Automatic extraction of contextual valence shifters.

Proceedings of Recent Advances in Natural Language Processing, pages 98–104, Hissar, Bulgaria, 7-13 September 2013. Automatic Extraction of Contextual Valence Shifters Noémi Boubel Thomas François Hubert Naets Cental, ILC (Université catholique de Louvain) {noemi.boubel,thomas.francois,hubert.naets}@uclouvain.be Abstract In opinion mining, many linguistic struc- tures, called contextual valence shifters, may modify the prior polarity of items. Some systems of sentiment analysis have tried to take these shifters into account, but few studies have focused on the identifica- tion of all these structures and their impact on polarized words. In this paper, we describe a method that automatically identifies contextual valence shifters. It relies on a chi-square test ap- plied to the contingency table represent- ing the distribution of a candidate shifter in a corpus of reviews of various opinions. The system depends on two resources in French – a corpus of reviews and a lexicon of valence terms – to build a list of French contextual valence shifters. We also intro- duce a set of rules used to classify the ex- tracted contextual valence shifters accord- ing to their impact on polarized words. They make use of the Pearson residuals in contingency tables to filter candidate shifters and classify them. We show that the technique reaches an F-measure of ei- ther 0.56 or 0.66, depending on how the categories of shifters are defined. 1 Introduction and State of the Art Most opinion mining systems rely on the extrac- tion of sentiment words to detect opinions. These words, which we will rather refer to as polarized words, convey useful information about the se- mantic orientation (positive or negative) of a text. However, the context in which these words appear may modify their valence in many ways. Although being of importance, this issue has been investi- gated only recently and is now the object of an increasing attention. Polanyi and Zaenen (2004) first postulated the existence of contextual valence shifters, which are contextual phenomena altering the prior polarity of a term. Afterwards, some of these phenomena (such as negative or conditional syntactic struc- tures) were dealt with on a case by case basis (Das and Chen, 2001; Na et al., 2004; Popescu and Et- zioni, 2005; Pang et al., 2002; Wilson et al., 2005; Wilson et al., 2006; Councill et al., 2010). Studies addressing the phenomenon as a whole flourished later. They aimed at best modelling the expression of opinions (Polanyi and Zaenen, 2004; Taboada et al., 2011; Hatzivassiloglou and Wiebe, 2000; Morsy and Rafea, 2012; Musat and Trausan-Matu, 2010), before embedding those in a classification system. The main purposes of these studies are to determine a list of contextual valence shifters that impact the polarity of a term as well as to define the nature of this impact. However, these lists are often manually built from linguistic intuitions and not learned from language data. Works relying on a corpus of texts to develop resources that best re- flect the actual role played by the linguistic context for opinion mining are few. Li et al. …

Content truncated to 500 words