Word order variances generally exist in different languages. In this paper, we hypothesize that cross-lingual models that fit into the word order of the source language might fail to handle target languages. To verify this hypothesis, we investigate whether making models insensitive to the word order of the source language can improve the adaptation performance in target languages. To do so, we reduce the source language word order information fitted to sequence encoders and observe the performance changes. In addition, based on this hypothesis, we propose a new method for fine-tuning multilingual BERT in downstream cross-lingual sequence labeling tasks. Experimental results on dialogue natural language understanding, part-of-speech tagging, and named entity recognition tasks show that reducing word order information fitted to the model can achieve better zero-shot cross-lingual performance. Furthermore, our proposed methods can also be applied to strong cross-lingual baselines, and improve their performances.
View on arXiv