Slide Slam L13
Words in the CLOUD: How orthographic similarity and bilingual experience facilitate foreign vocabulary learning
Jose A. Aguasvivas1,2, Alberto Testolin3,4, Marco Zorzi3, Manuel Carreiras1,2,5; 1BCBL, Basque Center on Cognition, Brain and Language, 2Universidad del País Vasco / Euskal Herriko Unibersitatea (UPV/EHU), 3Department of General Psychology, University of Padova, 4Department of Information Engineering, University of Padova, 5Ikerbasque, Basque Foundation for Science
Learning a foreign language as an adult is a rewarding but challenging endeavor that entails accruing a massive vocabulary to achieve adequate proficiency. The literature highlights that orthographic similarity and bilingual experience independently facilitate foreign vocabulary acquisition. Intuitively, a word that looks more similar to a known language should be easier to learn regardless of its meaning; and the knowledge of two languages—bilingualism—could provide an advantage in terms of sources from where to draw similarities. However, despite numerous efforts to formalize the development of the mental lexicon with computational approaches, the basis for these vocabulary learning effects remains largely unexplored. Here, we explored the combined effects of orthographic similarity and bilingual experience on foreign vocabulary learning using behavioral and computational approaches. First, we compared Spanish monolingual, Spanish-English, and Spanish-Basque bilingual participants (n = 40 per group) when learning an artificial vocabulary with varying orthographic similarity to Spanish. The vocabulary contained 24 orthographically similar and 24 dissimilar novel words, paired with black and white depictions of real objects. Participants performed a familiarization and five active learning blocks where they recognized and produced the novel vocabulary. Growth curve analyses on these data revealed that both bilingual groups outperformed the monolingual group, better recognizing and producing the novel words across the blocks regardless of their similarity to Spanish. Additionally, as expected, orthographically similar words were easier to recognize and produce than dissimilar words. As a second step, we developed the CLOUD—Constrained Learner of Orthography: Unified, Distributed, and Dynamic—model, a character-level recurrent neural network that can learn written vocabulary by implementing a unified, distributed, and dynamic view of the orthographic lexicon. We simulated adults’ orthographic lexicons by pre-training this architecture on monolingual and bilingual input using around 17,000 words. The monolingual and bilingual versions’ accuracy was matched using an adapted lexical decision task. After pre-training and matching the models, we tested the monolingual and bilingual models’ capacity to learn the novel words used in our behavioral task. Crucially, the models could emulate the orthographic similarity effects and showed an overall advantage of experience with bilingual input, as observed in the behavioral results. Past research has highlighted orthographic similarity and bilingual experience as independent catalysts of foreign vocabulary learning. The present study is first in unifying these seemingly disparate findings under a common computational framework, whereby distributed representations of orthographic word forms are stored in a unified space and dynamically modified by learning experiences. Our model could simulate participants’ behavior, corroborating the influence of orthographic similarity and showing a bilingual advantage for receptive and productive vocabulary. This conceptualization has implications regarding how consistent experience with specific words in different linguistic contexts (i.e., bilingual settings) can influence foreign vocabulary acquisition. Our work opens up exciting pathways for further investigating the cognitive and computational mechanisms of foreign vocabulary learning.