Domain Mismatch Doesn't Always Prevent Cross-Lingual Transfer Learning
This addresses the issue of domain mismatch for researchers and practitioners in cross-lingual NLP tasks like unsupervised translation and lexicon induction, though it is incremental as it builds on existing methods with a new initialization approach.
The paper tackles the problem of domain mismatch hindering cross-lingual transfer learning by showing that a simple initialization regimen using pre-trained embeddings on concatenated domain-mismatched corpora can recover a large portion of the losses, challenging prior claims that domain mismatch prevents effective transfer.
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.