Modeling Review Argumentation for Robust Sentiment Analysis

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 553–564, Dublin, Ireland, August 23-29 2014. Modeling Review Argumentation for Robust Sentiment Analysis Henning Wachsmuth Universit¨at Paderborn s-lab – Software Quality Lab Paderborn, Germany henningw@upb.de Martin Trenkmann, Benno Stein Bauhaus-Universit¨at Weimar Webis Group Weimar, Germany <1st>.<last>@uni-weimar.de Gregor Engels Universit¨at Paderborn s-lab – Software Quality Lab Paderborn, Germany engels@upb.de Abstract Most text classification approaches model text at the lexical and syntactic level only, lacking do- main robustness and explainability. In tasks like sentiment analysis, such approaches can result in limited effectiveness if the texts to be classified consist of a series of arguments. In this paper, we claim that even a shallow model of the argumentation of a text allows for an effective and more robust classification, while providing intuitive explanations of the classification results. Here, we apply this idea to the supervised prediction of sentiment scores for reviews. We combine existing approaches from sentiment analysis with novel features that compare the overall argumentation structure of the given review text to a learned set of common sentiment flow patterns. Our evalu- ation in two domains demonstrates the benefit of modeling argumentation for text classification in terms of effectiveness and robustness. 1 Introduction Text classification is a key technique in natural language processing and information retrieval that is ap- plied for several tasks. Standard classification approaches map a text to a vector of lexical and shallow syntactic surface-level features, from which class information is inferred using supervised learning (Man- ning et al., 2008). Even though the results of such approaches can hardly be explained, they have proven effective for narrow-domain texts with explicit class information (Joachims, 2001; Pang et al., 2002). However, surface-level features often do not help to classify out-of-domain texts correctly, because they tend to model the domain of the texts and not the classes to be inferred, as we observe in (Wachsmuth and Bujna, 2011) among others. Moreover, they are likely to fail on texts where the class information is implicitly represented by the argumentation of the writer. Such texts are in the focus of popular tasks like authorship attribution, automatic essay grading, and, above all, sentiment analysis. As an example, consider the short hotel review at the top and bottom of Figure 1. It contains more positive than negative statements. Hence, a surface-level analysis would probably classify the review to have a positive overall sentiment polarity. In fact, the argumentation of the review text reveals a clear negative sentiment. The analysis of argumentation is recently getting more attention (cf. Section 2 for details). With respect to sentiment, related approaches analyze discourse relations (Mukherjee and Bhattacharyya, 2012), iden- tify the different aspects mentioned in a text (Lazaridou et al., 2013), or the like. While these approaches can infer implicit class information from argumentative texts like reviews, they do not address the domain dependency problem of sentiment analysis (Wu et al., 2010). In addition, they still lack explainability, which limits end user acceptance in case of …

Content truncated to 500 words