the study by Dirix et al. (2016), this variable contributed relatively strongly to the obtained data patterns. For the random effects structure, we considered random slopes by participant for all predictors mentioned above.

To obtain the best fitting model, we performed a stepwise variable selection procedure in which one predictor was added at a time. For each significant predictor or interaction, it was evaluated whether inclusion of this predictor or interaction resulted in a better model (i.e., had a lower AIC compared to when this predictor was not part of the model). Next, the final model was trimmed by removing any remaining extreme outliers (defined as data points with standardized residuals exceeding 2.5 standard deviation units).

Tables 4, 5 summarize the final models for the word and non-word analyses, respectively. The final regression model for the word data in Table 4 revealed a significant interaction between English Neighbors and English Bigram Frequency, showing that response latencies are faster when the English bigram frequency and English neighborhood size is large compared to when the English neighborhood size is small. Figure 1 displays this interaction. Furthermore, English Frequency had a facilitatory effect on RT. Finally, the effect of PreviousRT shows that responses become slower when the response to the previous item was also slow.

TABLE 4

TABLE 4. Final model for the word data in Experiment 1 (English Lexical Decision with neighbors).

TABLE 5

TABLE 5. Final model for the non-word data in Experiment 1 (English Lexical Decision with neighbors).

FIGURE 1

FIGURE 1. The significant interaction between English Neighbors and Log English Bigram Frequency in English Lexical Decision with neighbors (Experiment 1).

The final model for the non-word data revealed a significant interaction between English Neighbors and Dutch Neighbors, indicating that responses times are slower when both English and Dutch neighborhood size are large. Figure 2 displays this interaction. The inhibitory effect of Previous RT shows that responses become slower when the response to the previous item was also slow.

FIGURE 2

FIGURE 2. The significant interaction between English Neighbors and Dutch Neighbors in English Lexical Decision with neighbors (Experiment 1).

Finally, as Tables 2, 3 reveal that our participants are considerably faster than the participants in van Heuven et al. (1998) effects of Dutch neighborhood size might occur only in the slower participants However, a median split of our data in a fast (mean RT = 519) and slow group (mean RT = 590) again revealed no effects of Dutch neighborhood density in both groups.

Discussion

The results of our ELD experiment, a replication of van Heuven et al. (1998, Experiment 4), are in line with our prediction that Dutch-English bilinguals have become better in English in the last 20 year. With respect to English and Dutch neighborhood effects and (sub)lexical factors, a shift toward English was observed in the response patterns for words and non-words. Only a small interaction effect of English and Dutch neighbors was observed in the non-words. We conclude that, apart from stimulus properties, relative L2 proficiency is an important determinant of neighborhood effects, explaining in part why studies involving similar tasks and designs may still obtain different results². Thus, future studies should be even more strict in their experimental manipulations and characterize their participant groups in as much detail as possible.

We consider the prominence of English in our participants to be a consequence of more intensive contact of the present generation of students with English in school, due to English university books, and/or due to the English-oriented Internet. Since 1993, Dutch children start to acquire English as their second language at least 2 years earlier than before, at the age of 10–11 rather than 12–13, due to a change in school systems and the strong increase in so-called ‘bilingual schools’ (Van Hell, 1998; Edelenbos and Vinjé, 2000). In addition, in the last decade, English has become default for Dutch students when they are searching the internet, and many Bachelor programs at Dutch universities are now taught in English. This has even instigated a debate on the role Dutch should play in the scientific education in various disciplines. We propose that this Dutch trend toward increased bilingualism, signaled by De Swaan, 2001 (p. 202), continues strongly today. Seen in this light, the comparison of the two studies shows how societal changes affect the L2 proficiency of participant populations, resulting in systematic differences in observed data patterns over time.

When the present generation of Dutch–English bilinguals have a stronger representation of English, its effect should become even more prominent when a stronger neighborhood manipulation is applied. Such a manipulation is that of hermit neighbors in Experiment 2 (English lexical decision) and Experiment 3 (English progressive demasking).

Experiment 2: English Lexical Decision With Hermits

In Experiment 1, we manipulated neighborhood size in terms of many vs. few neighbors. All studies so far used this specific manipulation, although they differed in other respects (e.g., participant groups, language pairs, and experimental techniques). However, Bowers et al. (2005) have argued that the critical and optimal neighborhood contrast to consider is not between words with many and few neighbors, but between words with one or more neighbors and with no neighbors. They pointed out that word processing models like IA and SOLAR predict little difference between words with few and many neighbors (Davis and Andrews, 1996), because there is no additional competition for words with many neighbors due to a normalization of the total amount of activity at the word level. Thus, in order to have a pure measurement of neighborhood size effects, words with one or more neighbors should be contrasted with words with no neighbors, the so-called hermit words.

Bowers et al. (2005) addressed this issue by having monolingual English participants learn new words (e.g., banara) that were neighbors of familiar hermit words (e.g., banana) and respond to these familiar words in a semantic categorization task. They observed that repeated exposure to the novel neighbor word made it more difficult to semantically categorize the familiar words. Interference effects even became larger with more training on the novel words. The authors concluded that the impact of the new neighbors on semantically classifying the hermit words is likely to reflect lexical competition and is in accordance with the predictions made by the IA and SOLAR models.

To include the strongest test of neighborhood effects possible in our study, we therefore contrasted word conditions with many or no neighbors at all in English and Dutch in Experiments 2 and 3. This resulted in four conditions: English words without any orthographic neighbors in English or Dutch, referred to as complete hermits; English words with neighbors in Dutch but not in English; English words with only English and no Dutch neighbors; and English words with neighbors in both languages. By comparing complete hermits to words that are hermits only in Dutch, we can directly assess the role of Dutch neighbors, while a comparison to hermits only in English should directly reflect effects of English neighbors. This should allow us to test the occurrence of between-language neighborhood size effects with a more pure contrast of neighborhood size than before.

As tasks, we chose ELD and progressive demasking, because in van Heuven et al. (1998; Experiments 1 and 4), both of these tasks included exclusively English (L2) words, while the non-words in ELD were also derived from English. In this case, it can be clearly seen to what extent the native language of our participants, Dutch, is able to affect non-native English language processing. A further reason to opt for the language-specific lexical decision task was that, in contrast to predictions, van Heuven et al. (1998) observed only a small effect of English neighbors in their English target word responses [(significant only in the participant analyses, F1); see Table 2]. It is therefore important to demonstrate that within-language neighborhood effects of English can be obtained by contrasting words with many and no neighbors. Finally, by applying the same neighborhood contrast to non-word stimuli in lexical decision, we cannot only collect control data for comparison with the word data, but also obtain more insight into neighborhood effects for targets without lexical representation and linked to a no-response.

Furthermore, while language-specific lexical decision requires a forced choice between two responses, in a paradigm such as progressive demasking, a target word must be identified in a background of noise (see Keuleers and Brysbaert, 2012). Thus, by conducting a progressive demasking experiment involving the same stimulus materials, we can assess the effect of task differences on the obtained result patterns. This will also help to better understand which mechanisms underlie performance in different language and task situations. Finally, if Experiments 2 and 3 with the hermit manipulation both demonstrate that, relative to Dutch (L1) neighbors, English (L2) neighbors exert a stronger effect on the RTs than in the replication by van Heuven et al. (1998, Experiment 1), this provides evidence that neighborhood effects are sensitive to subtle properties of the stimulus materials, in particular the degree to which they activate the background language Dutch (the strong L1) in a task requiring responses to the target language English (the weaker L2).

Note that in our hermit experiments, we applied exactly the same method for calculating within-language and between-language neighbors as van Heuven et al. (1998) and also preserved their four experimental neighborhood conditions. The only difference was that in our ‘small’ English and ‘small’ Dutch neighborhood conditions, neighborhood density was zero instead of one or more. This allowed us to see whether a stronger neighborhood contrast would lead to the same pattern of effects. Furthermore, the hermit conditions allowed a purer assessment of the independent effects of English and Dutch neighborhood density, particularly in the comparison of words with no neighbors at all to words with neighbors in one language only.