(A) bootstrapping technique for acquiring domain-specific features in sentiment analysis

Abstract

People are eager to know what others are thinking or feeling about subject matters such as products, politicians, and social issues, as witnessed by a rapid growth in online discussion groups and review sites (e.g., Yahoo forum, Amazon, New York Times). An automatic mining of subjective texts that convey people’s negative or positive sentiments towards specific objects are quite useful for individuals, governments, and companies. Reflecting the importance, many researchers have studied the area of sentiment analysis which includes the sub-tasks of sentiment-bearing text identification, polarity determination, and sentiment target identification. These tasks all exploit sentiment clue words (e.g., “angry”, “happy”): a sentence is deemed negative if it contains negative clues, for example. However, previous approaches to clue identifications fail to associate clues to specific topics or domains. The thesis addresses the problem of automatically generating domain-specific sentiment clues that are specific to a domain. The domain-specific nature of sentiment classification makes it essential to develop a clue lexicon for each domain, especially with news articles that cover diverse domains. Based on our observation that a sentiment clue is often syntactically related to a sentiment topic in a sentence, which is defined as a primary subject of sentiment expression, such as event, company, and person, we developed a novel method for automatically extracting sentiment clues in different domains. The main idea is to bootstrap from a small set of seed clues and generate new clues by utilizing linguistic dependencies and collocation information between sentiment clues and sentiment topics. A newly learned classifier with the new set of clues makes it possible to continue the bootstrapping process. We ran experiments to see the number of iterations required for convergence and show the technique is effective in building domain-specific sentiment classifiers.