Syntactic and Semantic Structure for Opinion Expression Detection

Proceedings of the Fourteenth Conference on Computational Natural Language Learning, pages 67–76, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Syntactic and Semantic Structure for Opinion Expression Detection Richard Johansson and Alessandro Moschitti DISI, University of Trento Via Sommarive 14 Povo, 38123 Trento (TN), Italy {johansson, moschitti}@disi.unitn.it Abstract We demonstrate that relational features derived from dependency-syntactic and semantic role structures are useful for the task of detecting opinionated expressions in natural-language text, significantly im- proving over conventional models based on sequence labeling with local features. These features allow us to model the way opinionated expressions interact in a sen- tence over arbitrary distances. While the relational features make the pre- diction task more computationally expen- sive, we show that it can be tackled effec- tively by using a reranker. We evaluate a number of machine learning approaches for the reranker, and the best model re- sults in a 10-point absolute improvement in soft recall on the MPQA corpus, while decreasing precision only slightly. 1 Introduction The automatic detection and analysis of opinion- ated text – subjectivity analysis – is potentially useful for a number of natural language processing tasks. Examples include retrieval systems answer- ing queries about how a particular person feels about a product or political question, and various types of market analysis tools such as review min- ing systems. A primary task in subjectivity analysis is to mark up the opinionated expressions, i.e. the text snippets signaling the subjective content of the text. This is necessary for further analysis, such as the determination of opinion holder and the polarity of the opinion. The MPQA corpus (Wiebe et al., 2005), a widely used corpus anno- tated with subjectivity information, defines two types of subjective expressions: direct subjective expressions (DSEs), which are explicit mentions of opinion, and expressive subjective elements (ESEs), which signal the attitude of the speaker by the choice of words. DSEs are often verbs of statement and categorization, where the opinion and its holder tend to be direct semantic arguments of the verb. ESEs, on the other hand, are less easy to categorize syntactically; prototypical examples would include value-expressing adjectives such as beautiful, biased, etc. In addition to DSEs and ESEs, the MPQA corpus also contains annotation for non-subjective statements, which are referred to as objective speech events (OSEs). Examples (1) and (2) show two sentences from the MPQA corpus where DSEs and ESEs have been manually annotated. (1) For instance, he [denounced]DSE as a [human rights violation]ESE the banning and seizure of satellite dishes in Iran. (2) This [is viewed]DSE as the [main impediment]ESE to the establishment of po- litical order in the country . The task of marking up these expressions has usually been approached using straightforward sequence labeling techniques using simple fea- tures in a small contextual window (Choi et al., 2006; Breck et al., 2007). However, due to the simplicity of the feature sets, this approach fails to take into account the fact that the semantic and pragmatic interpretation of sentences …

Content truncated to 500 words