A Discriminative Latent-Variable Model for Bilingual Lexicon Induction

TL;DR

This paper presents a new discriminative latent-variable model for bilingual lexicon induction that combines prior structures with representation-based methods, improving lexicon quality across multiple language pairs.

Contribution

It introduces an efficient Viterbi EM algorithm for training and shows how prior integration enhances bilingual lexicon induction performance.

Findings

Improved lexicon quality on six language pairs

Prior integration enhances induction accuracy

Efficient training via Viterbi EM algorithm

Abstract

We introduce a novel discriminative latent variable model for bilingual lexicon induction. Our model combines the bipartite matching dictionary prior of Haghighi et al. (2008) with a representation-based approach (Artetxe et al., 2017). To train the model, we derive an efficient Viterbi EM algorithm. We provide empirical results on six language pairs under two metrics and show that the prior improves the induced bilingual lexicons. We also demonstrate how previous work may be viewed as a similarly fashioned latent-variable model, albeit with a different prior.