A Discriminative Neural Model for Cross-Lingual Word Alignment

TL;DR

This paper presents a discriminative neural model for cross-lingual word alignment integrated into a Transformer-based translation system, demonstrating significant improvements in alignment accuracy and downstream NER projection with limited labeled data.

Contribution

It introduces a novel discriminative alignment model that outperforms unsupervised methods and supports future manual annotation efforts.

Findings

Achieved 11-27 F1 improvement over unsupervised baselines.

Enhanced Chinese NER projection accuracy using the alignments.

Validated the model's utility and feasibility through ablation and annotation experiments.

Abstract

We introduce a novel discriminative word alignment model, which we integrate into a Transformer-based machine translation model. In experiments based on a small number of labeled examples (~1.7K-5K sentences) we evaluate its performance intrinsically on both English-Chinese and English-Arabic alignment, where we achieve major improvements over unsupervised baselines (11-27 F1). We evaluate the model extrinsically on data projection for Chinese NER, showing that our alignments lead to higher performance when used to project NER tags from English to Chinese. Finally, we perform an ablation analysis and an annotation experiment that jointly support the utility and feasibility of future manual alignment elicitation.