Abstract: In order to improve the performance of sentiment analysis on the review texts of e-commerce sites, an approach was proposed to construct a domain affective ontology. The affective ontology, which was constructed, clearly described the semantic relation between a product and its components, a product and its attributes, as well as associations between product features and affective words. Particularly, collocations of feature words and affective words were extracted by POS patterns and their sentiment polarity was retroactively predicted by the polarity labels of review sentences and negations. Furthermore, node matching rules of ontology were put forward for sentiment analysis. The experimental results show that this domain affective ontology can provide the useful foundation for eliminating the domain-dependence of affective words and mining implicit features in product reviews.