Chinese Sentences Subjective And Emotional Integration Of Multi-granularity Classification Study

With the explosive growth of the Internet over the past years, especially Web 2.0, opinion mining is getting considerable attentions from the natural language processing (NLP) community. As key issues in opinions mining, subjectivity analysis focuses on determining whether a given sentence is either subjective or objective, while sentiment classification aims at classifying opinionated documents or sentences as expressing positive, negative or neutral opinions. The two tasks play a critical role in many opinion mining applications such as opinion question answering, opinion retrieval and opinion summarization system.This thesis presents a statistical and fuzzy set based framework for Chinese sentence-level sentiment analysis. In order to reduce the complexity and to improve the efficiency, different granularities of sentiment within sentences are explored and further combined to determine their subjectivity and sentiment orientation. Our research mainly concerns the following four respects:Subjectivity analysis is the first problem to be resolved in opinion mining. Toward large-scale applications, a comprehensive framework for Chinese sentence-level subjectivity analysis system is firstly proposed in this thesis, including:(1) The Chi-square technique is introduced for subjectivity feature selection. (2) A sentiment density-based approach is then applied to represent sentence-level subjectivity. (3) A Bayesian classifier is employed to detect subjective sentences with sentiment density intervals as its features. The preliminary experiments show the feasibility of the proposed method.Lexical subjectivity is the foundation of the phrase-level or sentence-level subjectivity analysis. After a survey of the multil-class and subjectivity strength, a novel method integrating log-linear model with the fuzzy sets is here proposed for the problem of lexical subjectivity classification. The new algorithm runs as follows:The log-linear model is firstly used to calculate the subjectivity weight of the candidate words, and then a fuzzy sets model combining subjectivity weights is then employed to construct the membership functions for evaluating their subjectivity strength. These membership functions are thus applied for Chinese sentence-level subjectivity calssification. Experiments indicate that the proposed method provides a well result.Lexical structural information makes a great contribution towards sentiment oritation analysis of Chinese words. To address the problem of unknown sentiment words in real text, word-internal structural features and local context information are combined under a morpheme-based staitical framework to predict the polarity of unknown sentiment words, which is further used for sentence-level sentiment classification. Compared with other methods, our method takes into account different granularity features.Finally, after an in-depth study of Chinese sentiment oritation at different levels, including morpheme-level, word-level and phrase-level, a multi-granularity fusion method is presented for Chinese sentence-level sentiment classification. Experimental results show that the combination of multiple granularity of sentiment information is beneficial to Chinese sentiment classification.