Integrating Knowledge for Subjectivity Sense Labeling

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 10–18, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Integrating Knowledge for Subjectivity Sense Labeling Yaw Gyamfi and Janyce Wiebe University of Pittsburgh {anti,wiebe}@cs.pitt.edu Rada Mihalcea University of North Texas rada@cs.unt.edu Cem Akkaya University of Pittsburgh cem@cs.pitt.edu Abstract This paper introduces an integrative approach to automatic word sense subjectivity annota- tion. We use features that exploit the hier- archical structure and domain information in lexical resources such as WordNet, as well as other types of features that measure the sim- ilarity of glosses and the overlap among sets of semantically related words. Integrated in a machine learning framework, the entire set of features is found to give better results than any individual type of feature. 1 Introduction Automatic extraction of opinions, emotions, and sentiments in text (subjectivity analysis) to support applications such as product review mining, sum- marization, question answering, and information ex- traction is an active area of research in NLP. Many approaches to opinion, sentiment, and sub- jectivity analysis rely on lexicons of words that may be used to express subjectivity. However, words may have both subjective and objective senses, which is a source of ambiguity in subjectivity and sentiment analysis. We show that even words judged in pre- vious work to be reliable clues of subjectivity have significant degrees of subjectivity sense ambiguity. To address this ambiguity, we present a method for automatically assigning subjectivity labels to word senses in a taxonomy, which uses new features and integrates more diverse types of knowledge than in previous work. We focus on nouns, which are challenging and have received less attention in auto- matic subjectivity and sentiment analysis. A common approach to building lexicons for sub- jectivity analysis is to begin with a small set of seeds which are prototypically subjective (or posi- tive/negative, in sentiment analysis), and then fol- low semantic links in WordNet-like resources. By far, the emphasis has been on horizontal relations, such as synonymy and antonymy. Exploiting vertical links opens the door to taking into account the infor- mation content of ancestor concepts of senses with known and unknown subjectivity. We develop novel features that measure the similarity of a target word sense with a seed set of senses known to be sub- jective, where the similarity between two concepts is determined by the extent to which they share in- formation, measured by the information content as- sociated with their least common subsumer (LCS). Further, particularizing the LCS features to domain greatly reduces calculation while still maintaining effective features. We find that our new features do lead to signif- icant improvements over methods proposed in pre- vious work, and that the combination of all features gives significantly better performance than any sin- gle type of feature alone. We also ask, given that there are many approaches to finding subjective words, if it would make sense for word- and sense-level approaches to work in tan- dem, or should we …

Content truncated to 500 words