Fine-Grained Sentiment Analysis Oriented In Product Domain

With the development of Internet information in the whole world, a large number of user reviews increase with exponential shape. Thus, how to analyze this sentimental information and extract user-needed information arouses the interest of Natural Language Processing (NLP) researchers. Sentiment Analysis as a branch of NLP with potential applications is gradually becoming a new hot research issue. Sentiment Analysis, also known as opinion mining, is to find relevant sources, extract related information with opinion, process this text and set summarization for further tasks.Sentiment Analysis involves a number of emotional and challenging research tasks. According to the different research field, Sentiment Analysis involves sentiment classification, sentiment extraction, opinion search and opinion summary. Sentiment extraction aims to extract text that is more important and valuable for users. Sentiment extraction has more practical value. Around the sentiment analysis, the thesis includes the following contents:(1) Research on the construction of sentiment resources. To solve the lack of sentiment resources, the thesis presents three methods to expand sentiment lexicon. The experimental results show that the so achieved sentiment lexicon is more effective.(2) Research on the extraction of product attributes. Product attribute extraction is a very valuable task for sentiment analysis. In order to improve the precision of product attributes extraction, we apply Conditional Random Field (CRF) model and Maximum Entropy (ME) model with the combination of several important linguistic features. We also analyze the role of each features such as Part of Speech (PoS), shallow parsing to improve our result.(3) Research on the sentimental domain adaptation. To deal with the lack of sentimental domain corpus, we propose an active learning method to label the data in CRF modeling. The experimental results show that the domain adaptation of active learning achieves the effective performance from electronic product domain to car domain.(4) Design and implementation of a sentiment analysis system. This system includes opinion objects extraction, opinion polarity recognition and sentiment text classification and integrates our proposed method in sentiment analysis research.