An Unsupervised Aspect-Sentiment Model for Online Reviews

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 804–812, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics An Unsupervised Aspect-Sentiment Model for Online Reviews Samuel Brody Dept. of Biomedical Informatics Columbia University samuel.brody@dbmi.columbia.edu Noemie Elhadad Dept. of Biomedical Informatics Columbia University noemie@dbmi.columbia.edu Abstract With the increase in popularity of online re- view sites comes a corresponding need for tools capable of extracting the information most important to the user from the plain text data. Due to the diversity in products and ser- vices being reviewed, supervised methods are often not practical. We present an unsuper- vised system for extracting aspects and deter- mining sentiment in review text. The method is simple and flexible with regard to domain and language, and takes into account the in- fluence of aspect on sentiment polarity, an is- sue largely ignored in previous literature. We demonstrate its effectiveness on both compo- nent tasks, where it achieves similar results to more complex semi-supervised methods that are restricted by their reliance on manual an- notation and extensive knowledge sources. 1 Introduction Online review sites continue to grow in popularity as more people seek the advice of fellow users regard- ing services and products. Unfortunately, users are often forced to wade through large quantities of writ- ten data in order to find the information they want. This has led to an increase in research in the areas of opinion mining and sentiment analysis, with the aim of providing systems that can automatically an- alyze user reviews and extract the information most relevant to the user. One example of such an application is generat- ing a summary of the important factors mentioned in the reviews of a product (see Lerman et al. 2009). Another application is comparing two similar prod- ucts. In this case, it is important to present to the user the aspects in which the products differ, rather than just provide a general star rating. A third exam- ple is systems for generating automatic recommen- dations, based on similarity between products, user reviews, and history of previous purchases. These types of application require an underlying frame- work to identify the important aspects of the prod- uct (also known as features or attributes), and the sentiment expressed by the review writer. Unsupervised Methods are desirable for this task, for two reasons. First, due to the wide range and va- riety of products and services being reviewed, the framework must be robust and easily transferable between domains. The second reason is the nature of the data. Online reviews are often short and unstruc- tured, and may contain many spelling and gram- matical errors, as well as slang or specialized jar- gon. These factors often present a problem to meth- ods relying exclusively on dictionaries, manually- constructed knowledge resources, and gazetteers, as they may miss out on an important aspect of the product or an indicator of sentiment. Unsupervised methods, on the other hand, are not influenced by the lexical …

Content truncated to 500 words