Bilingual Lexical Representation in a Self-Organizing Neural Network Model

Bilingual Lexical Representation in a Self-Organizing Neural Network Model Xiaowei Zhao (xzhao2@richmond.edu) Ping Li (pli@richmond.edu) Department of Psychology, University of Richmond Richmond, VA 23173 USA Abstract In this paper we present a self-organizing neural network model of bilingual lexical development. We focus on how the representational structure of the bilingual lexicon can emerge, develop, and change as a function of the learning history. Our results show that (1) distinct representations for the two lexicons can develop in our network during simultaneous acquisition, (2) the representational structure is highly dependent on the onset time of L2 learning if the two languages are learned sequentially, and (3) L2 representation becomes parasitic on L1 representation when L2 learning occurs late. The results suggest a dynamic developmental picture for bilingual lexical acquisition: the acquisition of two languages entails strong competition in a highly interactive context and limited plasticity as a function of the timing of learning. Keywords: SOM; DevLex; Bilingual Lexicon. Introduction Mechanisms underlying early bilingual lexical acquisition are so far poorly understood. This lack of knowledge may be partly due to the methodological limitations associated with studying young bilingual children at early stages of language development (e.g., Bialystok, 2001). Work in the monolingual context has shown that neural network models are ideally suited for identifying mechanisms of early lexical acquisition (e.g., Li, Farkas & MacWhinney, 2004; Regier, 2005). Unfortunately, the gap between neural networks and bilingualism is still wide open: to date, there have been only a handful of neural network models that are designed specifically to account for bilingual language processing and representation (see reviews in Li & Farkas, 2002; French & Jacquet, 2004; Thomas & van Heuven, 2005). Furthermore, no neural network model has been devoted to capture the impact of developmental time on bilingual children’s lexical representations. Our study here attempts to bridge the gap by examining bilingual lexical representations with a self- organizing neural network. An issue of enduring interest in bilingualism has been whether bilingual representation takes the form of a single, shared lexical storage or a separate, distinct storage in the mental lexicon (see French & Jacquet, 2004 and Kroll & Tokowicz, 2005 for recent reviews). The issue has been highly controversial, and has recently been further complicated by conflicting neuroimaging data (see Hernandez & Li, 2007), but researchers have come to recognize that a host of variables must be taken into consideration in dealing with this issue, such as bilingual proficiency, learning history (including age of acquisition), modality (comprehension vs. production), and word types (cognates vs. noncognates, abstract vs. concrete words). The DevLex and DevLex-II models have been developed to capture the interactive developmental dynamics in language acquisition. These models rely on simple but powerful computational principles of self-organization and Hebbian learning. We have applied them successfully to account for a variety of empirical phenomena in early monolingual lexical development (see Li et al., 2004; Li, Zhao & Macwhiney, 2007). Here we apply a variant of the DevLex-II model to the bilingual context and focus on how the representational structure of the bilingual lexicon can emerge, develop, and change as a function of the learning history. In particular, we manipulate the onset time of L2 lexical learning, in three scenarios: simultaneous – onset time of L2 co-occurs with that of L1, early learning – onset time of L2 is slightly delayed relative to that of L1, and late learning – onset time of L2 lags significantly behind that of L1. We hypothesize that the representational structure for the two lexicons in our model would differ as a function of the learning history defined by L2 onset time. In addition, through analyzing the model’s comprehension and production errors, we hope to show how the two developing lexicons compete and interact with each other. The Model Figure 1: DevLex-II (Li, Farkas, & MacWhinney, 2007) A Sketch of the Model DevLex-II is a multi-layer self-organizing neural network model as diagrammatically depicted in Figure 1 (see Li, et al. 2007 for details). It includes three basic levels for the representation and organization of linguistic information: phonological content, semantic content, and output sequence of the lexicon. The core of the model is a two-