A Joint Model of Text and Aspect Ratings for Sentiment Summarization

Proceedings of ACL-08: HLT, pages 308–316, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics A Joint Model of Text and Aspect Ratings for Sentiment Summarization Ivan Titov Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 titov@uiuc.edu Ryan McDonald Google Inc. 76 Ninth Avenue New York, NY 10011 ryanmcd@google.com Abstract Online reviews are often accompanied with numerical ratings provided by users for a set of service or product aspects. We propose a statistical model which is able to discover corresponding topics in text and extract tex- tual evidence from reviews supporting each of these aspect ratings – a fundamental problem in aspect-based sentiment summarization (Hu and Liu, 2004a). Our model achieves high ac- curacy, without any explicitly labeled data ex- cept the user provided opinion ratings. The proposed approach is general and can be used for segmentation in other applications where sequential data is accompanied with corre- lated signals. 1 Introduction User generated content represents a unique source of information in which user interface tools have facil- itated the creation of an abundance of labeled con- tent, e.g., topics in blogs, numerical product and ser- vice ratings in user reviews, and helpfulness rank- ings in online discussion forums. Many previous studies on user generated content have attempted to predict these labels automatically from the associ- ated text. However, these labels are often present in the data already, which opens another interesting line of research: designing models leveraging these labelings to improve a wide variety of applications. In this study, we look at the problem of aspect- based sentiment summarization (Hu and Liu, 2004a; Popescu and Etzioni, 2005; Gamon et al., 2005; Nikos’ Fine Dining Food 4/5 “Best fish in the city”, “Excellent appetizers” Decor 3/5 “Cozy with an old world feel”, “Too dark” Service 1/5 “Our waitress was rude”, “Awful service” Value 5/5 “Good Greek food for the $”, “Great price!” Figure 1: An example aspect-based summary. Carenini et al., 2006; Zhuang et al., 2006).1 An aspect-based summarization system takes as input a set of user reviews for a specific product or ser- vice and produces a set of relevant aspects, the ag- gregated sentiment for each aspect, and supporting textual evidence. For example, figure 1 summarizes a restaurant using aspects food, decor, service, and value plus a numeric rating out of 5. Standard aspect-based summarization consists of two problems. The first is aspect identification and mention extraction. Here the goal is to find the set of relevant aspects for a rated entity and extract all textual mentions that are associated with each. As- pects can be fine-grained, e.g., fish, lamb, calamari, or coarse-grained, e.g., food, decor, service. Sim- ilarly, extracted text can range from a single word to phrases and sentences. The second problem is sentiment classification. Once all the relevant as- pects and associated pieces of texts are extracted, the system should aggregate sentiment over each as- pect to provide the user with an average numeric or symbolic rating. Sentiment classification is a …

Content truncated to 500 words