Domain Adaptation in NLP Based on Hybrid Generative and Discriminative Model

Abstract

This study investigates the domain adaptation problem for nature language processing tasks in the distributional view. A novel method is proposed for domain adaptation based on the hybrid model which combines the discriminative model with the generative model. The advantage of the discriminative model is to have lower asymptotic error, while the advantage of the generative model can easily incorporate the unlabeled data for better generalization performance. The hybrid model can integrate their advantages. For domain transfer, the proposed method exploits the difference of the distributions in different domains to adjust the weights of the instances in the training set so that the source labeled data is more adaptive to the target domain. Experimental results on several NLP tasks in different domains indicate that our method outperforms both the traditional supervised learning and the semi-supervised method.