Sentiment Flow - A General Model of Web Review Argumentation

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 601–611, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Sentiment Flow – A General Model of Web Review Argumentation Henning Wachsmuth and Johannes Kiesel and Benno Stein Faculty of Media, Bauhaus-Universität Weimar, Germany {henning.wachsmuth,johannes.kiesel,benno.stein}@uni-weimar.de Abstract Web reviews have been intensively studied in argumentation-related tasks such as sen- timent analysis. However, due to their fo- cus on content-based features, many sen- timent analysis approaches are effective only for reviews from those domains they have been specifically modeled for. This paper puts its focus on domain indepen- dence and asks whether a general model can be found for how people argue in web reviews. Our hypothesis is that people ex- press their global sentiment on a topic with similar sequences of local sentiment inde- pendent of the domain. We model such sentiment flow robustly under uncertainty through abstraction. To test our hypoth- esis, we predict global sentiment based on sentiment flow. In systematic experiments, we improve over the domain independence of strong baselines. Our findings suggest that sentiment flow qualifies as a general model of web review argumentation. 1 Introduction The web is full of user-generated reviews on prod- ucts, services, and works of art, like those from Amazon, TripAdvisor, and Rotten Tomatoes. Such web reviews provide facts, positive opinions, and negative opinions on different aspects. By that, the reviews express, implicitly or explicitly, an overall opinion on the topic in question. From an abstract viewpoint, the argumentation of a web review can thus be seen as a composition of local sentiments used to justify some global sentiment. Both local and global sentiment of reviews are in the focus of numerous sentiment analysis ap- proaches (cf. Section 2 for details). Many of these approaches model reviews primarily with content- based features, derived from the words in the re- views. The use of words, however, varies strongly across domains, as illustrated in Figure 1 for a product, a hotel, and a movie review. As a con- sequence, sentiment analysis suffers from domain dependence (Wu et al., 2010), i.e., high effective- ness is often achieved only in the domain an ap- proach has been specifically modeled for. To adapt to other domains, prior knowledge about these domains or about domain-independent features is needed (Prettenhofer and Stein, 2010). This paper considers the question as to whether the overall argumentation of web reviews can be modeled in a general way in order to increase do- main independence in sentiment analysis. We ob- serve that people structure web reviews largely sequentially—in contrast to the complex struc- tures of many other argumentative texts. While the reviewed aspects differ between domains, our as- sumption is that the overall argumentation of a web review is generally represented by a sequence of local sentiments, called the review’s sentiment flow (Mao and Lebanon, 2007). In particular, we hypothesize that, under an adequate model, similar sentiment flows express similar global sentiments, also across domains. All reviews in Figure …

Content truncated to 500 words