Abstract: Word alignment is an important part of statistical machine translation systems.Previous works obtain word alignment through sequential models,which do not take into account the structure information and linguistic features of the language,leading to bad word alignments violating linguistic characteristics.This paper proposed a novel self-correction method for word alignments,aiming to correct the alignment errors which violate linguistic characteristics by exploiting linguistic prior knowledge.First,we conducted a coarse correction on short alignments obtained by binary segmentation based on punctuation method.Second,we proposed a fine-grained correction method for each short alignment based on statistical features.Third,corrected short alignments were merged to original alignments.This process does not rely on any third-party word aligner and additional parallel corpus.Experimental results show that our method significantly improves the accuracy machine translation results.

Key words: Self-correction,Word alignment,Coarse-to-fine