A sentence-level approach to opinion search and sentiment classification problems using natural language low-level features: models and applications

Opinion mining includes two active research directions; opinion search and sentiment classification. The former is the art of searching for subjective statements (peoples' expression of feelings) on any particular subject matter such as products, political decisions and news topics. The subjective expressions are then ranked according to their relevance. In the latter, preferential stances in the subjective expressions are classified into positive, negative, and sometimes neutral stances that can be used as a resourceful bank of intelligence for different applications. In this thesis, we investigate a sentence-level approach to both opinion search and sentiment classification problems. The approach uses different natural language processing techniques for selecting low-level and fine-grained subjective features in order to realize more accurate results. For the opinion search problem, we propose two sentence-level topic-opinion relevance models to investigate their performance on a standard blog data set. The first model uses predicate-argument structures of sentences to improve contextual and semantic relevance between natural language query topics and the retrieved documents. The second model uses discourse representation structures to detect sentence-level opinion dependencies between the key entities of a query topic and the opinion words. The results show that using these low-level linguistic and subjective features gives up to 15.6% improvement over the TREC baselines. Finally, for the sentiment classification problem, we adapt our sentence-level approach to classify sentiment polarity from the consumer products domain. We propose a Sentence Polarity Shift model that identifies sentiment shift patterns and captures only the consistent polarities from the patterns. The premise is that there are patterns in the way people often communicate their sentiments and these patterns can be predicted by capturing the sentiment shifts of each sentence. Our empirical evidence indicates that using sentences with consistent sentiment polarities combined with additional sentence-level and document-level features is likely to reduce sentiment classification loss and improve the performance of a sentiment classifier by 8.5%