A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1630–1639, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations Angeliki Lazaridou University of Trento angeliki.lazaridou@unitn.it Ivan Titov Saarland University titov@mmci.uni-saarland.de Caroline Sporleder Trier University csporled@coli.uni-sb.de Abstract We propose a joint model for unsuper- vised induction of sentiment, aspect and discourse information and show that by in- corporating a notion of latent discourse re- lations in the model, we improve the pre- diction accuracy for aspect and sentiment polarity on the sub-sentential level. We deviate from the traditional view of dis- course, as we induce types of discourse re- lations and associated discourse cues rel- evant to the considered opinion analysis task; consequently, the induced discourse relations play the role of opinion and as- pect shifters. The quantitative analysis that we conducted indicated that the integra- tion of a discourse model increased the prediction accuracy results with respect to the discourse-agnostic approach and the qualitative analysis suggests that the in- duced representations encode a meaning- ful discourse structure. 1 Introduction With the rapid growth of the Web, it is becoming increasingly difficult to discern useful from irrel- evant information, particularly in user-generated content, such as product reviews. To make it easier for the reader to separate the wheat from the chaff, it is necessary to structure the available informa- tion. In the review domain, this is done in aspect- based sentiment analysis which aims at identify- ing text fragments in which opinions are expressed about ratable aspects of products, such as ‘room quality’ or ‘service quality’. Such fine-grained analysis can serve as the first step in aspect-based sentiment summarization (Hu and Liu, 2004), a task with many practical applications. Aspect-based summarization is an active re- search area for which various techniques have been developed, both statistical (Mei et al., 2007; Titov and McDonald, 2008b) and not (Hu and Liu, 2004), and relying on different types of supervi- sion sources, such as sentiment-annotated texts or polarity lexica (Turney and Littman, 2002). Most methods rely on local information (bag-of-words, short ngrams or elementary syntactic fragments) and do not attempt to account for more complex interactions. However, these local lexical repre- sentations by themselves are often not sufficient to infer a sentiment or aspect for a fragment of text. For instance, in the following example taken from a TripAdvisor1 review: Example 1. The room was nice but let’s not talk about the view. it is difficult to deduce on the basis of local lexical features alone that the opinion about the view is negative. The clause let’s not talk about the view could by itself be neutral or even positive given the right context (e.g., I’ve never seen such a fancy ho- tel room, my living room doesn’t look that cool... and let’s not talk about the view). However, the contrast relation signaled by the connective but makes it clear that the second clause has a …

Content truncated to 500 words