Aligning Very Small Parallel Corpora Using Cross-Lingual Word Embeddings and a Monogamy Objective

TL;DR

This paper introduces an unsupervised method using cross-lingual word embeddings and a monogamy objective to improve word alignment on very small parallel corpora, outperforming traditional count-based methods.

Contribution

The paper presents a novel unsupervised approach that adapts cross-lingual embeddings for small datasets, enhancing alignment accuracy.

Findings

Outperforms fast-align on small datasets (25-500 sentences)

Fine-tuning CLWEs improves baseline performance

Effective on very limited parallel data

Abstract

Count-based word alignment methods, such as the IBM models or fast-align, struggle on very small parallel corpora. We therefore present an alternative approach based on cross-lingual word embeddings (CLWEs), which are trained on purely monolingual data. Our main contribution is an unsupervised objective to adapt CLWEs to parallel corpora. In experiments on between 25 and 500 sentences, our method outperforms fast-align. We also show that our fine-tuning objective consistently improves a CLWE-only baseline.