Low-resource bilingual lexicon extraction using graph based word embeddings

TL;DR

This paper introduces a graph-based method for extracting bilingual lexicons in low-resource settings, outperforming traditional Word2Vec approaches by leveraging translation pairs from unsupervised alignments.

Contribution

The paper proposes a novel graph-based approach to generate bilingual word vectors from limited data, improving translation accuracy in low-resource scenarios.

Findings

Graph-based vectors outperform Word2Vec in low-resource conditions

Linear transformations effectively translate words between languages

Method is successful for Spanish-Nahuatl bilingual lexicon extraction

Abstract

In this work we focus on the task of automatically extracting bilingual lexicon for the language pair Spanish-Nahuatl. This is a low-resource setting where only a small amount of parallel corpus is available. Most of the downstream methods do not work well under low-resources conditions. This is specially true for the approaches that use vectorial representations like Word2Vec. Our proposal is to construct bilingual word vectors from a graph. This graph is generated using translation pairs obtained from an unsupervised word alignment method. We show that, in a low-resource setting, these type of vectors are successful in representing words in a bilingual semantic space. Moreover, when a linear transformation is applied to translate words from one language to another, our graph based representations considerably outperform the popular setting that uses Word2Vec.