Abstract: The task of sentiment classification is domain-specific,i. e. ,a classifier learning from the annotated data from a domain often performs dramatically badly on the data from a different domain. We presented a novel approach for cross-domain sentiment classification. Specifically, we first generalized four general categories of the opinion targets; o- verall, software, hardware, and service and classified all sentences into these categories. hhen, some sentences with the category information were annotated in the source domain and a classifier for opinion target categorization was developed with the annotated data to classify all the sentences in both the source and target domain. Third, the four categories of opinion targets were considered as four different views which arc employed in a standard co-training algorithm to per- form cross-domain sentiment classification. Experimental results across several domains of Chinese reviews demonstrate the effectiveness of the proposed approach.

Key words: Opinion target, Co-training, Maximum entropy, Cross-domain sentiment classification