Reshaping Word Embedding Space With Monolingual Synonyms for Bilingual Lexicon Induction

Recent research on Bilingual Lexicon Induction (BLI) involves mapping monolingual word embeddings (WEs) into a shared space and obtaining word translations by retrieving the nearest cross-lingual neighbors. However, we observe that word vectors of some synonyms in the monolingual embedding space are located far apart, resulting in a reduced isomorphism between two monolingual WE spaces, which brings greater challenges in the WE spaces' mapping. To address this issue, we propose a novel method for enhancing BLI with synonyms. Specifically, we first build a monolingual synonym lexicon, which contains synonym pairs obtained from Wordnet. We then generate pseudo translation pairs that include synonyms of the original seed lexicon. Finally, the BLI model is jointly trained on the original and pseudo translation pairs, and the WE space is reshaped by bringing monolingual synonyms closer and aligning them with their cross-lingual translations simultaneously. The extensive experiments show that our method significantly outperforms the strong baseline models in word translation.