Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Bilinguals exhibit semantic convergence while remaining communicatively efficient
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Maya Taliaferro1, Nathaniel Imel1, Noga Zaslasky1, Esti Blanco-Elorrieta1; 1New York University
Semantic categories vary across languages, such that translationally equivalent words do not always refer to the same things (Berlin & Kay, 1969; Van Hell & De Groot, 1998; Malt, 1995). This variation is thought to be constrained by pressures for efficient communication (Kemp, Xu, & Regier, 2018), and more specifically, has been characterized by the Information Bottleneck (IB) tradeoff between communicative accuracy and informational complexity (Zaslavsky, Kemp, Regier, & Tishby, 2018). This raises two key questions: (1) how are distinct category systems managed when two languages coexist within a single mind, as in the case of bilinguals, and (2) is this type of integration of category knowledge shaped by pressure for efficiency? Prior work suggests that bilingual category systems may converge (Ameel et al., 2005, 2009; White et al., 2017), thus deviating for the monolingual systems. However, the precise structure of category boundaries and whether bilinguals are also constrained by efficiency remains largely unknown. Here, we address these questions by conducting four theory-driven experiments to test how bilinguals integrate category knowledge, instantiate a cognitively grounded IB model, and evaluate its predictions in monolingual and bilingual English and Mandarin speakers. We used images of household objects arranged along six continua between distinct categories (e.g., plate vs. bowl). For each continuum, participants completed two tasks designed to obtain each group’s communicative behavior: a free naming task and a comprehension task. Bilingual participants completed the tasks in both English and Mandarin on separate days, with the order of languages counterbalanced. Participants also completed a similarity rating task and a prior elicitation task, and the data from these tasks was used to instantiate two IB models: a shared model, that generates predictions for naming and comprehension behavior based on a shared similarity space and prior across language groups; and a group-specific model that considers group-specific similarities and priors for generating these predictions. Crucially, the models are not trained on the naming and comprehension data they aim to predict. We find convergence in bilingual naming patterns, with highly correlated similarity structures and priors across groups, suggesting broadly shared underlying non-linguistic representations. Across all language groups and continua, category systems are highly efficient under the shared IB model, and significantly more efficient compared to hypothetical counterparts. We found no significant advantage for the group-specific IB model, further supporting a common representational space across groups. Furthermore, the shared IB model largely captures the naming and comprehension patterns for both bilinguals and monolinguals, suggesting that pressure for efficiency shapes not only monolingual category systems but also converged bilingual systems. These findings provide substantial evidence for bilingual convergence in category systems and shed new insight as to the cognitive mechanisms that underlie how bilinguals integrate two non-overlapping category systems. Pressure for efficiency emerges as a key organizing constraint on the structure of bilingual category systems, while bilinguals appear to maintain similar non-linguistic representations as monolinguals. An important direction for future work is to examine how pressure for efficiency interacts with social and environmental factors in shaping bilingual category systems.
Topic Areas: Multilingualism, Computational Approaches