Extracting Opinion Expressions with semi-Markov Conditional Random Fields

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1335–1345, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Extracting Opinion Expressions with semi-Markov Conditional Random Fields Bishan Yang Department of Computer Science Cornell University bishan@cs.cornell.edu Claire Cardie Department of Computer Science Cornell University cardie@cs.cornell.edu Abstract Extracting opinion expressions from text is usually formulated as a token-level sequence labeling task tackled using Conditional Ran- dom Fields (CRFs). CRFs, however, do not readily model potentially useful segment-level information like syntactic constituent struc- ture. Thus, we propose a semi-CRF-based ap- proach to the task that can perform sequence labeling at the segment level. We extend the original semi-CRF model (Sarawagi and Co- hen, 2004) to allow the modeling of arbitrar- ily long expressions while accounting for their likely syntactic structure when modeling seg- ment boundaries. We evaluate performance on two opinion extraction tasks, and, in contrast to previous sequence labeling approaches to the task, explore the usefulness of segment- level syntactic parse features. Experimental results demonstrate that our approach outper- forms state-of-the-art methods for both opin- ion expression tasks. 1 Introduction Accurate opinion expression identification is crucial for tasks that benefit from fine-grained opinion anal- ysis (Wiebe et al., 2005): e.g., it is a first step in characterizing the sentiment and intensity of the opinion; it provides a textual anchor for identifying the opinion holder and the target or topic of an opin- ion; and these, in turn, form the basis of opinion- oriented question answering and opinion summa- rization systems. In this paper, we focus on opin- ion expressions as defined in Wiebe et al. (2005) — subjective expressions that denote emotions, senti- ment, beliefs, opinions, judgments, or other private states (Quirk et al., 1985) in text. These include direct subjective expressions (DSEs): explicit men- tions of private states or speech events expressing private states; and expressive subjective expressions (ESEs): expressions that indicate sentiment, emo- tion, etc. without explicitly conveying them. Follow- ing are two example sentences labeled with DSEs and ESEs. (1) The International Committee of the Red Cross, [as usual][ESE], [has refused to make any statements][DSE]. (2) The Chief Minister [said][DSE] that [the demon they have reared will eat up their own vitals][ESE]. As a type of information extraction task, opinion expression extraction has been successfully tackled in the past via sequence tagging methods: Choi et al. (2006) and Breck et al. (2007), for example, ap- ply conditional random fields (CRFs) (Lafferty et al., 2001) using sophisticated token-level features. In token-level sequence labeling, labels are assigned to single tokens, and the label of each token depends on the current token and the label of the previous to- ken (we consider the usual first-order assumption). Segment-based features — features that describe a set of related contiguous tokens, e.g., a phrase or constituent — might provide critical information for identifying opinion expressions; they cannot, how- ever, be readily and naturally represented in the CRF model. 1335 Our goal in …

Content truncated to 500 words