Bilingual lexical interactions in an unsupervised neural network model

Abstract

In this paper we present an unsupervised neural network model of bilingual lexical development and interaction. We focus on how the representational structures of the bilingual lexicons can emerge, develop, and interact with each other as a function of the learning history. The results show that: (1) distinct representations for the two lexicons can develop in our network when the two languages are learned simultaneously; (2) the representational structure is highly dependent on the onset time of the second language (L2) learning if the two languages are learned sequentially; and (3) L2 representation becomes parasitic on the representation of the first language when the learning of L2 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 involves limited plasticity as a function of the timing of L2 learning.

Keywords:

• bilingual interaction
• lexical development
• neural network
• DevLex

Acknowledgements

We would like to thank two anonymous reviewers and the editor for their valuable comments and suggestions on the earlier versions of this article. Preparation of this article was made possible by a grant from the National Science Foundation (BCS-0642586) to Ping Li and the Colgate University faculty research grant to Xiaowei Zhao.

Notes

1. In some models, weights may be prewired and cannot be changed, which is the case, for example, in earlier models like the IA model of McClelland and Rumelhart (1981).

2. We have also obtained some preliminary results of applying our model to bilingual acquisition, on which the current study is based (see Zhao and Li 2007, 2008).

3. The detailed procedure of implementing this new neighborhood function is described in Li, Zhao, and MacWhinney (2007).

4. We excluded homographs in our simulations because the unique semantic representations for them are difficult to get; and excluded phrases because they include more than one word. See Bates et al. (1994) for reasons for excluding the other four types of words from a normal analysis of vocabulary development.

5. In separate simulations, we obtained similar results when English was L1 and Chinese was L2.

6. The number in the Chinese phonetic transcription indicates the tone of the corresponding word.

7. Initially, high density may bring a certain advantage to the learning of novel words in the dense areas, in that once a novel word is learned, its close neighbors may be more easily mapped to the semantic category to which they belong. However, the disadvantages caused by strong competition and high confusion could overwhelm the advantages eventually.

8. We constructed separate models in which either Chinese or English was the L2 (with the same modeling parameters). Given that the results from these models were very similar, we report here mainly the results from modeling Chinese as L2 and English as L1.