Domain adaption based on lda and word embedding in SMT

Abstract

Current methods about domain adaption in SMT mostly assume that a small in-domain sample is need at training time. However, the fact target domain may not be known at training time so that it may not satisfy the fact translation or is far away from user needs. We instead propose a more suitable method to avoid this situation. Our methods mainly contain two sections (1) Firstly, we use word embedding and LDA model to divide the training corpus into some similar semantic subdomains. (2) Secondly, for an actual source sentences we can select a more suitable translation system by semantic clues. We implement experiments on two language pairs. We can observe consistent improvements over three baselines.