Edited by:
Reviewed by:
*Correspondence:
This article was submitted to Language Sciences, a section of the journal Frontiers in Psychology.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Second language learners face a dual challenge in vocabulary learning: First, they must learn new names for the 100s of common objects that they encounter every day. Second, after some time, they discover that these names do not generalize according to the same rules used in their first language. Lexical categories frequently differ between languages (
Second language acquisition research has often highlighted the role of learners’ language history as a strong predictor of ultimate second language (L2) attainment in syntax and phonology (e.g.,
A closer examination of lexical semantics reveals, though, that the development of the lexicon may be more analogous to that of syntax and phonology than such divergent outcomes suggest. Recent research in lexical categorization has moved beyond the size of learners’ vocabularies and investigated more subtle aspects of word knowledge such as lexical category boundaries in both native and L2 speakers of a language. The studies reviewed below have found significant variation in lexical categorization patterns among native speakers, simultaneous bilinguals, and sequential bilinguals as a function of predictors such as age of onset, language learning experience, and usage patterns. In this paper we examine determinants of L2 lexical acquisition in more detail, with emphases on L2 immersion experience and its interaction with both individual bilinguals’ language histories and the word use patterns of the linguistic communities in which both first and second language are acquired.
Decades of research have indicated differences in lexical categorization across languages (such as the seminal comparison of color categories by
Learners of a second language, including children who acquire two languages simultaneously, are thus faced with a major incongruity between languages. For example, Chinese (referring to Mandarin Chinese throughout this paper) and English differ in the principal features by which containers are categorized. Native Chinese speakers use
Lexical categorization is a valuable tool for identifying variation in lexical semantic mappings among speakers, and with this more sensitive measure of lexical semantic variation, second language lexical proficiency may no longer be sufficiently described by the accumulation of a list of words as tested by most picture naming, lexical decision, and fluency tasks. Instead, lexical semantic mappings are more precisely probed when many similar objects are named, which allows inferences about the boundaries of a given speaker’s lexical category. For instance, the researcher can examine which drinking vessels are named
Recent work has investigated whether and how bilinguals can maintain native-like lexical semantic representations in each language despite these differences.
These cross-language transfer and convergence effects can be thought of in terms of how exposure to one language might change mappings from objects’ representative features to words in the other language of the bilingual speaker. Theoretical models of lexical semantic representation, such as
At least two computational models have attempted to simulate bilingual lexical categorization (
Although there are only a few quantitative accounts of bilingual lexical categorization, a number of likely predictors for development of lexical categories are apparent from the broader study of second language acquisition. The extent of L2 immersion, age of second language onset, time spent learning the second language in a formal setting (classroom training), and patterns of language use (the extent to which the languages are intermixed in use) all appear to be involved in non-native learners’ degree of success in learning a second language. Further, because name choice for an object may vary across speakers (e.g.,
While many studies in second language acquisition explore the influence of language history variables on lexical learning, fewer studies have evaluated a combination of such variables simultaneously and properly controlled for interaction among the variables and statistical obstacles to measuring effects of variables individually, as outlined by
The value of L2 immersion is uncontroversial in second language acquisition research with respect to many components of L2 acquisition. Recent findings in lexical categorization suggest that as in other domains of language acquisition, native-like L2 lexical categorization is supported by L2 immersion.
Within-category variation arises constantly as part of the natural environment, as one may have occasion to sit in several different
Age of second language onset as a predictor of eventual second language attainment remains a controversial topic, as evidence for and against a sensitive period for language acquisition is weighed alongside varying levels of other confounding age-related variables (such as years of L2 exposure, motivation, and socialization; see a recent review in
However, these tests do not account for between-language variation in lexical semantic mappings and may overlook non-native word uses by older speakers who rely on direct translation for L2 learning. The relationship between age effects and native-like lexical categorization performance is not entirely clear. Although
The possible departure from conventional “earlier is better” wisdom about age of onset raises questions about whether simultaneous bilinguals are unique in their degree of convergence between languages. If late bilinguals show diminished convergence, more native-like representations may be learnable in both L1 and L2 independently, even when marginal cross-language transfer is observable.
The latter result does not strongly contradict the
The type of language experience gained in an immersion environment can vary substantially among bilinguals. Simultaneous bilinguals, such as those in
The monolingual or bilingual context of the language environment or the extent to which speakers switch between languages changes the degree of cognitive control necessary for language production. Specifically, highly bilingual environments raise the potential for frequent code-switching and increase activation of the non-target language, which must then be actively inhibited from production (
Evidence supporting the view of language change through use can be found in a recent study of phonological accent in the native language.
The contrast observed in
Because native, monolingual speakers of a language also show significant variation in lexical categorization patterns, even monolingual infants acquiring their native language are exposed to variable input for many objects’ names. In the relatively familiar domain of household containers,
As we have discussed earlier, bilinguals’ lexical categorization patterns in either language are, indeed, jointly predicted by the native (monolingual) patterns of the two languages (
It is evident that in many instances of simultaneous and sequential bilingualism, the category information provided to bilinguals by the native-speaker communities of each language is variable and yet still bears a significant influence on their production in both languages. With relatively few lexical category stimulus sets normed for native speakers of more than one language and tested on sufficiently advanced bilinguals of both languages, the exact degree and means of this cross-language influence remains to be explored. However, native category norms that represent the full distribution of names produced and thus the degree of name agreement and variation among native speakers may allow an elaborated view of cross-language competition and transfer. The extent to which L1 representations are vulnerable to change may vary as a function of their own entrenchment, with greater native naming agreement representing more robust L1 representations. Conversely, objects named with greater consistency in L2 (high L2 native agreement) could be associated with better learning outcomes as compared to objects for which L2 speakers show little agreement.
In the present study, we aim to disentangle the respective roles of four broad categories of individual language history variables in predicting native-likeness of L2 lexical semantics: L2 environment (non-immersion vs. immersion), age of L2 onset, years of L2 classroom study, and L2 usage pattern [code-switching frequency (CSFreq)]. Collinearity between age and immersion predictors has been shown to cause serious confounds in studies of second language acquisition (see
By measuring several language history variables together, accounting for the earliest L2 exposure (that is, L2 onset before immersion), and using categorization as a more sensitive measure to inter-personal lexical semantic variation, we aim to make better statistical estimates of each variable’s effect. We offer a simultaneous measure of four variables based primarily on the self-reports of Chinese–English bilinguals resident in Beijing, China and in Pennsylvania, United States.
We also introduce linguistic community norms for word use in L1 and L2, derived from native speakers of each language, as possible predictors of bilinguals’ lexical categorization patterns. The contribution of such norms has rarely been considered in predicting L2 performance (except see
These non-immersed and immersed participants are compared in an L2 (English) lexical categorization task that has proved highly sensitive to variation in lexical semantic mapping for other populations of bilinguals. Based on the simultaneous evaluation of all four language history variables and the linguistic community norms, we evaluate participants’ English native-likeness on the lexical categorization task. We offer an interactive account of how various aspects of one’s native language, second language, and language learning history jointly influence the lexical semantic mappings that defines object naming, a behavior that occurs often in our daily experience.
Two groups of bilingual students, one in the United States and one in China, participated in this study. In the U.S., Chinese–English bilingual undergraduate and graduate students were recruited from the Introduction to Psychology subject pool and through posters around the campus community at Penn State University (State College, PA, USA). In China, Chinese–English bilingual undergraduate and graduate students were recruited through an online campus message board (BBS) and through personal referrals at Beijing Normal University (Beijing, China). Generally speaking, the students at Penn State were slightly younger (mostly undergraduates) than those at Beijing Normal (mostly graduate students), were first exposed to English at a slightly earlier age, and had higher self-rated proficiencies in English.
Although many of the bilingual participants reported some degree of training in a third language, most rated themselves at very low proficiency. Participants who self-reported a proficiency of 2.5 or greater in the third language on a 7-point scale (averaged across four ratings: reading, writing, speaking, listening) or failed to provide a proficiency rating in their third language were not included in the data. In total, 57 participants from Beijing Normal and 68 participants from Penn State met the inclusion criterion. Third languages included French, German, Russian, Mongolian, Japanese, Korean, Taiwanese, and Cantonese.
Penn State students ranged in age from 18 to 23 (
We also drew on a set of native-speaker norming data from functionally monolingual participants who had participated in a previous version of the lexical categorization task, using the same stimuli (
All participants completed a language history questionnaire (LHQ;
Early trials at Penn State revealed that several participants failed to complete the code-switching section of the LHQ or claimed to never code-switch, a self-report that may (in some cases) underestimate the true rate of code-switching in cultural environments that stigmatize language mixing. An additional code-switching questionnaire (CSQ) was added to subsequent sessions to specifically probe participants’ code-switching and was administered according to the dominant language environment. A single item on the CSQ was used to obtain a point-estimate of participants’ overall CSFreq: “Do you use English words when speaking Chinese, or do you use Chinese words when speaking English?” rated on a five-point ordinal scale with response options from “never” to “very often.”
Sixty-seven photographs of common household objects were used to elicit category names from monolingual and bilingual participants. These objects were drawn from a stimulus set (called the dish set) used by
An Operation-Span (O-Span) test was also used to screen the bilingual participant groups for systematic differences in working memory, a cognitive factor that might be confounded with language proficiency or language transfer. The O-Span includes mathematical and verbal components (
After giving informed consent in the local language, participants completed the LHQ, also in the local language (Chinese or English). They then completed an unrelated English receptive vocabulary task (results not presented here) to establish an English language mode to the extent possible in both the Chinese- and English-immersed participants. After the vocabulary test, all participants performed the English picture naming task. The Chinese O-Span was then completed and used to shift participants into a Chinese language mode before naming the objects again in Chinese. Finally, the CSQ was completed last. Participants in the US completed English and Chinese tasks on separate days, 1–2 weeks apart (range: 6–21 days; mean: 9 days) and counter-balanced for order. Sessions in China could not be scheduled separately and all tasks were completed on the same day, with English first, followed by Chinese. We reasoned that the English task was less likely to influence Chinese naming in a Chinese immersion environment, and intervening Chinese tasks (namely, the O-Span) would help to reduce any language priming effects.
In the picture naming tasks, participants were instructed to name aloud photographs of objects depicted on the computer. They were asked not to name the objects’ contents, as illustrated by two photographic examples: a grocery bag full of vegetables (called
Participant responses were transcribed from audio recordings by high-proficiency Chinese–English bilinguals in the United States who were able to comprehend Chinese responses and phonetically accented English responses. Transcribers were not able to view the objects during transcription to prevent bias on ambiguous recordings. Transcribed responses were subsequently reduced to head nouns (e.g., “a small blue bowl” is reduced to “bowl”) for comparison with the native norming data. Skipped trials, inaudible responses, and irrelevant responses (e.g., “I don’t know”) were entered as blanks and treated as missing data.
Four biographical variables were included for each subject: Age of first exposure to English (AOEE), LOR in the English immersion environment (LOR), self-reported frequency of code-switching between Chinese and English (CSFreq) and the total number of years spent learning English (current age minus the age of first exposure, YrsLearn). For participants who failed to complete some language history and code-switching questions, missing data for the CSFreq variable were replaced with the sample mean (3% of the participants included in the analysis). Participants who did not report AOEE were excluded from the analysis (eight participants), and an additional set of early childhood bilinguals (AOEE < 5 years, six participants) were removed from the US participant data to maintain comparability with the sample in China (AOEE range 5–15 years).
Given the above exclusion/inclusion criteria for data analysis, our data analyses presented in the Results section were based on a total of 30 participants from China and 33 from State College (see
Demographics and language histories of participants before and after screening.
Sample | Age (SD) | AOEE (SD) | EngProf (SD) | CSFreq (SD) | LOR (SD) | |
Beijing Normal | 57 | 22.9 (1.8) | 11.6 (1.9) | 4.1 (1.0) | 1.1 (1.2) | 0 |
Penn State | 68 | 20.9 (2.9) | 8.8 (3.3) | 4.7 (1.1) | 1.8 (1.2) | 3.8 (5.2) |
Beijing Normal | 30 | 22.8 (1.7) | 11.5 (2.2) | 4.2 (1.1) | 0.95 (1.1) | 0 |
Penn State | 33 | 21.8 (3.3) | 9.8 (2.6) | 4.7 (1.0) | 1.9 (0.8) | 2.2 (2.6) |
In the following sections, we present a set of analyses that examine the lexical categorization patterns of the Chinese–English bilingual participants at three different levels, as follows. (1) The group-wise analysis compares the overall patterns of transfer and convergence between Chinese and English as spoken by the bilingual participants. This analysis looks at the overall trends in naming distributions generated by sub-groups, which is defined by their degrees of L2 immersion (see details below). This analysis allows direct correlations of the bilinguals’ overall patterns with the monolingual norms. (2) The subject-wise analysis focuses on individual bilingual participants’ language histories and how these variables predict their individual differences in L2 naming patterns. (3) The item-wise analysis examines naming performance on each object of the stimulus set, controlling for variation in individuals’ language histories and examining the impact of linguistic community norms on the bilinguals’ accuracy in producing the native preferred L2 names.
For group-wise comparison, participants were organized by three discrete values of LOR to describe three types of immersion conditions observed in our sample: No Immersion, Short-term, and Long-term. No Immersion was defined by LOR = 0, describing participants who have never lived in an English immersion environment. English-immersed participants were divided into two groups by a median split (median non-zero LOR = 1.3 years). Short- and Long-term Immersion were defined as the samples below and above the median, respectively.
A cross-language correlation matrix was calculated for each bilingual and monolingual group according to the method of
The cross-language correlations revealed that native, monolingual speakers of Chinese and English correlate in their categorization of this set of objects at
Surprisingly, the No Immersion group showed the highest convergence between their two languages (0.95), a relatively low correlation with the monolingual Chinese (0.80) compared to their recently immersed peers (Short Immersion, 0.85,
Participants’ picture naming responses were compared to a set of English native norms to generate a score for each participant describing the English native-likeness of their lexical categories. Each of a participant’s responses was awarded a score based on the proportion of native monolingual speakers who produced that same response in the norms (following
We estimated a linear regression model for the English native-likeness scores over participants’ language histories to determine the relationships between language background and attained L2 lexical category proficiency. Previous analyses from smaller datasets showed several two-way interactions between the language history variables and an inter-dependency of the significance of these interactions in the model (see
In an attempt to improve the parsimony of this model without discarding important interaction effects, an automatic Akiake information criterion (
To understand the highly interactive terms of the subject-wise model, we generated several estimated marginal means plots based on the model’s predicted English native-likeness scores across a range of values for the two-way interactions between L2 immersion (LOR) and each remaining predictor (while holding other predictors constant at the mean value). These two-way interactions were all highly significant (LOR × YrsLearn:
The first plot (
Age of earliest English exposure (AOEE;
Participants’ self-reported CSFreq was also a significant predictor in the model and significantly interacted with immersion. As
In this analysis looking at how native naming consensus for objects impacts the likelihood of naming objects correctly, we compared each response by the participants to the single dominant name
In the first logistic regression, we entered the same language history variables used in the subject-wise analysis to determine how adequate these variables were for identifying variation in native-like categorization for different objects. The logistic regression model including only participants’ language history information contained several statistically significant predictors, but offered a very poor fit to the data (Nagelkerke
In the next analysis, we added four language variables which described the native speaker norms for every given object: naming agreement in Chinese (L1), naming agreement in English (L2), number of alternative names produced by the Chinese norming group, and number of alternative names produced by the English norming group. Due to computational limitations, this model was estimated with up to four-way interactions and reduced using the same
As in the subject-wise analysis, the model contained many interaction terms that impeded interpretation without isolating a few of the variables. Again, we sought to describe how immersion experience affected the role of these language variables in predicting the participants’ success in producing native-like English names for objects. A binomial logistic regression predicts the probability that an outcome will occur, in this case the probability that the participant will produce the English native-like dominant name for a given object. Again, we estimated plots in which the individual variables (this time, language variables) interacted with three levels of immersion while holding all other variables constant at a mean value.
In a follow-up analysis, we asked how L1 and L2 norms might interact with one another in predicting a learner’s success in producing the L2 dominant name. Several interaction terms between these norming variables were highly significant, so we examined the cross-language relationships between L1 and L2 agreement and number of alternate L1 and L2 names. This analysis offers a closer examination of two interesting effects from the preceding results: (1) native speaker agreement in each language appears to compete in predicting L2 native-likeness and (2) an increasing number of names in English seems to be associated with greater L2 native-likeness.
This performance diminishes as English agreement decreases. In general, high levels of L2 (English) agreement are associated with successful learning across varying levels of L1 (Chinese) agreement. However, the worst performance by the learners occurs when Chinese agreement is high and English agreement is low, confirming that L1 patterns can have a strong negative effect on L2 native-likeness when L2 input is inconsistent. Many of these items came from the
For comparison,
Participants’ observed performance across the different numbers of alternative names in the English and Chinese norms, however, differed significantly from the regression model’s estimated accuracy rates.
In this study we examined the relative effects of four language history variables in predicting learners’ outcomes in L2 lexical categorization native-likeness. Highly significant interactions were found among these variables, supporting the idea that language history (e.g., age of L2 onset) variables should not be evaluated in isolation from other variables. Significant age of L2 onset effects were observed, but these effects were tempered by the positive contribution of increased immersion experience. A surprising observation was that increased experience with L2 prior to immersion was actually associated with reduced native-likeness of L2 lexical categorization. Finally, we found that for bilinguals with long-term L2 immersion, patterns of language use (i.e., code-switching habits) were a significant predictor of L2 native-likeness, but for learners with less immersion experience (including no immersion experience), language use was a less important predictor of L2 native-likeness.
We further explored how the naming norms of the linguistic communities of both languages influenced the learners’ success in acquiring native-like L2 lexical semantic mappings. Both L1 (Chinese) and L2 (English) norms were significant predictors of the learners’ L2 native-likeness, consistent with previous findings in other domains of second language acquisition, such as phonology. Further, we identified unique effects for agreement among native speakers and the number of alternate names produced in the norming samples. The result of an item-wise analysis revealed that a large amount of the between-object variation in naming was captured by these native speaker naming norms, indicating both the lasting impact of L1 mappings on L2 production and the sensitivity of L2 learners to the native speaker norms of the L2. Below we present a more detailed discussion of how L2 naming patterns are influenced by the learner variables and input (linguistic community norm) variables.
The most surprising finding of this study was that the number of years spent studying English outside an immersion environment was negatively related to L2 native-likeness in the lexical categorization task, even after controlling for the length of eventual L2 immersion. This outcome was not predicted by any past research nor intuition. This novel contrast between years of non-immersed and immersed learning in learners who have significant experience in both environments suggests that L2 training outside of an immersion environment may ultimately reinforce lexical semantic mappings that significantly differ from those of L2 native speakers. There is little doubt that immersion experience is beneficial to second language learning, and second language acquisition research has long promoted this view, but the present study adds the unique corollary that L2 learning without immersion may, in fact, hinder native-likeness. This effect may be due to the entrenchment of L1 structures in learners’ L2 as a result of impoverished input. Common classroom techniques for learning translation equivalents or naming highly prototypical objects encourage learners to export their inferences about object categories from L1 to L2 by way of one-to-one translation. However, native-like L2 mappings only become available to the learner with more diverse input from an immersion environment or (potentially) another immersive instructional setting such as the highly enriched virtual environments that may be simulated in computer games (see
Second language lexical learning has often been regarded as a qualitatively different type of acquisition from phonology and syntax that tend to show strong age effects. One theoretical account,
Age of second language onset effects may also be confounded with the negative pre-immersion learning effect. In the present study, we surveyed participants’ earliest exposure to English as a second language rather than their earliest immersion experience in an English language community. Although there were significant advantages for earlier learners over later learners, these advantages were limited in the sense that for every year of earlier acquisition, the same effects could be gained by an additional year of L2 immersion. With a small age of onset advantage on the one hand, and a non-immersed L2 learning disadvantage on the other hand, one may ask whether earlier L2 instruction is indeed beneficial for lexical semantic native-likeness. Addressing this question requires considering the multiple influence of both age effects, amount of total training, and the eventual onset of immersion (if at all). In a later section on implications for L2 instruction, we address these issues in further detail.
Whereas age effects and training effects focus specifically on the conditions under which learners begin acquiring a new language, eventual native-likeness may just as well depend on how that language is used at later stages, such as in an L2 immersion environment. Switching from one language to another may be common, even difficult to avoid, in bilingual environments, but considerably more variation in individual CSFreq could be observed among bilinguals in relatively monolingual-like environments. While some bilinguals may use each language in a distinct context (e.g., home vs. work), others may switch frequently. Research in first language lexical attrition has highlighted the role of bilinguals’ specific language use patterns in re-shaping L1 (
In the present study we observe a complementary effect. Increased code-switching is associated with greater L2 native-likeness. However, this effect interacts with L2 immersion such that it applies only after a significant period of immersion (illustrated at 4.7 years in
The causal relationship between CSFreq and native-likeness cannot be determined from our results, however. One explanation would argue that increasing an advanced learner’s code-switching leads to improvements in L2 native-likeness by promoting simultaneous activation and therefore increasing opportunities for lexical semantic remapping. On the other hand, bilinguals with greater L2 native-likeness may already be more involved in bilingual social settings (as opposed to seeking out L1 contexts) and increase their rate of code-switching as a result. Future research could investigate the short-term effects of code-switching in an experimental procedure, but the long-term causal relationship between these variables remains unknown.
Although the learner-oriented variables as discussed above proved useful in predicting overall performance in lexical categorization, they were rather inadequate in predicting native-like naming for individual objects. Language-specific variation, on the other hand, proved extremely important in predicting trial-by-trial accuracy of participants’ object naming, even after controlling for inter-participant differences in the learner variables. These effects have been revealed by our item-wise analyses. One lesson from these effects is that any kind of overall attainment score in lexical categorization masks significant variation in mastery for individual words, with some words posing much greater challenges for the learner (see also
We found a competing relationship between the level of native-speaker agreement in L1 and L2 in predicting the native-likeness of learners’ L2 responses. The role of L2 agreement in learners’ responses indicates that these learners are sensitive to variation in native speakers’ lexical categories for these objects. In the alternative case, where learners rely only on a general majority name for objects, we should see little effect of the L2 agreement variable, as learners would be more consistent than native speakers. Instead, learners respond proportionally to native speakers in their level of naming agreement. Further, the interaction between immersion and L2 agreement demonstrates that the advantage for high L2 agreement increased with greater immersion: These objects show greater improvement than low L2 agreement objects, which did not improve much even with almost 5 years of immersion.
Conversely, agreement among native-speakers of the L1 significantly impeded native-like naming in the L2, indicating that L1 learners were more resistant to revising their lexical semantic mappings in L2 when L1 native speakers were more consistent, likely showing a higher degree of confidence about the object’s category membership. The interaction between immersion and L1 agreement demonstrates that this L1 disadvantage predicts learners’ improvement with greater immersion experience. Low L1 agreement words significantly increase in their native-likeness with longer periods of immersion, while higher L1 agreement words show less improvement, highlighting these lexical semantic mappings’ resistance to restructuring.
Re-examining the observed accuracy rates for the learners across both L1 and L2 agreement levels, we found an antagonistic interaction between these variables. When L1 agreement was especially high, learners struggled to produce native-like L2 names, even at relatively high levels of L2 agreement. However, when L1 agreement was relatively low, L2 agreement was a better predictor of L2 learners’ native-likeness, apparently becoming more salient in the absence of strong L1 cues. The statistical model did not find such a strong interaction, instead identifying the same opposing main effects of L1 and L2 agreement but without an effect of the very small (though significant) interaction term. It remains to be seen whether the observed interaction is a byproduct of the objects in our particular task or whether the model simply underestimates the importance of this interaction. In either case, the important roles of L1 and L2 agreement norms are apparent, either independently or interactively.
The number of alternate names for an object produced by native speakers in L1 and L2 were also significant predictors of L2 learners’ native-likeness and were highly interactive with one another. If learners have only one name for an object in their native language, the model indicated that they would be equally likely to produce the dominant L2 name, regardless of alternatives. However, the observed naming behavior indicated that this trend overlooked a significant variation from L2 norms in the learners’ naming. For this subset of objects with only one name in L1, the lowest probability of producing a native-like L2 name occurred when the L2 provided two name alternatives, with greater L2 native-likeness occurring when only one L2 name was available or when three or more L2 names were available. This pattern suggests two mechanisms: (1) the attraction of the 1-to-1 translation, as learners struggle with competing pairs of L2 names, and (2) the competition within a distribution of L2 names, as learners’ performance improves with a greater number of name alternatives, showing some indifference to the L2 alternatives when there are several.
In the remaining conditions, when learners have multiple L1 words for an object, a greater number of L2 names appears to offer an advantage in selecting the dominant name. One potential explanation for this effect is the proportion of input that each alternative name comprises for L2 learners. Because the present model looks at both agreement in the dominant name and the number of alternatives, the latter provides an indirect measure of the native-speaker agreement levels for each alternate (non-dominant) name. As the number of alternative names increases, the remaining portion of the norm is divided into smaller parts relative to the dominant name, and thus each alternative name becomes a less salient competitor.
Under the foregoing explanation, we would expect the lowest L2 learner performance to occur when naming agreement is low
As the number of names in the sample of L1 native speakers increased, L2 learners’ native-likeness declined, suggesting that the relative frequency of the dominant name was less important for L1 than the full array of available names. This observation makes intuitive sense, as we would expect the learners to have a more stable, entrenched knowledge of their native language. In the case of L1, participants may simply be sensitive to the presence of names regardless of agreement level, or alternate names in L1 may reflect a more general uncertainty about the identity of an object and, apart from language, its membership in semantic categories with other objects. If the function of an entirely novel object is unknown, even native speakers will have a difficult time settling on the best name for that object because lexical categorization does not strictly adhere to similarity of physical features like size and shape.
Finally, performance on six objects that had two competing names in both L1 and L2 was observed to be the worst overall among the stimuli. This effect is not replicated in the modeled plots because, again, it depends not only on the number of names but on the combination of name agreement and number of names, while the model parametrically varies each of these factors. Indeed, the item-wise observations are consistent with the proposal that the distribution of naming agreement between the two objects in L2 drives the general disadvantage for two-named objects.
In the introduction, we explored how theories of lexical and semantic representation could be extended to understanding patterns of lexical categorization. The present study does not directly implement any specific theoretical model, as we observe only the naturally occurring shifts in lexical categorization by Chinese–English bilinguals over their varying language learning and language immersion experiences. Nonetheless, connectionist theories such as the Distributed Feature Model (
Specifically, important factors in connectionist training paradigms, such as amount, frequency, and consistency of input are readily translated into the lexical categorization terms used in this study. We quantify the amount of L2 experience (LOR), frequency of the dominant name relative to other names (naming agreement), and alternate names, finding compelling parallels between the associative learning principles that underlie connectionist models and the estimated effects of these variables on L2 categorization. For example, the (weak) age of onset effect observed in the present study concurs with entrenchment accounts of age effects in models of lexical acquisition (e.g.,
Entrenchment also provides some explanation for the relative disadvantage in re-mapping L2 categories for objects with high L1 agreement, as high agreement confers greater training frequency for the dominant L1 name (for a given object presentation). The role of L2 linguistic community variables in predicting learners’ native-likeness confirms that learners are sensitive to the relative frequency of several alternate names, showing improved performance when the agreement for the dominant name increases and decreased performance when alternate name competitors increase in frequency (e.g., two names distributed 60–40% versus three names distributed 60–20–20%). We also found support for the interactive relationship between L1 and L2 mappings, as suggested by the models proposed by
The present study offers several new insights into the role of language history, language training, and language use in second language lexical semantic learning. Most importantly, we find that greater time spent studying a second language before immersion predicted lower levels of eventual L2 native-likeness, likely due to the entrenchment of L1-like lexical-semantic mappings. Although we do find an age of onset effect, even after controlling for immersion and duration of language training, the magnitude of this age effect is proportional to the benefits of immersion, and the benefits to L2 native-likeness from early age of onset are small relative to the effects of more pre-immersion training.
On the extreme end, one might propose that pre-immersion language instruction is actually counter-productive to native-like lexical semantic development, and second language education would be best postponed until immersion opportunities arise. However, this viewpoint is impractical for most non-immigrant learners, and likely over-stated, as our analysis of language-specific variables (native-speaker agreement and alternative names) show that learners are, in fact, highly sensitive to the inconsistent input that describes native-like lexical categorization. Lexical semantic learning in non-immersion environments might therefore be improved by introducing learners to a greater variety of referents and the naturally diverse naming patterns associated with those referents, allowing them to develop more native-like intuitions about the relationships between objects that define lexical categories. The method of using a diverse set of naming patterns in second language instruction clearly contradicts the traditional classroom teaching method, in which training focuses primarily on one-to-one translations; such a focus underestimates cross-languages differences, and by our findings, encourages the use L1 patterns for L2 words and therefore impedes learners’ later ability to acquire native-like lexical semantic mappings.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We wish to thank Anqi Li for supervising data collection and coding, as well as Peiyao Chen, Patrick Clark, Anqi Li, Jessica Wen, Han Wu, Zhichao Xia, Tianyang Zhang, Dan Zhong, and Lijuan Zou for assistance with participant recruitment, data collection, and language consulting. We also thank Hua Shu at Beijing Normal University for providing lab facilities and equipment for this study. This research was supported by National Science Foundation grants (BCS-1057855; OISE-0968369).
Two participants in the US and four in China were excluded for response rates below 50% on one or more of the naming tasks. Non-response rates were approximately the same between the English task (four participants in Beijing) and the Chinese task (three participants in Beijing and two in State College), suggesting that most of these missing data were attributable to participant inattention. Two participants in China were removed for naming accuracy scores more than 2.5 standard deviations below the mean (see Subject-wise Analysis).
In two cases, the naming agreement score for an object was tied between two names. For each case, we randomly selected one name as the “dominant” name for the purpose of the comparison. This uncertainty, however, is preserved in the L2 Name Agreement variable included in the logistic regression.