Research On The Key Technologies Of Text Sentiment Information Extraction

With the development and popularization of Web2.0 technology, the model of Internet users have changed from the simple Internet information obtainment to the Internet creation, which realized the two-way communication of Internet users and Websites. As a result, a number of subjective texts generated from the Internet. These texts contain various sentiment and attitudes of people. How to analyze and process these massive texts has become a hot topic of researches, text sentiment information extraction and analysis arises under this background. It has a broad application prospect in domains of user comments analysis and decision making, public opinion monitoring and information forecast, thus, there are many scholars and institutions have put in study. But the study is still in a development stage and many methods are not mature now, so this subject deeply studies the text sentiment information extraction in the product reviews domain, the specific research contents are as follows:(1) Opinion Target Extraction. In the product reviews, opinion target is vital information. For the task of opinion target extraction in the product reviews, this subject proposes a domain-independent opinion target extraction method, called the M-Score algorithm. The thought of this algorithm is derived from the Pointwise Mutual Information algorithm, and its biggest advantage is the domain-independent feature, which is easier to be transplanted. Firstly, the candidate opinion targets are extracted by the use of Conditional Random Field model. Then the M-Score algorithm is employed for domain-related processing of the candidate opinion targets. Finally, the further screening is done to obtain the final opinion targets. The experiments adopt different domain corpus to verify the algorithm, and the experiment results proved the validity of the algorithm well.(2) Appraisal Expression Extraction. For the task of appraisal expression <opinion target, opinion words> extraction, this subject proposes a method based on semantic analysis and dependency parsing. This method introduces four types of semantic features and twenty dependency templates. The introduction of semantic features makes up the imperfect of dependency templates and the instability of syntactic analysis, and the dependency parsing eliminates the lexical ambiguity at the same time. The combination of semantic analysis and dependency parsing makes the extraction more accurate, and the experimental results verified the effectiveness of the proposed method.(3) The construction of sentiment lexicon. After the extraction of opinion target, the sentiment tendency analysis is needed. This subject constructs a new sentiment lexicon based on How Net and Pointwise mutual information for the task of opinion target sentiment tendency analysis. The new sentiment lexicon makes the necessary correction and supplement to the existing sentiment lexicons, and it improves the calculation method of word similarity and the calculation rules of polarity intensity. This subject uses this sentiment lexicon to analyze the polarity and calculate the polarity intensity of opinion words which modified the opinion target. Compared with the existing sentiment lexicon, this new sentiment lexicon achieved different degrees of improvement in precision, recall and F-measure.