Domain Mismatch Doesn't Always Prevent Cross-Lingual Transfer Learning

TL;DR

This paper demonstrates that simple initialization strategies can significantly mitigate the negative effects of domain mismatch in cross-lingual transfer learning, enabling effective zero-shot performance across various tasks.

Contribution

The study shows that pre-training embeddings on concatenated, domain-mismatched corpora and using these as initializations can recover much of the performance loss caused by domain differences.

Findings

Proper initialization recovers large performance portions lost due to domain mismatch.

Pre-training on combined corpora improves results in UBLI, UNMT, and word similarity tasks.

Challenging prior claims, the work shows domain mismatch is less prohibitive with simple methods.

Abstract

Cross-lingual transfer learning without labeled target language data or parallel text has been surprisingly effective in zero-shot cross-lingual classification, question answering, unsupervised machine translation, etc. However, some recent publications have claimed that domain mismatch prevents cross-lingual transfer, and their results show that unsupervised bilingual lexicon induction (UBLI) and unsupervised neural machine translation (UNMT) do not work well when the underlying monolingual corpora come from different domains (e.g., French text from Wikipedia but English text from UN proceedings). In this work, we show that a simple initialization regimen can overcome much of the effect of domain mismatch in cross-lingual transfer. We pre-train word and contextual embeddings on the concatenated domain-mismatched corpora, and use these as initializations for three tasks: MUSE UBLI, UN Parallel UNMT, and the SemEval 2017 cross-lingual word similarity task. In all cases, our results challenge the conclusions of prior work by showing that proper initialization can recover a large portion of the losses incurred by domain mismatch.