Shujie Liu, Chi-ho Li, Mu Li, and Ming Zhou
In this paper, based on the investigation of domain adaptation for feature weight, we propose to use co-training framework to handle domain adaptation for feature weight, i.e., we use the translation results from another heterogeneous decoder as pseudo references, and add them to the development data set for minimum error rate training, so as to bias the feature weight to the domain of test data set. Furthermore, we use minimum Bayes-Risk combination for pseudo reference selection, which can pick proper translation results from the translation candidates from both decoders, so as to smooth the training process. Experimental results show that our co-training method with minimum Bayes-Risk combination can yield significant improvements in target domain.
In Journal of Software