A Study On The Key Problems In Text Emotion Analysis

With the booming of social networks, there are many subjective docu-ments apperaring on the Web such as fourm posts, tweets, online reviews and so on. The subjective information plays an important role in Web mining. Sen-timent analysis, also called opinion mining, is the field of study that analyzes people’s attitudes (support or oppose, like or dislike, etc.) and emotions (hap-piness, anger, sadness, fear, etc.) towards entities such as products, services, organizations, individuals, events and issues. This thesis focuses on some key issues of sentiment analysis in different text levels.In word-level sentiment analysis, we take account of the fact that the sen-timent orientations of words may change in different domains. This thesis pro-poses Affinity Propagation algorithm (AP) to determine word semantic orien-tations in specific domains. On the basis of the Activation Force model, AP ini-tially builds word affinity network for the initial corpus, which calculates struc-tual link similarities between words through their semantic activation relations. AP then represents each word as an informative affinity vector to calculate it-s semantic similarities with some seed sentiment words, thereby propagating sentiment information through the entire network. The experiment results have demonstrated that AP can construct domain-specific sentiment lexicons effi-ciently.Document-level sentiment analysis is usually regared as a classification task. This thesis explores how feature selection affects the performance of document sentiment classification. Inspired by Linear Discriminative Analy-sis, the unsupervised Sentiment Discriminative Analysis (SDA) is presented, which builds objective fuction through local scatter matrices of each document and calculates features’ SDA scores by solving optimal linear classifier. On the other hand, the Activation Force model is applied to Sentiment Strength Calculation (SSC) to compute the sum of affinities between a feature and prior sentiment words. SDA preserves local sentiment structure between documents, while SSC emphasizes feature’s global sentiment orientation in the corpus. Therefore, they supplement each other to some degree. This thesis incorporates linearlly SDA and SSC into the final Unsupervised Sentiment-bearing Feature Selection algorithm (USFS).Document-level sentiment classification may be sometimes too coarse-grained to meet the needs of some applications, it is thus necessary to conduct element-level sentiment analysis. Element-level sentiment analysis is a more fine-grained technique which focuses on the sentiment elements in the quin-tuple definition. Aspect extraction is a main task in element-level sentiment analysis, where confidence estimation is crucial to ensure the extraction perfor-mance. This thesis presents a two-step estimation method:estimating mutually the candidate product features and dependency patterns by Prevalence and Re-liability in Pattern-based Bootstrapping; clustering the candidate features in-to aspect clusters and filtering them according to Compactness and Texture. The experiments on electronic product reviews have verified that the proposed method can ensure precision and recall simultaneously in aspect extraction. In addition, on the basis of aspect extraction, this thesis aggregates iteratively the sentiment strengths of opinion collcations for each aspect to calculate its global reputation.Finally, this thesis exploits an online demo system called Sentiment Sum-marization on Product Aspects (SSPA), which integrates word, sentence and element levels of sentiment analysis to convert the unorganized review set in-to the sturctural aspect-based sentiment summary. All the aspects about some product type are listed on the top layer. Under the hierarchy of each aspect, positive and negative review sentences are ranked respectively in a descend-ing order of sentiment strengths. With the help of SSPA, it is convenient for consumers as well as manufacturers to browse market information in different aspects of some type of product.