Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A Computational Approach to How Bilinguals Reconcile Cross-Linguistic Categorization
Poster Session D, Saturday, September 13, 5:00 - 6:30 pm, Field House
Aditi Singh1, Maya Taliaferro1, Esti Blanco-Elorrieta1; 1New York University
To categorize the world around us, we use labels that aid memory, recognition, and generalization. While some concepts have clear boundaries, others are more ambiguous, resulting in different category boundaries across languages—even for direct cross-linguistic translations. Bilinguals need to reconcile cross-linguistic differences in categorization, but how this process is carried out remains unknown. In this study, we investigated whether bilingual individuals have separate conceptual systems for each language—categorizing objects similarly to monolingual speakers of each language—or whether they develop a single, unified conceptual system whereby concepts are categorized along the same boundaries regardless of the language being used. 24 English monolinguals, 24 Mandarin monolinguals, and 25 Mandarin-English bilinguals performed a two-alternative forced choice (2AFC) task and were asked to categorize objects that gradually transitioned from one concept (e.g., bowl) to another (e.g., plate). Our results showed that even for concepts that had different monolingual boundaries, bilinguals categorized objects identically across languages and did not map their responses to each language’s categorization rule. These findings confirmed that although categorization differs across languages for the two monolingual groups, it remains unified for bilinguals irrespective of the language used. Interestingly, the unified bilingual boundary did not reflect an even bidirectional influence of both languages. Theoretically, this could be due to several sociolinguistic factors, including Age of Acquisition, Proficiency, and Exposure in each language. Given the multilayered nature of the bilingual experience, these parameters are hard to control for (or parametrically vary) in longitudinal developmental studies. For this reason, we created a cognitively plausible deep-learning neural network model that was trained to perform the same 2AFC task and that could be used to characterize the influence of Language Exposure in a bilingual’s conceptual space. Specifically, this neural network model was based on feature-based cognitive theories of categorization—which break down objects into their defining features—and learned to take in an image and predict its label. To simulate a bilingual with greater exposure to one language over the other we manipulated the ratio of Mandarin to English training examples across three conditions: Mandarin-dominant (90%–10%), English-dominant (10%–90%), and balanced (50%–50%). Our model successfully replicated the behavioral results by converging on a single, blended category boundary as the optimal solution when guided by cognitively plausible constraints. Crucially, balanced language exposure led to a bilingual category boundary that lay exactly midway between the monolingual boundaries while unbalanced exposure biased the categorization boundary toward the dominant language. This demonstrated that although bilinguals rely on a single conceptual boundary across languages, where this boundary lies is causally shaped by Exposure. Taken together, our behavioral and computational results demonstrate that bilinguals flexibly navigate cross-linguistic category conflicts by developing a shared, yet dynamic, conceptual space across languages that is systematically influenced by exposure to each language.
Topic Areas: Multilingualism, Computational Approaches