Learning with Compositional Semantics as Structural Inference for Subsentential Sentiment Analysis

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 793–801, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Learning with Compositional Semantics as Structural Inference for Subsentential Sentiment Analysis Yejin Choi and Claire Cardie Department of Computer Science Cornell University Ithaca, NY 14853 {ychoi,cardie}@cs.cornell.edu Abstract Determining the polarity of a sentiment- bearing expression requires more than a sim- ple bag-of-words approach. In particular, words or constituents within the expression can interact with each other to yield a particu- lar overall polarity. In this paper, we view such subsentential interactions in light of composi- tional semantics, and present a novel learning- based approach that incorporates structural in- ference motivated by compositional seman- tics into the learning procedure. Our exper- iments show that (1) simple heuristics based on compositional semantics can perform bet- ter than learning-based methods that do not in- corporate compositional semantics (accuracy of 89.7% vs. 89.1%), but (2) a method that integrates compositional semantics into learn- ing performs better than all other alterna- tives (90.7%). We also find that “content- word negators”, not widely employed in pre- vious work, play an important role in de- termining expression-level polarity. Finally, in contrast to conventional wisdom, we find that expression-level classification accuracy uniformly decreases as additional, potentially disambiguating, context is considered. 1 Introduction Determining the polarity of sentiment-bearing ex- pressions at or below the sentence level requires more than a simple bag-of-words approach. One of the difficulties is that words or constituents within the expression can interact with each other to yield a particular overall polarity. To facilitate our discus- sion, consider the following examples: 1: [I did [not]¬ have any [doubt]− about it.]+ 2: [The report [eliminated]¬ my [doubt]−.]+ 3: [They could [not]¬ [eliminate]¬ my [doubt]−.]− In the first example, “doubt” in isolation carries a negative sentiment, but the overall polarity of the sentence is positive because there is a negator “not”, which flips the polarity. In the second example, both “eliminated” and “doubt” carry negative sentiment in isolation, but the overall polarity of the sentence is positive because “eliminated” acts as a negator for its argument “doubt”. In the last example, there are effectively two negators – “not” and “eliminated” – which reverse the polarity of “doubt” twice, result- ing in the negative polarity for the overall sentence. These examples demonstrate that words or con- stituents interact with each other to yield the expression-level polarity. And a system that sim- ply takes the majority vote of the polarity of indi- vidual words will not work well on the above exam- ples. Indeed, much of the previous learning-based research on this topic tries to incorporate salient in- teractions by encoding them as features. One ap- proach includes features based on contextual va- lence shifters1 (Polanyi and Zaenen, 2004), which are words that affect the polarity or intensity of sen- timent over neighboring text spans (e.g., Kennedy and Inkpen (2005), Wilson et al. (2005), Shaikh et al. (2007)). Another approach encodes frequent sub- sentential patterns (e.g., McDonald …

Content truncated to 500 words