Research On Key Technologies Of Opinion Mining Towards Product Reviews

Nowadays, there is a wealth of information on the Internet, making the Internet become the main means for people to access to information. People can express their opinion through the internet, for example through the forums. Before people buy a product, they may find related comments about this product online. Many product forums now provide people with relevant information. However, people simply go through the user review one by one, eventually get the impression about the product, which is no doubt a tedious and time-consuming task. If such information can be effectively aggregated, it will improve the efficiency of people's search. Considering on such demand, we have constructed an opinion mining system towards mobile phone reviews. The main research contents include the following:Firstly, this paper describes the useful reviews filtering module. This module is the first step of opinion mining system, because the origin reviews are intermingled with the good and the bad. The useless reviews may have side effect to the follow-up steps. Based on the definition of useful reviews, this paper use support vector machine, introducing one feature that the co-occurrence of product feature and sentiment word. Experiment results show that the introduction of the co-occurrence feature can improve classification results.Secondly, this paper describes the sentiment block identification and sentiment analysis module. This module is the core module of the whole system. This paper treats this problem as a word sequence labeling problem, and uses Conditional Random Fields to identify sentiment blocks, integrating template feature. Experiments show that such methods can identify not only blocks with sentiment words, but also blocks without sentiment words, such blocks are mostly colloquial expression. In addition, the introduction of template feature can further improve the results.Finally, this paper introduces the word of mouth aggregation module. This module mainly contains two steps: product feature word-sentiment block pairs mining and word of mouth aggregation. The method of product feature word-sentiment block pairs mining is: For a given sentiment block, we treat the nearest product feature word as the word paired with this block. After this step, system groups all the reviews according to different product types, in each group we aggregate the word of mouth.