Approaches to Large-Scale User Opinion Summarization for the Web

Introduction

The research literature in automatic text summarization is extensive and spans nearly seven decades. The focus in this research has traditionally centered most strongly on summarization of news articles (Gong & Liu, 2001; Nenkova, et al., 2006; Haghighi & Vanderwende, 2009; Nenkova & McKeown, 2011), but in recent years has expanded in diverse directions to, most notably, opinion summarization (Hu & Liu, 2004; Blair-Goldensohn, et al., 2008; Lerman, et al., 2009; Lu, et al., 2009). Opinion summarization is of immense interest to corporations and governments, who can act more effectively when they understand the viewpoints of their customers or citizens that are being laid bare each day in Internet forums and social networks, but are so disparate and numerous that they cannot be efficiently digested without some service in between the raw data and the ultimate consumer.

Part of the impetus of the disparate evolution in text summarization’s domain can be attributed to its application to real-world problems. While news summarization is interesting, in practice it is often unnecessary due to the concise and structured nature of newswire writing and the common existence of an already-included high quality succinct summary: the headline. Further, news summarization (as far as the task has commonly been defined in the literature) is arguably not a “real-world problem”; the datasets are small, the payoff is low, and the task is generally accomplished easily by humans.

In contrast, the areas that summarization research is expanding to are primarily those that can be considered real-world problems. Here, the payoffs are high, the datasets are often huge (making the problems interesting by virtue of fitting into the in vogue area of “big data'”) and the tasks are those that, rather than simplifying or optimizing the efficiency of an existing practice, could not be done without the help of algorithmic – often machine learning based – approaches. The principal example is opinion summarization (Liu & Zhang, 2012). This broad field encompasses disparate research tasks that include, inter alia, topic modeling, sentence scoring, clustering, sentiment analysis, subjectivity prediction, and user modeling. Building a powerful and mature system for real-world use in this area necessitates expertise in at least a large subset of all of these areas, plus the experience and knowledge required to make them work together synoptically.

What it does not necessarily require, however, is “advanced” or overly complicated models. Peter Norvig and others have argued convincingly that in general more data with simple models beats less data with complex models (Halevy, et al., 2009). While Norvig’s statement was with respect to machine learning in general, and in particular to cases where large amounts of data are available, the sentiment at least appears to be particularly apt with respect to text summarization. Models that have existed for several years tend to perform on par with more recent complicated methods and the simpler models are often easier to implement and scale much better. When these simple methods are coupled with domain-specific targeted assumptions about the structure of the input and the types of meta information that are available, impressive results can be easily achieved. In this Chapter, we explain powerful yet simple techniques for opinion summarization that build on these ideas.