Research On Model And Method For Opinion Target Recognition

With the development of Internet technology, electronic commerce has become a more and more indispensable portion in the daily life of people, followed by the rapid growth of user comments and reviews. These comments and reviews contain important information about how the consumers evaluate the commodity with respect to the related functions and properties. Effective utilization of these information can bring benefits to improving product quality and understanding the real needs of consumers, which stimulates the generation and technological development of opinion target recognition. An opinion target is the entity composing of one or more words on which opinion holders express opinion in the customer reviews, and opinion target recognition is to extract the real opinion entities accurately. From the methodological point, opinion target recognition can be classified into supervised learning, unsupervised learning and semi supervised learning; from the point of application, opinion target recognition can be classified into single domain problem and cross domain problem. This thesis mainly focuses on the supervised learning opinion target recognition in a single domain, and analyze the advantages and disadvantages of each model by experiments. Research methods in this thesis can be summarized as the following three categories:Firstly, methods based on unsupervised learning. This thesis employ the rule-based methods to recognize extract nouns and noun phrases which appear with high frequency in the datasets using association-rule mining, and then uses word semantic relativity to get the candidates of true opinion targets in the candidate set. On this basis, this thesis introduces a methods based on syntactic analysis and double propagation algorithm, which respectively aims at recognizing the opinion targets made up of noun phrases and those with low frequency in the datasets.Secondly, methods based on temporal models. In general, customer reviews are essentially a contextual sequence of words, which can be modeled effectively by temporal models. This thesis extracts surface features, statistic features and syntax features of the sentence and learn the interrelation between them by Conditional Random Field model. Experimental results show that feature combination has a great impact, and the model can achieve a quite great performance with appropriate features.Finally, methods based on recurrent neural networks. Recurrent neural networks are end-to-end models, which get rids of extra preprocessing techniques and task specific feature engineering efforts. This thesis compares the performance of several existing recurrent neural networks, which shows the deficiencies that it is difficult for existing recurrent neural networks to capture the temporal dependencies output label sequences. To address this problem, a new output perception recurrent neural network is proposed in this thesis. It is shown that the proposed model can beat the conventional recurrent neural networks not only in effectiveness but also in the rate of convergence.