The same entity in a document often occurs in multiple expressions,forming a series of complex referen-tial relations.To deal with this challenge to Document-level translation,this paper focuses on integrating reference resolution into Document-level machine translation.First,we design the reference representation for the reference chain;Then,we introduce a new translation mechanism to fuse reference representation by combining soft constraint and hard constraint.Experiments on English-German and Chinese-English datasets show that compared with the best sentence-level translation system,our methods can significantly improve translation performance.In addition,experiments show the our methods lead to a strong improvement in the task of evaluating the translation quality of English-German pronouns.
维普
万方数据