Adapting to Non-Centered Languages for Zero-shot Multilingual Translation

  • 2022-09-09 07:34:12
  • Zhi Qu, Taro Watanabe
  • 4

Abstract

Multilingual neural machine translation can translate unseen language pairsduring training, i.e. zero-shot translation. However, the zero-shot translationis always unstable. Although prior works attributed the instability to thedomination of central language, e.g. English, we supplement this viewpoint withthe strict dependence of non-centered languages. In this work, we propose asimple, lightweight yet effective language-specific modeling method by adaptingto non-centered languages and combining the shared information and thelanguage-specific information to counteract the instability of zero-shottranslation. Experiments with Transformer on IWSLT17, Europarl, TED talks, andOPUS-100 datasets show that our method not only performs better than strongbaselines in centered data conditions but also can easily fit non-centered dataconditions. By further investigating the layer attribution, we show that ourproposed method can disentangle the coupled representation in the correctdirection.

 

Quick Read (beta)

loading the full paper ...