From Words to Senses: A Case Study of Subjectivity Recognition

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 825–832 Manchester, August 2008 From Words to Senses: A Case Study of Subjectivity Recognition Fangzhong Su School of Computing University of Leeds, UK fzsu@comp.leeds.ac.uk Katja Markert School of Computing University of Leeds, UK markert@comp.leeds.ac.uk Abstract We determine the subjectivity of word senses. To avoid costly annotation, we evaluate how useful existing resources es- tablished in opinion mining are for this task. We show that results achieved with existing resources that are not tailored to- wards word sense subjectivity classifica- tion can rival results achieved with super- vision on a manually annotated training set. However, results with different re- sources vary substantially and are depen- dent on the different definitions of subjec- tivity used in the establishment of the re- sources. 1 Introduction In recent years, subjectivity analysis and opinion mining have attracted considerable attention in the NLP community. Unlike traditional information extraction and document classification tasks which usually focus on extracting facts or categorizing documents into topics (e.g., “sports”, “politics”, “medicine”), subjectivity analysis focuses on de- termining whether a language unit (such as a word, sentence or document) expresses a private state, opinion or attitude and, if so, what polarity is ex- pressed, i.e. a positive or negative attitude. Inspired by Esuli and Sebastiani (2006) and Wiebe and Mihalcea (2006), we explore the auto- matic detection of the subjectivity of word senses, in contrast to the more frequently explored task of determining the subjectivity of words (see Sec- tion 2). This is motivated by many words being c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. subjectivity-ambiguous, i.e. having both subjec- tive and objective senses, such as the word positive with its two example senses given below.1 (1) positive, electropositive—having a positive electric charge;“protons are positive” (objective) (2) plus, positive—involving advantage or good; “a plus (or positive) factor” (subjective) Subjectivity labels for senses add an additional layer of annotation to electronic lexica and allow to group many fine-grained senses into higher-level classes based on subjectivity/objectivity. This can increase the lexica’s usability. As an example, Wiebe and Mihalcea (2006) prove that subjectiv- ity information for WordNet senses can improve word sense disambiguation tasks for subjectivity- ambiguous words (such as positive). In addition, Andreevskaia and Bergler (2006) show that the performance of automatic annotation of subjectiv- ity at the word level can be hurt by the presence of subjectivity-ambiguous words in the training sets they use. Moreover, the prevalence of different word senses in different domains also means that a subjective or an objective sense of a word might be dominant in different domains; thus, in a sci- ence text positive is likely not to have a subjective reading. The annotation of words as subjective and objective or positive and negative independent of sense or domain does not capture such distinctions. In this paper, we validate whether word sense subjectivity labeling can be achieved with existing resources for subjectivity analysis at the …

Content truncated to 500 words