Jointly Modeling Aspects and Opinions with a MaxEnt-LDA Hybrid

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 56–65, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Jointly Modeling Aspects and Opinions with a MaxEnt-LDA Hybrid Wayne Xin Zhao†, Jing Jiang‡, Hongfei Yan†, Xiaoming Li† †School of Electronics Engineering and Computer Science, Peking University, China ‡School of Information Systems, Singapore Management University, Singapore {zhaoxin,yhf}@net.pku.edu.cn, jingjiang@smu.edu.cn, lxm@pku.edu.cn Abstract Discovering and summarizing opinions from online reviews is an important and challeng- ing task. A commonly-adopted framework generates structured review summaries with aspects and opinions. Recently topic mod- els have been used to identify meaningful re- view aspects, but existing topic models do not identify aspect-specific opinion words. In this paper, we propose a MaxEnt-LDA hy- brid model to jointly discover both aspects and aspect-specific opinion words. We show that with a relatively small amount of train- ing data, our model can effectively identify as- pect and opinion words simultaneously. We also demonstrate the domain adaptability of our model. 1 Introduction With the dramatic growth of opinionated user- generated content, consumers often turn to online product reviews to seek advice while companies see reviews as a valuable source of consumer feedback. How to automatically understand, extract and sum- marize the opinions expressed in online reviews has therefore become an important research topic and gained much attention in recent years (Pang and Lee, 2008). A wide spectrum of tasks have been studied under review mining, ranging from coarse-grained document-level polarity classification (Pang et al., 2002) to fine-grained extraction of opinion expres- sions and their targets (Wu et al., 2009). In partic- ular, a general framework of summarizing reviews of a certain product is to first identify different as- pects (a.k.a. features) of the given product and then extract specific opinion expressions for each aspect. For example, aspects of a restaurant may include food, staff, ambience and price, and opinion expres- sions for staff may include friendly, rude, etc. Be- cause of the practicality of this structured summary format, it has been adopted in several previous stud- ies (Hu and Liu, 2004; Popescu and Etzioni, 2005; Brody and Elhadad, 2010) as well as some commer- cial systems, e.g. the “scorecard” feature at Bing shopping1. Different approaches have been proposed to iden- tify aspect words and phrases from reviews. Previ- ous methods using frequent itemset mining (Hu and Liu, 2004) or supervised learning (Jin and Ho, 2009; Jin et al., 2009; Wu et al., 2009) have the limitation that they do not group semantically related aspect expressions together. Supervised learning also suf- fers from its heavy dependence on training data. In contrast, unsupervised, knowledge-lean topic mod- eling approach has been shown to be effective in au- tomatically identifying aspects and their representa- tive words (Titov and McDonald, 2008; Brody and Elhadad, 2010). For example, words such as waiter, waitress, staff and service are grouped into one as- pect. We follow this promising direction and extend ex- isting topic models to jointly identify both aspect and opinion words, especially …

Content truncated to 500 words