For a few dollars less: Identifying review pages sans human labels

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 494–502, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics For a few dollars less: Identifying review pages sans human labels Luciano Barbosa Dept. of Computer Science University of Utah Salt Lake City, UT 84112, USA. lbarbosa@cs.utah.edu Ravi Kumar Bo Pang Yahoo! ResearchAndrew Tomkins 701 First Ave Sunnyvale, CA 94089, USA. {ravikumar,bopang,atomkins}@yahoo-inc.com Abstract We address the problem of large-scale auto- matic detection of online reviews without us- ing any human labels. We propose an efficient method that combines two basic ideas: Build- ing a classifier from a large number of noisy examples and using the structure of the web- site to enhance the performance of this classi- fier. Experiments suggest that our method is competitive against supervised learning meth- ods that mandate expensive human effort. 1 Introduction Shoppers are migrating to the web and online re- views are playing a critical role in affecting their shopping decisions, online and offline. According to two surveys published by comScore (2007) and Horrigan (2008), 81% of web users have done on- line research on a product at least once. Among readers of online reviews, more than 70% reported that the reviews had a significant influence on their purchases. Realizing this economic potential, search engines have been scrambling to cater to such user needs in innovative ways. For example, in response to a product-related query, a search engine might want to surface only review pages, perhaps via a “fil- ter by” option, to the user. More ambitiously, they might want to dissect the reviews, segregate them into novice and expert judgments, distill sentiments, and present an aggregated “wisdom of the crowds” opinion to the user. Identifying review pages is the indispensable enabler to fulfill any such ambition; nonetheless, this problem does not seem to have been addressed at web scale before. Detecting review webpages in a few, review-only websites is an easy, manually-doable task. A large fraction of the interesting review content, however, is present on pages outside such websites. This is where the task becomes challenging. Review pages might constitute a minority and can be buried in a multitude of ways among non-review pages — for instance, the movie review pages in nytimes. com, which are scattered among all news articles, or the product review pages in amazon.com, which are accessible from the product description page. An automatic and scalable method to identify reviews is thus a practical necessity for the next-generation search engines. The problem is actually more gen- eral than detecting reviews: it applies to detecting any “horizontal” category such as buying guides, fo- rums, discussion boards, FAQs, etc. Given the nature of these problems, it is tempt- ing to use supervised classification. A formidable barrier is the labeling task itself since human la- bels need time and money. On the other hand, it is easier to generate an enormous number of low- quality labeled examples through purely automatic methods. This prompts the question: Can …

Content truncated to 500 words