Iterative Constrained Clustering for Subjectivity Word Sense Disambiguation

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 269–278, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Iterative Constrained Clustering for Subjectivity Word Sense Disambiguation Cem Akkaya, Janyce Wiebe University of Pittsburgh Pittsburgh PA, 15260, USA {cem,wiebe}@cs.pitt.edu Rada Mihalcea University of North Texas Denton TX, 76207, USA rada@cs.unt.edu Abstract Subjectivity word sense disambiguation (SWSD) is a supervised and application- specific word sense disambiguation task disambiguating between subjective and objective senses of a word. Not sur- prisingly, SWSD suffers from the knowl- edge acquisition bottleneck. In this work, we use a “cluster and label” strategy to generate labeled data for SWSD semi- automatically. We define a new algo- rithm called Iterative Constrained Cluster- ing (ICC) to improve the clustering purity and, as a result, the quality of the gener- ated data. Our experiments show that the SWSD classifiers trained on the ICC gen- erated data by requiring only 59% of the labels can achieve the same performance as the classifiers trained on the full dataset. 1 Introduction Subjectivity lexicons (e.g., (Turney, 2002; Whitelaw et al., 2005; Riloff and Wiebe, 2003; Yu and Hatzivassiloglou, 2003; Kim and Hovy, 2004; Bloom et al., 2007; Andreevskaia and Bergler, 2008; Agarwal et al., 2009)) play an important role in opinion, sentiment, and subjectivity analysis. These systems typically look for the presence of clues in text. Recently, in (Akkaya et al., 2009), we showed that subjectivity clues are fairly ambiguous as to whether they express subjectivity or not – words in such lexicons may have both subjective and objective usages. We call this problem subjectivity sense ambiguity. Consider the following sentence containing the clue “attack”: (1) He was attacked by Milosevic for at- tempting to carve out a new party from the Socialists. Knowing that “attack” is a subjectivity clue with negative polarity will help a system recognize the negative sentiment in the sentence. But for (2), the same information is simply misleading, because the clue is used with an objective meaning. (2) A new treatment based on training T-cells to attack cancerous cells ... Any opinion analysis system which relies on a subjectivity lexicon will be misled by subjectiv- ity clues used with objective senses (false hits). In (Akkaya et al., 2009), we introduced the task, Subjectivity Word Sense Disambiguation, which is to automatically determine which word instances in a corpus are being used with subjective senses, and which are being used with objective senses. SWSD can be considered as a coarse-grained and application-specific word sense disambigua- tion task. We showed that sense subjectivity in- formation about clues can be fed to subjectiv- ity and sentiment analysis resulting in substantial improvement for both subjectivity and sentiment analysis by avoiding false hits. Although SWSD is a promising tool, it suf- fers from the knowledge acquisition bottleneck. SWSD is defined as a supervised task, and fol- lows a targeted approach common in the WSD lit- erature for performance reasons. This means, for each target clue, a different …

Content truncated to 500 words