Inhibitory and facilitative effects of lexical neighbors in spoken word recognition: The role of language experience

Abstract

The study evaluated whether the direction (inhibitory or facilitative) of the phonological neighborhood density effect in English spoken word recognition was modulated by the relative strength of competitor activation (neighborhood type) in two groups of English-dominant learners of Spanish who differed in language experience. Classroom learners and heritage learners of Spanish identified spoken English words from dense (e.g., BEAR) and sparse (e.g., BOAT) phonological neighborhoods presented in moderate noise. The phonological neighborhood was separately manipulated at word onset (cohort) and word offset (rhyme). Classroom learners were overall slower in recognizing spoken words from denser neighborhoods. Strongly active (onset) neighbors exerted inhibitory effects in both classroom and heritage learners. Critically, weakly active (offset) neighbors exerted inhibitory effects in classroom learners but facilitative effects in heritage learners. The results suggest that the activation of both within and cross-language neighbors should be considered in determining the direction of neighbor effects in bilingual lexical processing.

Introduction

Achieving proficiency in a second language can be a matter of choice or circumstance. For those who choose to learn a second language (L2), the process typically involves classroom-based instruction with or without L2 immersion opportunities. In contrast, for those who learn an L2 as a result of circumstance, the process typically involves naturalistic exposure to a heritage language in the home. The resulting language learning experience differs on many other aspects, including the timing of L2 exposure (late in classroom learners, early in heritage learners) and input modality (spoken and written input for classroom learners, spoken input for heritage learners; for a review, see Montrul, 2012). Both groups of learners maintain dominance of the majority language and achieve varying degrees of L2 proficiency. Functionally, the majority of classroom learners remains monolingual, whereas heritage learners become bilingual.

Although much research has evaluated how such differences in language learning experience may affect L2 outcomes (Dewey, 2004; Hartshorne et al., 2018), there has been considerably less work on how the distinct paths to L2 proficiency may affect lexical processing in the dominant language (Botezatu et al., 2022b; Linck et al., 2009). In the current study, we considered whether the distinct language experience of classroom and heritage learners of Spanish affects the relative strength of within and cross-language lexical neighbor activation, influencing the direction (inhibitory or facilitative) of the effect of lexical neighbors during spoken word recognition in the dominant language (i.e., English).

Spoken word recognition involves mapping of the unfolding speech signal onto the lexicon (for a review, see Mirman, 2016). Interactive activation models such as the TRACE model (McClelland & Elman, 1986) propose that this mapping occurs incrementally, as listeners temporarily activate a broad network of similar-sounding words in their lexicon that partially match the input, resulting in competition among multiple lexical candidates. The competition represents an index of the relative difficulty of mapping the speech signal onto the lexicon and may be operationalized in terms of phonological neighborhood density (Luce, 1986), a measure of the number and average frequency of phonologically similar lexical candidates that partially match the unfolding speech (Luce & Pisoni, 1998).

Phonological neighbors have generally been reported to exert inhibitory effects on spoken word recognition (e.g., Dufour & Frauenfelder, 2010; Luce & Pisoni, 1998; Magnuson et al., 2007), but facilitative effects have also been reported (Vitevitch & Rodríguez, 2005). Inhibitory effects have been related to the activation of a larger number of phonologically similar neighbors during recognition of spoken words from denser than sparser neighborhoods, resulting in slower and less accurate recognition of words from denser neighborhoods. Phonological neighbors may also exert an indirect facilitative effect in spoken word recognition through excitatory feedback connections that support shared phonological units. However, this facilitative effect is generally weaker than the direct inhibitory effect. Phonological neighbors have also been shown to exert facilitative effects in visual word recognition (see Andrews, 1997, for a review) and spoken word production (Kittredge et al., 2008; Vitevitch & Sommers, 2003), suggesting that the inhibitory or facilitative direction of the effect may differ across tasks. In an attempt to provide a unified account of the inhibitory and facilitative effects of lexical neighbors, Chen and Mirman (2012) used a simple interactive activation and competition model to simulate phonological, orthographic, and semantic neighbor effects in word recognition and production tasks. The authors used the same model and parameter values across simulations to discover the computational principle that determined whether neighbor effects were facilitative or inhibitory across domains. For the spoken domain, the architecture of the interactive activation and competition model used to simulate phonological neighbor effects in spoken word recognition tasks was closely related to the TRACE model. The results revealed that the overall strength of neighbor activation determined the direction (inhibitory or facilitative) of neighbor effects across simulations. Specifically, Chen and Mirman (2012) found that the inhibitory effect outweighed the facilitative effect for strongly active neighbors, whereas the facilitative effect outweighed the inhibitory effect for weakly active neighbors. In the current study, we seek to determine whether this computational principle (i.e., the strength of neighbor activation determines the inhibitory or facilitative direction of neighbor effects) is supported in the spoken domain by performance in two groups of English-dominant learners of Spanish who differ in the strength of within and cross-language competitor activation as a function of language experience.

The current study

The study evaluated whether the direction (inhibitory or facilitative) of the phonological neighborhood density effect in English spoken word recognition was modulated by the relative strength of competitor activation (i.e., neighborhood type) in two groups of English-dominant learners of Spanish who differed in language experience: classroom learners of Spanish who were exposed to Spanish in an educational setting after the age of 10 and were functionally English monolingual and heritage learners of Spanish who were exposed to Spanish in the home environment prior to the age of 5 and were functionally bilingual. We chose to test classroom and heritage learners of Spanish for this study because both groups activate a broad set of within-language competitors during spoken word recognition in the dominant language (Marian & Spivey, 2003a, 2003b) and may experience at least occasional coactivation of their nondominant language (Blumenfeld & Marian, 2007; Canseco-Gonzalez et al., 2010; Ju & Luce, 2004). Critically, the two groups differ in age of Spanish acquisition and Spanish proficiency, which have been shown to influence the number of cross-language competitors and strength of their activation (Blumenfeld & Marian, 2013; Canseco-Gonzalez et al., 2010). Bilinguals who acquire their L2 relatively early (i.e., before the age of 6) experience cross-language competition during spoken word recognition (Canseco-Gonzalez et al., 2010). This group comparison is critical in determining whether the direction of the phonological neighborhood density effect in spoken word recognition is only dependent on the strength of competitor activation or whether the relationship is mediated by participants’ language experience.

The spoken domain provides a unique opportunity to test this question because the activation of phonological neighbors that partially match the unfolding speech differs in timing and strength as a result of the incremental mapping of the speech signal onto the lexicon. Simulations with the TRACE model (McClelland & Elman, 1986) and empirical evidence from eye-tracking studies using the visual-world paradigm (Allopenna et al., 1998; Magnuson et al., 2007) suggest that the overall amount of activation and competition for cohort neighbors, which share the onset of the spoken word (operationalized as the first two phonemes), is much higher than for rhyme neighbors, which share the offset of the spoken word (operationalized as the vowel and coda). Early in the time course of spoken word recognition, cohorts are strongly active and compete more strongly for selection than rhyme neighbors. Late in the time course of spoken word recognition, rhymes might be more active than cohorts, but both are weakly active at that point. Most research on bilingual spoken word recognition has focused on competition dynamics among cohort neighbors, revealing robust within- (Bruggeman & Cutler, 2019; Canseco-Gonzalez et al., 2010; Marian & Spivey, 2003b; Shin et al., 2015) and across-language effects (Blumenfeld & Marian, 2007; Canseco-Gonzalez et al., 2010; Marian & Spivey, 2003a, 2003b). Competition among offset neighbors has remained largely unexplored in bilinguals. The available evidence suggests that within-language rhyme competition in bilinguals may be either similar to that experienced by monolinguals (Shin et al., 2015) or very weak (Bruggeman & Cutler, 2019). Therefore, it is critical to consider whether differences in language experience may affect rhyme competition.

We hypothesize that if language experience modulates the overall strength of neighbor activation, then both monolinguals and bilinguals should exhibit an inhibitory effect for within-language neighbors that compete strongly for selection (i.e., cohort neighbors) but that the direction (inhibitory or facilitative) of the effect exhibited by within-language neighbors that exert weak competition (i.e., rhyme neighbors) may differ between groups as a function of degree of interference from cross-language competitors. If the overall strength of neighbor activation determines the direction of neighbor effects in lexical processing, as proposed by Chen and Mirman (2012), then regardless of participants’ language experience, cohort neighbors should exert an inhibitory effect on spoken word recognition because they compete strongly for selection, whereas rhyme neighbors should exert a facilitative effect on spoken word recognition because they represent weaker competitors. Alternatively, if the strength of neighbor activation does not determine the direction of neighbor effect in lexical processing, then both cohort and rhyme neighbors should exert inhibitory effects on spoken word recognition, with the magnitude of the effect being larger for cohort than for rhyme neighbors. The results are expected to hold beyond individual differences in language proficiency, cognitive control, and working memory.

Method

Participants

Forty-five English-dominant classroom learners of Spanish (15 male, mean age = 20.3, age range = 18–27 years) and 37 English-dominant heritage learners of Spanish (8 male, mean age = 19.3, age range = 18–24 years) were recruited from the University of Missouri and the University of California, Riverside and completed the study for payment. Classroom learners lived in either unilingual (Midwest; n = 29) or diverse (West Coast; n = 16) linguistic environments on the United States and learned Spanish through classroom instruction after age 6. Heritage learners lived in a linguistically diverse environment on the U.S. West Coast, where they grew up in Spanish-speaking households and were exposed to English before the age of 6, being schooled almost exclusively in English. All participants reported normal hearing and vision and no history of neurological, language, reading, or learning disorders. Participants varied on lab-based measures of language proficiency (see Table 1) and cognitive resources (see Table 2). Both groups exhibited higher proficiency in English than Spanish and reported a strong preference for speaking and reading English over Spanish.

Table 1. Mean (standard error) psycholinguistic data

Table 2. Mean (standard error) for demographic and cognitive measures

Materials and procedure

Participants completed a set of tasks in E-Prime 2.0 (Psychology Software Tools Incorporated, 2012) that measured English spoken word recognition, language production proficiency, working memory, and cognitive control. A brief description of each task is provided below. For full details of the experimental procedures, see Botezatu, Guo, et al. (2022).

English spoken word recognition

Spoken word recognition was measured using a spoken-to-written word-matching task. Phonological neighborhood density effects have been observed using this task in a variety of populations, such as monolinguals (Botezatu et al., 2015; Botezatu & Mirman, 2019), bilinguals (Botezatu et al., 2022a, 2022b) and individuals with poststroke aphasia (Botezatu & Mirman, 2019), although the effect has not been observed in other close-set formats (Sommers et al., 1997). Spoken words were 120 monosyllabic English words that varied systematically in phonological neighborhood density, which was separately manipulated at word onset (i.e., first two phonemes, cohort neighborhood) and word offset (i.e., vowel and coda, rhyme neighborhood). In the cohort condition, density was operationalized as the number, t(58) = 4.4, p < .001, and summed frequency of onset neighbors, t(58) = 4.23, p < .001, while the equivalent measures of the offset neighborhood were held constant (all ps > .05). In the rhyme condition, density was operationalized as the number, t(58) = 10.47, p < .001, and summed frequency of offset neighbors, t(58) = 6.86, p < .001, while the equivalent measures of the onset neighborhood were held constant (all ps > .05). The manipulations of neighborhood density (dense versus sparse) and neighborhood type (cohort versus rhyme) resulted in four conditions of interest: dense cohorts (e.g., BEAR), sparse cohorts (e.g., BOAT), dense rhymes (e.g., BEAN), and sparse rhymes (e.g., BENCH). There were 30 trials in each condition. The conditions were matched on number of phonemes, lexical frequency, and auditory file length (all ps > .05). The conditions were further matched on the number and summed frequency of Spanish orthographic neighbors and on the number of Spanish phonological neighbors (all ps > 0.05; see Table A1 in the Appendix). The analysis of the summed frequency of Spanish phonological neighbors yielded a significant interaction between neighborhood density and neighborhood type, F(1, 26) = 5.09, p = .033. Follow-up analyses revealed a trend toward a significant difference in the summed frequency of Spanish phonological neighbors at the level of cohorts, t(59) = 2.26, p = .06, such that on average, dense cohorts tended to have higher summed frequency of Spanish phonological neighbors than sparse cohorts. This difference was driven by a few extreme values, as most items had no Spanish neighbors. In contrast, no difference on the summed frequency of Spanish phonological neighbors was detected at the level of rhymes, t(59) = -0.77, p = .46. The list of experimental items is found in Table A2 in the Appendix.

On each trial, participants listened to one of 120 monosyllabic English words presented at a signal-to-noise ratio of -2dB over headphones and used a mouse to select the word they heard from among an array of response options presented on the screen after a 1,000-ms delay. Each array of response options included the target word, two close distractors (i.e., a cohort neighbor and a rhyme neighbor), two distant distractors (i.e., an onset neighbor of the rhyme neighbor and an offset neighbor of the cohort neighbor), and an undecided response option “?” that were presented at a random location on each trial. Response latencies and accuracy rates were collected on each trial.

Proficiency

Proficiency was assessed using measures of verbal fluency (semantic and phonemic), expressive vocabulary, and word/nonword reading, as well as self-ratings. Semantic fluency measured the average number of correct responses (excluding repetitions) produced across four named semantic categories (i.e., animals, clothing, fruit, furniture), whereas phonemic fluency provided an equivalent measure in response to three named letters (i.e., “F,” “A,” and “S” for English; “P,” “M,” and “R” for Spanish). The Multilingual Naming Test (MiNT; Gollan et al., 2011) provided a measure of expressive vocabulary, indexed as the average correct-response latencies and accuracy rates in naming 68 line drawings of increasing difficulty. Speeded word reading was assessed using the Test of Word Reading Efficiency in English (TOWRE; Torgesen et al., 1999) and a Spanish equivalent (Miller et al., 2006), each measuring the number of words of increasing difficulty (out of 104) read aloud correctly within 45 s. Phonemic decoding was tested in English only using the Phonemic Decoding subtest of the TOWRE (Torgesen et al., 1999), which measured the number of nonwords of increasing difficulty (out of 63) read correctly within 45 s. Participants self-reported their language background and proficiency using the Language Experience and Proficiency Questionnaire (Marian et al., 2007). Proficiency was computed for each language by adding the z scores of the semantic fluency, phonemic fluency, MiNT accuracy, sight word reading, and phonemic decoding (English only) measures in the relevant language.

Working memory

The operation span task (Turner & Engle, 1989) was administered in English, the dominant language, to measure verbal working memory. On each of the 60 equation–word pairs presented in sets of two to six, participants made a button-press judgment of correctness in response to an arithmetic equation (e.g., 8 / 4 + 3 = 1) and then stored in memory a visually presented English word (e.g., FATHER) for recall at the end of a set. The working memory score reflected the number of correctly recalled words from trials with correct equation judgements. The only restriction on recall was that participants could not list the last word of a set in the initial position unless that was the only item from that list that they recalled.

Cognitive control

The Simon task (Simon & Rudell, 1967) was used to measure cognitive control. On each of the 126 trials, participants pressed one of two color-coded keys on the standard keyboard/Chronos Box in response to the color (not the location) of a red or blue square that appeared on the screen in three locations: in the middle of the screen (neutral), on the same side as the correct-response key (congruent), and on the opposite side of the correct-response key (incongruent). Cognitive control was operationalized as the difference in response latency between incongruent and congruent trials after excluding incorrect responses, absolute outliers (response latencies faster than 200 ms and slower than 2000 ms), and relative outliers (average response latency ±3 standard deviations).

Data analysis

The latencies of correct-response spoken word recognition were analyzed using linear mixed-effects models (Baayen et al., 2008; Barr et al., 2013). Response accuracy was analyzed using mixed-effects logistic regression (Baayen et al., 2008; Barr et al., 2013; Jaeger, 2008). Analyses were implemented in R version 3.6.0 (R Development Core Team, 2016; http://cran.us.r-project.org/) using the lme4 package version 1.1-11 (Bates et al., 2016).

For each analysis, the base model included the fixed effect of neighborhood density (dense versus sparse) and a maximal random-effect structure consisting of random effects of participants and items and by-participant random slopes of density. Neighborhood type (cohort versus rhyme) was entered in the analysis as a fixed effect and was evaluated in terms of its overall effect on response latencies and accuracy rates, then in terms of its interaction with density (dense versus sparse) and group (classroom versus heritage speakers). Improvement in model fit for each of these steps was evaluated using the likelihood ratio test (χ² test with degrees of freedom equal to the number of parameters added). Models controlled for individual differences in English and Spanish proficiency, working memory, and cognitive control. To ease the interpretation of the models, control measures were centered before being entered in the analysis. Parameter estimates for the full models are presented in Table 3. Simplified models that do not account for individual difference variables are reported in Table A3 in the Appendix.

Table 3. Predictor models of spoken word recognition latency and accuracy

Note. Full model formula: RT (or accuracy) ~ English proficiency + Spanish proficiency + Cognitive control + Working memory + Density*Neighborhood type*Group + (Density | participant) + (1 | item).

Results

First, incorrect-response trials (representing 22.3% of the data) were excluded from the response latency data but not from the response accuracy data. Then, relative outliers (trials with average response latency ± 3 standard deviations from participants and items means, representing 2.2% of the data) were excluded. Overall, participants were equally fast, b = 37.59, SE = 29.58, 95% CI [-20.39, 95.57], t = 1.27, p = .204, and accurate b = -0.12, SE = 0.17, 95% CI [-0.46, 0.22], z = -0.70, p = .485, in recognizing spoken English words from dense and sparse phonological neighborhoods. The effect of density on response latencies was modulated by group, b = -20.13, SE = 6.84, 95% CI [-33.52, -6.73], t = -2.94, p = .003. The standard inhibitory effect of phonological neighborhood density in spoken word recognition was replicated in classroom learners of Spanish but not heritage learners of Spanish.

There was also a trend toward a significant interaction between density and neighborhood type on response latencies, b = 52.99, SE = 29.58, 95% CI [-4.98, 110.97], t = 1.79, p = .073, which revealed that the density effect was larger in cohort than in rhyme neighborhoods. There was also an interaction between neighborhood type and group on response accuracy, χ²(3) = 9.70, p = .021, b = -0.09, SE = 0.03, 95% CI [-0.15, -0.03], t = -2.99, p = .003, such that heritage learners were more accurate in responding to rhymes, whereas classroom learners were equally accurate in both conditions. Critically, there was a three-way interaction among density, neighborhood type, and group on response latencies, χ²(3) = 13.11, p = .004, b = 14.14, SE = 6.83, 95% CI [0.76, 27.52], t = 2.07, p = .038, which revealed that dense cohort neighborhoods were associated with inhibitory effects in both groups. In contrast, dense rhyme neighborhoods yielded inhibitory effects in classroom learners but facilitative effects in heritage learners (see Figure 1, Panel A).

Figure 1. Spoken word recognition latencies (Panel A) and accuracy rates (Panel B) to English words from dense versus sparse cohort and rhyme neighborhoods in heritage and classroom learners of Spanish. Error bars indicate standard errors.

Discussion

The study evaluated whether the overall strength of neighbor activation determines the direction of phonological neighborhood density effects in spoken word recognition by experimentally manipulating the phonological neighborhood of target words at word onset and word offset (in separate conditions). Two groups of speakers who differed in the strength of within and cross-language competitor activation as a function of language experience were tested. Strongly active (i.e., cohort) neighbors exerted inhibitory effects in both classroom and heritage learners, whereas weakly active (i.e., rhyme) neighbors exerted inhibitory effects in classroom learners but facilitative effects in heritage learners.

The results support the hypothesis that language experience modulates the overall strength of neighbor activation. Strongly active (i.e., cohort) competitors (Allopenna et al., 1998) have been shown to exert robust inhibitory effects in English spoken word recognition (Magnuson et al., 2007). The current study replicates this pattern with monolinguals and bilinguals, suggesting that both groups experience competition from onset neighbors, leading to robust inhibitory effects of dense cohort neighborhoods. Rhyme neighbors have been shown to be weakly active competitors in monolinguals (Allopenna et al., 1998). Conflicting evidence from research on bilingual spoken word recognition points to either robust rhyme competition in the L2 (Shin et al., 2015) or no competition in either the L1 or the L2 (Bruggeman & Cutler, 2019). In the current study, we found a trend toward a significant interaction between neighborhood density and neighborhood type, suggesting that, overall, participants exhibited smaller density effects for rhymes than cohorts. This suggests that overall participants experienced weaker activation of rhyme competitors, consistent with previous reports from monolinguals (Allopenna et al., 1998) and bilinguals (Bruggeman & Cutler, 2019).

Consideration of group differences in language experience suggests that the overall weaker activation of rhyme competitors results in inhibitory effects of dense rhyme neighborhoods in classroom learners, who were functionally monolingual speakers of English, but in facilitative effects of dense rhyme neighborhoods in heritage learners, who were functionally bilingual. The reversal of the density effect in weakly active competitors (i.e., rhymes) experienced by heritage learners may be interpreted in the context of a dynamic language system, where weakly active offset neighbors may become more open to the influence of neighbors from the nontarget language. Previous studies reported that bilinguals who acquire an L2 at a younger age, as is the case for heritage learners in this study, experience cross-language competition during spoken word recognition (Canseco-Gonzalez et al., 2010). Heritage learners also had higher Spanish proficiency than classroom learners in this study (all ps < .05), resulting in higher coactivation of Spanish competitors (Blumenfeld & Marian, 2013). The current results suggest that in the context of weakly active within-language neighbors, the direction (inhibitory or facilitative) of the phonological neighborhood density effect may be influenced by the degree of cross-language competition.

The presence of noise in the speech signal might have influenced the overall activation of rhyme competitors in the two participant groups. The addition of noise in spoken word recognition has been shown to boost the activation of word offset neighbors in monolingual listeners (Brouwer & Bradlow, 2011, 2016; McQueen & Huettig, 2012), which may explain why classroom learners, who were functionally monolingual, exhibited an inhibitory effect of phonological neighbors in both cohort and rhyme conditions. In contrast, bilingual listeners have been reported to strongly rely on word-initial information during spoken word recognition in noise (Coumans et al., 2014; Hintz et al., 2021), which may have contributed to the weak activation of rhyme competitors in the group of heritage learners, who were functionally bilingual. Additionally, noise has been shown to increase cross-language lexical competition during spoken word recognition (Guediche et al., 2023).

Limitations and future directions

The current study suggests that differences in language experience may mediate the relationship between the strength of neighbor activation and the direction of neighbor effects in lexical processing, pointing to the dynamic nature of the bilingual lexicon. We consider the limitations of the current study. First, the spoken-to-written word-matching task used a close-set response format, which makes it more difficult to observe neighborhood density effects (Sommers et al., 1997). Although we had successfully observed phonological neighborhood density effects using this task in a variety of participant populations (Botezatu et al., 2015; 2022a, 2022b; Botezatu & Mirman, 2019), future studies should test whether the results are replicated using open-set response formats. Second, the current study did not fully control of cross-language neighbors (i.e., the dense cohort condition had higher summed frequency of Spanish phonological neighbors than the sparse cohort condition due to a few extreme values). Future studies should evaluate whether results are replicated in conditions that manipulate within-language competitors while fully controlling for cross-language neighbors as well as in conditions in which within- and cross-language competitors are covaried. Third, as noted in the previous paragraph, the presence of noise in the speech signal might have influenced the overall activation of rhyme competitors. Future research should also test whether results are replicated when stimuli are not presented in noise and whether the pattern persists when evaluating the time course of spoken word recognition using eye-tracking, computer-mouse tracking, or electrophysiological methods, which provide a higher temporal resolution than behavioral methods. A fourth limitation is the relatively small sample of participants in each group. Results should be replicated with larger participant samples.