Aspect Extraction with Automated Prior Knowledge Learning

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 347–358, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Aspect Extraction with Automated Prior Knowledge Learning Zhiyuan Chen Arjun Mukherjee Department of Computer ScienceBing Liu University of Illinois at Chicago Chicago, IL 60607, USA {czyuanacm,arjun4787}@gmail.com,liub@cs.uic.edu Abstract Aspect extraction is an important task in sentiment analysis. Topic modeling is a popular method for the task. However, unsupervised topic models often generate incoherent aspects. To address the is- sue, several knowledge-based models have been proposed to incorporate prior knowl- edge provided by the user to guide mod- eling. In this paper, we take a major step forward and show that in the big data era, without any user input, it is possi- ble to learn prior knowledge automatically from a large amount of review data avail- able on the Web. Such knowledge can then be used by a topic model to discover more coherent aspects. There are two key challenges: (1) learning quality knowl- edge from reviews of diverse domains, and (2) making the model fault-tolerant to handle possibly wrong knowledge. A novel approach is proposed to solve these problems. Experimental results using re- views from 36 domains show that the pro- posed approach achieves significant im- provements over state-of-the-art baselines. 1 Introduction Aspect extraction aims to extract target entities and their aspects (or attributes) that people have expressed opinions upon (Hu and Liu, 2004, Liu, 2012). For example, in “The voice is not clear,” the aspect term is “voice.” Aspect extraction has two subtasks: aspect term extraction and aspect term resolution. Aspect term resolution groups ex- tracted synonymous aspect terms together. For ex- ample, “voice” and “sound” should be grouped to- gether as they refer to the same aspect of phones. Recently, topic models have been extensively applied to aspect extraction because they can per- form both subtasks at the same time while other existing methods all need two separate steps (see Section 2). Traditional topic models such as LDA (Blei et al., 2003) and pLSA (Hofmann, 1999) are unsupervised methods for extracting la- tent topics in text documents. Topics are aspects in our task. Each aspect (or topic) is a distribution over (aspect) terms. However, researchers have shown that fully unsupervised models often pro- duce incoherent topics because the objective func- tions of topic models do not always correlate well with human judgments (Chang et al., 2009). To tackle the problem, several semi-supervised topic models, also called knowledge-based topic models, have been proposed. DF-LDA (Andrze- jewski et al., 2009) can incorporate two forms of prior knowledge from the user: must-links and cannot-links. A must-link implies that two terms (or words) should belong to the same topic whereas a cannot-link indicates that two terms should not be in the same topic. In a similar but more generic vein, must-sets and cannot-sets are used in MC-LDA (Chen et al., 2013b). Other re- lated works include (Andrzejewski et al., 2011, Chen et al., 2013a, Chen et al., …

Content truncated to 500 words