Abstract

The demand for information about sentiment expressed in texts has stimulated a growing interest into automatic sentiment analysis in Natural Language Processing (NLP). This dissertation is motivated by an unmet need for high-performance domain-independent sentiment taggers and by pressing theoretical questions in NLP, where the exploration of limitations of specific approaches, as well as synergies between them, remain practically unaddressed. This study focuses on sentiment tagging at the sentence level and covers four genres: news, blogs, movie reviews, and product reviews. It draws comparisons between sentiment annotation at different linguistic levels (words, sentences, and texts) and highlights the key differences between supervised machine learning methods that rely on annotated corpora (corpus-based, CBA) and lexicon-based approaches (LBA) to sentiment tagging. Exploring the performance of supervised corpus-based approach to sentiment tagging, this study highlights the strong domain-dependence of the CBA. I present the development of LBA approaches based on general lexicons, such as WordNet, as a potential solution to the domain portability problem. A system for sentiment marker extraction from WordNet's relations and glosses is developed and used to acquire lists for a lexicon-based system for sentiment annotation at the sentence and text levels. It demonstrates that LBA's performance across domains is more stable than that of CBA. Finally, the study proposes an integration of LBA and CBA in an ensemble of classifiers using a precision-based voting technique that allows the ensemble system to incorporate the best features of both CBA and LBA. This combined approach outperforms both base learners and provides a promising solution to the domain-adaptation problem. The study contributes to NLP (1) by developing algorithms for automatic acquisition of sentiment-laden words from dictionary definitions; (2) by conducting a systematic study of approaches to sentiment classification and of factors affecting their performance; (3) by refining the lexicon-based approach by introducing valence shifter handling and parse tree information; and (4) by development of the combined, CBA/LBA approach that brings together the strengths of the two approaches and allows domain-adaptation with limited amounts of labeled training data.