Supervised and Unsupervised Methods in Employing Discourse Relations for Improving Opinion Polarity Classification

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 170–179, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Supervised and Unsupervised Methods in Employing Discourse Relations for Improving Opinion Polarity Classification Swapna Somasundaran Univ. of Pittsburgh Pittsburgh, PA 15260 swapna@cs.pitt.edu Galileo Namata Univ. of Maryland College Park, MD 20742 namatag@cs.umd.edu Janyce Wiebe Univ. of Pittsburgh Pittsburgh, PA 15260 wiebe@cs.pitt.edu Lise Getoor Univ. of Maryland College Park, MD 20742 getoor@cs.umd.edu Abstract This work investigates design choices in modeling a discourse scheme for im- proving opinion polarity classification. For this, two diverse global inference paradigms are used: a supervised collec- tive classification framework and an un- supervised optimization framework. Both approaches perform substantially better than baseline approaches, establishing the efficacy of the methods and the underlying discourse scheme. We also present quan- titative and qualitative analyses showing how the improvements are achieved. 1 Introduction The importance of discourse in opinion analy- sis is being increasingly recognized (Polanyi and Zaenen, 2006). Motivated by the need to en- able discourse-based opinion analysis, previous research (Asher et al., 2008; Somasundaran et al., 2008) developed discourse schemes and created manually annotated corpora. However, it was not known whether and how well these linguistic ideas and schemes can be translated into effective com- putational implementations. In this paper, we first investigate ways in which an opinion discourse scheme can be computation- ally modeled, and then how it can be utilized to improve polarity classification. Specifically, the discourse scheme we use is from Somasundaran et al. (2008), which was developed to support a global, interdependent polarity interpretation. To achieve discourse-based global inference, we ex- plore two different frameworks. The first is a supervised framework that learns interdependent opinion interpretations from training data. The second is an unsupervised optimization frame- work which uses constraints to express the ideas of coherent opinion interpretation embodied in the scheme. For the supervised framework, we use It- erative Collective Classification (ICA), which fa- cilitates machine learning using relational infor- mation. The unsupervised optimization is imple- mented as an Integer Linear Programming (ILP) problem. Via our implementations, we aim to empirically test if discourse-based approaches to opinion analysis are useful. Our results show that both of our implemen- tations achieve significantly better accuracies in polarity classification than classifiers using local information alone. This confirms the hypothesis that the discourse-based scheme is useful, and also shows that both of our design choices are effective. We also find that there is a difference in the way ICA and ILP achieve improvements, and a simple hybrid approach, which incorporates the strengths of both, is able to achieve significant overall im- provements over both. Our analyses show that even when our discourse-based methods bootstrap from noisy classifications, they can achieve good improvements. The rest of this paper is organized as follows: we discuss related work in Section 2 and the discourse scheme in Section 3. We present our discourse-based implementations in Section 4, ex- periments in Section 5, discussions in Section 6 and conclusions …

Content truncated to 500 words