Sentiment Analysis of Figurative Language using a Word Sense Disambiguation Approach

International Conference RANLP 2009 - Borovets, Bulgaria, pages 370–375 Sentiment Analysis of Figurative Language using a Word Sense Disambiguation Approach Vassiliki Rentoumi∗† vrentoumi@iit.demokritos.gr George Giannakopoulos∗† ggianna@iit.demokritos.gr Vangelis Karkaletsis∗ vangelis@iit.demokritos.gr George A. Vouros† georgev@aegean.gr Abstract In this paper we propose a methodology for sen- timent analysis of figurative language which ap- plies Word Sense Disambiguation and, through an n-gram graph based method, assigns polar- ity to word senses. Polarity assigned to senses, combined with contextual valence shifters, is exploited for further assigning polarity to sen- tences, using Hidden Markov Models. Evalua- tion results using the corpus of the Affective Text task of SemEval’07, are presented together with a comparison with other state-of-the-art meth- ods, showing that the proposed method provides promising results, and positive evidence support- ing our conjecture: figurative language conveys sentiment. 1 Introduction Metaphors and expansions are very common phenom- ena in everyday language. Exploring such cases, we aim at revealing new semantic aspects of figurative lan- guage. Detection of sentiments and implicit polarity, is within our focus. We thus propose a methodology for sentiment analysis that can be valuable in detect- ing new semantic aspects of language. Recent stud- ies have shown that subjectivity is a language prop- erty which is directly related to word senses [14]. We believe that subjectivity and thus, senses related to meanings that convey subjectivity, are valuable indica- tors of sentiment. In order to prove this, we are led to the exploration of non-literal senses. Work presented in [11] provides evidence that non-literal senses, such as metaphors and expanded senses, tend to indicate subjectivity, triggering polarity. In order to capture the polarity that figurative language conveys [12], it is necessary to resolve ambiguities and detect the cases where words have non-literal meaning. Detecting this property for words, we can further assign polarity to the enclosing context, as it is shown in [12]. Towards this goal, other elements in discourse such as valence shifters [9] affect evaluative terms such as figurative ex- pressions and modify the polarity of the whole context. In this paper we introduce a methodology for polar- ity detection that applies word sense disambiguation (WSD), exploits the assessed word senses and assigns ∗National Centre for Scientific Research ”Demokritos”, Inst. of Informatics and Telecommunications, Greece †University of the Aegean, Dpt. of Information and Com- munication System Enginneering, Greece polarity to the enclosing sentences, exploiting other contextual features as well. In Section 2 we briefly present related work while in Section 3 we present the theoretical background of this work, together with evidence for the conjectures made. Section 4 presents a detailed description of the overall methodology and of the specific methods used. Section 5 details the evaluation of the specific techniques used as well as the overall evaluation of the system. Section 6 concludes this article with a brief presentation of future research. 2 Related Work Sentiment analysis aims to determine the exact polar- ity of a subjective expression. Specifically in [3] there is an effort using semi-supervised machine learning methods to …

Content truncated to 500 words