The Key Technology Research On Opinion Target Extraction

After entering the21st century, more and more reviews in the internet are available with the development of the web. How to deal with the vast amounts of information automatically becomes an important issue, which makes the appearing of the study of sentiment analysis.Opinion target extraction is a basic subtask in sentiment analysis and it aims to recognize the object on which the opinion expressed. In this thesis, we mainly focus on the task of opinion target extraction. Our study mainly includes following aspects:First, we conduct study on opinion target extraction on Chinese text. The adopted method considers the task as a standard sequence labeling problem and the extraction model is trained with Conditional Random Fields (CRFs). The difference between the English and Chinese opinion target extraction is analyzed in detail. Besides the features that are popularly used in the English task, we add the features which are captured on the basis of the distinguishing features of the Chinese which improves the whole extraction performance.Second, we address opinion target extraction using a shallow semantic parsing framework. Although the existing methods on opinion target extraction have got a very good result, they fail to consider the syntactic information. To fully use the syntactic knowledge, we present a simplified shallow semantic parsing approach to extract opinion targets. Various of syntactic features are employed in our system to improves the whole extraction performance.Third, we conduct research on domain adaptation problem in opinion target extraction. To reduce the annotation cost on the labeled data in opinion target extraction we propose a domain adaptation approach to this task under the shallow semantic parsing framework. Specifically, we employ two groups of features in our systems, that is the syntactic features and the non‐syntactic features in order to study their contribution on the domain adaptation issue.