Joint Inference for Fine-grained Opinion Extraction

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1640–1649, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Joint Inference for Fine-grained Opinion Extraction 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 This paper addresses the task of fine- grained opinion extraction – the identi- fication of opinion-related entities: the opinion expressions, the opinion hold- ers, and the targets of the opinions, and the relations between opinion expressions and their targets and holders. Most ex- isting approaches tackle the extraction of opinion entities and opinion relations in a pipelined manner, where the inter- dependencies among different extraction stages are not captured. We propose a joint inference model that leverages knowledge from predictors that optimize subtasks of opinion extraction, and seeks a glob- ally optimal solution. Experimental re- sults demonstrate that our joint inference approach significantly outperforms tradi- tional pipeline methods and baselines that tackle subtasks in isolation for the problem of opinion extraction. 1 Introduction Fine-grained opinion analysis is concerned with identifying opinions in text at the expression level; this includes identifying the subjective (i.e., opin- ion) expression itself, the opinion holder and the target of the opinion (Wiebe et al., 2005). The task has received increasing attention as many nat- ural language processing applications would ben- efit from the ability to identify text spans that cor- respond to these key components of opinions. In question-answering systems, for example, users may submit questions in the form “What does en- tity A think about target B?”; opinion-oriented summarization systems also need to recognize opinions and their targets and holders. In this paper, we address the task of identifying opinion-related entities and opinion relations. We consider three types of opinion entities: opinion expressions or direct subjective expressions as de- fined in Wiebe et al. (2005) — expressions that ex- plicitly indicate emotions, sentiment, opinions or other private states (Quirk et al., 1985) or speech events expressing private states; opinion targets — expressions that indicate what the opinion is about; and opinion holders — mentions of whom or what the opinion is from. Consider the follow- ing examples in which opinion expressions (O) are underlined and targets (T) and holders (H) of the opinion are bracketed. S1: [The workers][H1,2] were irked[O1] by [the government report][T1] and were worried[O2] as they went about their daily chores. S2: From the very start it could be predicted[O1] that on the subject of economic globalization, [the developed states][T1,2] were going to come across fierce opposition[O2]. The numeric subscripts denote linking relations, one of IS-ABOUT or IS-FROM. In S1, for in- stance, opinion expression “were irked” (O1) IS- ABOUT “the government report” (T1). Note that the IS-ABOUT relation can contain an empty tar- get (e.g. “were worried” in S1); similarly for IS- FROM w.r.t. the opinion holder (e.g. “predicted” in S2). We also allow an opinion entity to be involved in multiple relations (e.g. “the developed states” in S2). Not surprisingly, …

Content truncated to 500 words