Subjectivity Recognition on Word Senses via Semi-supervised Mincuts

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 1–9, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Subjectivity Recognition on Word Senses via Semi-supervised Mincuts Fangzhong Su School of Computing University of Leeds fzsu@comp.leeds.ac.uk Katja Markert School of Computing University of Leeds markert@comp.leeds.ac.uk Abstract We supplement WordNet entries with infor- mation on the subjectivity of its word senses. Supervised classifiers that operate on word sense definitions in the same way that text classifiers operate on web or newspaper texts need large amounts of training data. The re- sulting data sparseness problem is aggravated by the fact that dictionary definitions are very short. We propose a semi-supervised mini- mum cut framework that makes use of both WordNet definitions and its relation structure. The experimental results show that it outper- forms supervised minimum cut as well as stan- dard supervised, non-graph classification, re- ducing the error rate by 40%. In addition, the semi-supervised approach achieves the same results as the supervised framework with less than 20% of the training data. 1 Introduction There is considerable academic and commercial in- terest in processing subjective content in text, where subjective content refers to any expression of a pri- vate state such as an opinion or belief (Wiebe et al., 2005). Important strands of work include the identification of subjective content and the determi- nation of its polarity, i.e. whether a favourable or unfavourable opinion is expressed. Automatic identification of subjective content of- ten relies on word indicators, such as unigrams (Pang et al., 2002) or predetermined sentiment lex- ica (Wilson et al., 2005). Thus, the word positive in the sentence “This deal is a positive development for our company.” gives a strong indication that the sentence contains a favourable opinion. How- ever, such word-based indicators can be misleading for two reasons. First, contextual indicators such as irony and negation can reverse subjectivity or po- larity indications (Polanyi and Zaenen, 2004). Sec- ond, different word senses of a single word can ac- tually be of different subjectivity or polarity. A typ- ical subjectivity-ambiguous word, i.e. a word that has at least one subjective and at least one objec- tive sense, is positive, as shown by the 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) We concentrate on this latter problem by automat- ically creating lists of subjective senses, instead of subjective words, via adding subjectivity labels for senses to electronic lexica, using the exam- ple of WordNet. This is important as the prob- lem of subjectivity-ambiguity is frequent: We (Su and Markert, 2008) find that over 30% of words in our dataset are subjectivity-ambiguous. Informa- tion on subjectivity of senses can also improve other tasks such as word sense disambiguation (Wiebe and Mihalcea, 2006). Moreover, Andreevskaia and Bergler (2006) show that the performance of auto- matic annotation of subjectivity at the word level can be hurt by …

Content truncated to 500 words