Towards Context-Based Subjectivity Analysis

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1180–1188, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Towards Context-Based Subjectivity Analysis Farah Benamara IRIT-CNRS Toulouse benamara@irit.fr Baptiste Chardon Synapse D´eveloppement Toulouse baptiste.chardon @synapse-fr.com Yannick Mathieu LLF-CNRS Paris yannick.mathieu@ linguist.jussieu.fr Vladimir Popescu IRIT-CNRS Toulouse popescu@irit.fr Abstract We propose a new subjectivity classi- fication at the segment level that is more appropriate for discourse-based sen- timent analysis. Our approach automati- cally distinguish between subjective non- evaluative and objective segments and be- tween implicit and explicit opinions, by using local and global context features. 1 Introduction Subjectivity and polarity classification is one of the most studied research area in opinion analy- sis (Pang and Lee, 2004; Wiebe and Riloff, 2005; Wilson et al., 2009). The first task generally dis- tinguish between objective and subjective state- ments. Polarity classification is then performed in order to extract positive, negative and possi- bly neutral statements. These two tasks are co- dependent, since subjectivity analysis filters out statements that contain no opinion. A common approach in these tasks is to rely on the prior polarity of words and expressions as encoded in external lexical resources. However, as (Polanyi and Zaenen, 2006) stated, identifying prior polarity alone may not suffice to improve sentiment analysis at a finer grain, we need both local and global context. Context provided locally can help in two ways. First, it can be used in sub- jectivity word sense disambiguation (SWSD) in order to determine if a given word has a subjective or an objective sense (Akkaya et al., 2009). It can also be used to identify valence shifters (viz. nega- tions, modalities and intensifiers) that strengthen, weaken or reverse the prior polarity of a word or an expression (Kennedy and Inkpen, 2006; Wilson et al., 2009). Global context on the other hand can be used to identify implicit opinions and to improve the recognition of the overall stance. Few research efforts have been undertaken on using discourse as features for sentence / clause- based opinion analysis. Among them, (Pang and Lee, 2004) assume that subjective and objec- tive sentences are more likely to appear together, (Asher et al., 2008) have developed an annota- tion schema for a fine-grained contextual opin- ion analysis using discourse relations, (Taboada et al., 2008) have used a Rhetorical Structure Theory discourse parser in order to calculate se- mantic orientation by weighting the nuclei more heavily, and finally, (Somasundaran, 2010) has proposed a discourse-level treatment to improve sentence-based polarity classification and to rec- ognize the overall stance. More recently, (Zhou et al., 2011) proposed an unsupervised method to recognize RST-based discourse relations for elim- inating intra-sentence polarity ambiguities. How- ever, no work has investigated so far how dis- course structure can be used to enhance subjectiv- ity analysis (SA). Using discourse for SA raises new issues: Is sentence/clause subjectivity-based analysis ap- propriate? Is binary subjective vs. objective clas- sification enough for capturing how opinions are expressed within discourse? and finally, how can rhetorical …

Content truncated to 500 words