Research of Cascaded Conditional Random Fields Model for Sentence Sentiment Analysis Based on Isotonic Constraints

摘要

The sentence sentiment analysis is a key task in sentiment analysis. Existing methods ignored the contextual information, the negative effect of the redundancy between labels, or the relationship from sentiment words to annotation labels. Aiming at these problems, this paper present a novel cascaded model based on isotonic constraints, which respectively classify sentiment polarities and strength in different layers. Different from traditional cascaded model, the proposed method incorporates a kind of domain knowledge about sentiment words through enforcing a set of monotonic constraints on the CRF parameters. Experimental results indicate that the proposed algorithm has strong discrimination ability between different labels, and thus validate the effectiveness of our model in sentence sentiment analysis for Chinese texts.