Cross-Lingual Word Embeddings and the Structure of the Human Bilingual Lexicon

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 110–120 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 110 Cross-lingual word embeddings and the structure of the human bilingual lexicon Paola Merlo University of Geneva Paola.Merlo@unige.ch Maria A. Rodriguez University of Geneva Maria.AnduezaRodriguez@unige.ch Abstract Research on the bilingual lexicon has uncov- ered fascinating interactions between the lex- icons of the native language and of the sec- ond language in bilingual speakers. In par- ticular, it has been found that the lexicon of the underlying native language affects the or- ganisation of the second language. In the spirit of interpreting current distributed repre- sentations, this paper investigates two models of cross-lingual word embeddings, compar- ing them to the shared-translation effect and the cross-lingual coactivation effects of false and true friends (cognates) found in humans. We find that the similarity structure of the cross-lingual word embeddings space yields the same effects as the human bilingual lexi- con. 1 Introduction Research on the bilingual lexicon has uncov- ered fascinating interactions between the L1 (na- tive language) and L2 (second language) lexicons showing that both production and comprehension coactivate lexical items in both languages, indi- cating that bilinguals store lexical representations from their native and their second language in the same space (Kroll and Dijkstra, 2012; Williams, 2014).1 This paper presents the first bilingual investiga- tion of models of cross-lingual word embeddings and asks whether the bilingual spaces they define 1Throughout this paper, and following the current liter- ature on the topic, we use the term ‘bilingual’ loosely, to refer to any speaker of more than one language. Although there has been much research on all aspects of bilingualism, and at all stages of proficiency, the effects we model here have been found in experiments testing speakers who began to learn their second language after their first, usually in a school context, and who are at an advanced level of profi- ciency. The form German-English below will indicate, for example, a native speaker of German who learnt English as a second language (Williams, 2014). have similar properties to the human bilingual lex- icon. Among the many questions and results in the vast bilingualism literature, we concentrate on coactivation effects in items with shared transla- tions. We also study interference or facilitatory ef- fects in form-meaning mapping, the case of false friends, words that share form but differ in mean- ing across the two languages, and true friends, words that share both form and meaning. We find that the similarity structure of cross-lingual word embeddings matches well with known experimen- tal findings of the human bilingual lexicon. 2 The structure of the bilingual lexicon The core findings about the bilingual lexicon con- firm that the two languages occupy an integrated space and they interact with each other (Wolter, 2001). For example, both in the monolingual and bilingual lexicon the best predictor of the time to recognise a word is the number of similarly spelled words, within and across languages (John- …

Content truncated to 500 words