Detecting Fine-Grained Cross-Lingual Semantic Divergences without Supervision by Learning to Rank

TL;DR

This paper presents a novel ranking-based training strategy for multilingual BERT models to detect fine-grained semantic divergences across languages without supervision, improving accuracy in cross-lingual NLP tasks.

Contribution

It introduces a ranking-based training approach using synthetic divergent examples and a new dataset for English-French semantic divergence detection, advancing unsupervised fine-grained divergence analysis.

Findings

Ranking-based training improves divergence detection accuracy.

Token-level predictions help distinguish divergence granularity.

Models outperform strong similarity-based baselines.

Abstract

Detecting fine-grained differences in content conveyed in different languages matters for cross-lingual NLP and multilingual corpora analysis, but it is a challenging machine learning problem since annotation is expensive and hard to scale. This work improves the prediction and annotation of fine-grained semantic divergences. We introduce a training strategy for multilingual BERT models by learning to rank synthetic divergent examples of varying granularity. We evaluate our models on the Rationalized English-French Semantic Divergences, a new dataset released with this work, consisting of English-French sentence-pairs annotated with semantic divergence classes and token-level rationales. Learning to rank helps detect fine-grained sentence-level divergences more accurately than a strong sentence-level similarity model, while token-level predictions have the potential of further distinguishing between coarse and fine-grained divergences.