Sentiment Analysis is a new task of natural language processing. The aim of this task is to recognize, analyze and understand the opinions in the texts, which involoves the key technologies of pattern recognition, machine learning, information retrieval, information extraction, etc. Thus, the research on sentiment analysis has significant academic value. Furthermore, with the development of Web 2.0 and the rise of Web 3.0, automatic sentiment analysis of texts will be beneficial for public opinion analysis, market research, society information security and so on. Therefore, it is also very useful for real applications. Under this background, this dissertation focuses on the sentiment classification task in sentiment analysis. The main contributions are summarized as follows: (1). Sentence-Level Sentiment Classification: To capture the contextual constraints on the sentence sentiment, this dissertation regards the the sentences in a passage as a sentiment flow instead of isolated points. Therefore, the orginal sentence-level sentiment classification task is converted into a sequential labeling task and the contextual information can be considered. At the same time, to capture label redundancy, this dissertation introduces redundant labels into the original sentimental label set and adds redundant label features into statistic model, so that the performance of sentiment classification can be improved. Experiments on sereval sentiment analysis tasks (including subjective classification, polarity classification and sentiment strength rating) prove the effectiveness of our approach. (2). Document-Level Sentiment Classification: There are two difficulties for identifying the overall sentiment in a document: 1). A document may contain multiple sentiments for different object’s facets (loacal sentiments). How to summarize the overall sentiment of the document from these diversified local sentiments? 2). When a document-level sentiment classifier is trained, only the overall sentiment labels can be observed in the training dataset, and the local sentiments (may be inconsistent with the overall sentiment) are unoberved, so that some feature weights may be biased after training because these features are not connected with their real sentiment labels. How to revise these biased feature weights? For the first problem, we regard the overall sentiment is an intergration of all the local sentiments in a document with different weights. Therefore, we present a novel ...
修改评论