Towards Domain Independent Product Review Analysis

The large amount of product reviews published by the users are quite useful for those who wants to buy the product, and also valuable for the manufactures to improve the product quality. However, tremendous amount of reviews available on Internet make useful information inaccessible. With the help of a system that capable performing review analysis based on Natural Language Processing technology, users can get a better understanding about the product.The problem of product review analysis generally consists of two partments: products' entity extraction and opinion judgement, each of which has been addressed by many researchers. However, when applied to a new product domain, most product analysis systems suffer from high cost of building the training samples for both two sub tasks. This thesis addressed the problem of developing a domain independent review analysis system. For the first sub-task, a semi-supervised extraction algorithm was proposed. Given the review documents of a specific domain, the algorithm can extract the product description words with only a few predefined seed words. For the second, I proposed a novel unsupervised sentimental classifier, which is based on search engine's potential to indicate how the words are related to one and another. The experiment showed that the proposed method is effective and domain independent.