Matching Reviews to Objects using a Language Model

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 609–618, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Matching Reviews to Objects using a Language Model Nilesh Dalvi Ravi Kumar Bo Pang Andrew Tomkins Yahoo! Research 701 First Ave Sunnyvale, CA 94089 {ndalvi,ravikumar,bopang,atomkins}@yahoo-inc.com Abstract We develop a general method to match un- structured text reviews to a structured list of objects. For this, we propose a lan- guage model for generating reviews that incorporates a description of objects and a generic review language model. This mix- ture model gives us a principled method to find, given a review, the object most likely to be the topic of the review. Extensive experiments and analysis on reviews from Yelp show that our language model-based method vastly outperforms traditional tf- idf-based methods. 1 Introduction Consider a user searching for reviews of “Casablanca Moroccan Restaurant.” The search engine would like to obtain as many reviews of this restaurant as possible, both to offer a high- quality result set for even obscure restaurants, and to enable advanced applications such as aggrega- tion/summarization/categorization of reviews and recommendation of alternate restaurants. To solve this problem, it faces two high-level challenges: first, identify the restaurant review pages on the Web; and second, given a review, identify the restaurant that is being reviewed. There has been previous work addressing the first challenge (Sec- tion 2). We focus in this paper on the second. The Web is replete with restaurant reviews available on top restaurant verticals such as Yelp and CitySearch, as well as newspaper articles, newsgroup discussions, blog posts, small local re- view aggregators and so forth. Ideally, the search engine would like to obtain reviews from all pos- sible sources. While identifying the subject mat- ter of a given review on the large sites may be amenable to structured extraction through wrapper induction or related techniques, it is typically not cost-effective to apply such techniques to smaller “tail” sites, and purely unstructured sources re- quire alternate approaches altogether. In this pa- per, we explore the setting of matching reviews to objects using only their textual content. Note that matching reviews to objects is a pervasive prob- lem beyond the restaurant domain. Shopping ver- ticals like to aggregate camera reviews, entertain- ment verticals wish to collect movie reviews, and so on. We use restaurant reviews as a running ex- ample, but the techniques are general. More specifically, the problem we consider in this paper is the following. Given a list of struc- tured objects (restaurants/cameras/movies) and a text review, identify the object from the list that is the topic of the review. Our focus on tex- tual content allows us to expand the universe of sources from which we can extract reviews to in- clude sources that are purely textual, such as fo- rum posts, blog posts, newsgroup postings, and the like. In fact, even among collections of “struc- tured” sources like review aggregators, there are no highly accurate unsupervised techniques to match a known …

Content truncated to 500 words