Abstract: A three-level filter method was proposed to extract the opinion targets for product reviews on the Internet. In the first level, a bootstrapping framework was adopted to extract candidate opinion targets and opinion words from opinion texts. In the second level, the association between the opinion target and opinion word was used to estimate the association confidence of every candidate opinion target and candidate opinion word. The opinion targets with high association confidence were extracted to improve recognition accuracy. In the third level, an uncorrelated domain was adopted to calculate the domain confidence of every opinion target in the rest set which was for mining the opinion targets of low frequency. The experimental results on three public datasets demonstrate the effectiveness of the proposed approach.