Revisiting Machine Translation for Cross-lingual Classification

TL;DR

This paper demonstrates that using a stronger machine translation system for cross-lingual classification can significantly improve performance of translate-test methods, challenging the focus on multilingual models.

Contribution

It shows the importance of MT quality and mismatch mitigation in cross-lingual classification, emphasizing the value of MT-based baselines over multilingual models.

Findings

Stronger MT systems improve translate-test performance.

Mismatch mitigation enhances cross-lingual transfer.

Task dependency influences optimal approach choice.

Abstract

Machine Translation (MT) has been widely used for cross-lingual classification, either by translating the test set into English and running inference with a monolingual model (translate-test), or translating the training set into the target languages and finetuning a multilingual model (translate-train). However, most research in the area focuses on the multilingual models rather than the MT component. We show that, by using a stronger MT system and mitigating the mismatch between training on original text and running inference on machine translated text, translate-test can do substantially better than previously assumed. The optimal approach, however, is highly task dependent, as we identify various sources of cross-lingual transfer gap that affect different tasks and approaches differently. Our work calls into question the dominance of multilingual models for cross-lingual classification, and prompts to pay more attention to MT-based baselines.