Multilingual BERT Based Word Alignment By Incorporating Common Chinese Characters

Word alignment is an important task of detecting translation equivalents between a sentence pair. Although word no longer necessarily needed for neural machine translation, it’s still useful in wealth applications, e.g., bilingual lexicon induction, constraint decoding, and so on. However, the most well-known aligners are Giza++ fastAlign, both which implementations traditional IBM models. To keep pace with advance NMT, there has been surge interest replacing models We follow this trend but aim to boost performance Japanese Chinese, share large portion Chinese characters. Our key idea leverage these common characters languages as indicator inferring alignment; i.e., source target words should be likely aligned. Following idea, we propose three methods that mBERT-based alignment, including reward factor, representation contrastive training. Furthermore, annotate release golden dataset Japanese-Chinese alignment. Experiments on show our outperform several strong baselines terms AER score verify effectiveness exploiting