Summize Text Summarization Talk @ UMD

Today, Eric gave a talk on our summarization technology at the University of Maryland's cloud computing speaker series.
 

Above is Eric giving his talk: Yes, a few of us at Summize thought it was funny that he was wearing a tie. Below is an abstract of the talk.

Text Summarization of Review Sentiments

The Web provides a unique outlet for individuals to have their voices heard on a variety of different topics. Often this happens in the form of subjective opinions, with authors expressing strong sentiment either in favor or against a product, person, issue, etc. The number of such opinions on the web is growing at an impressive rate, by a few hundred thousand per week in the blogosphere alone, to say nothing of sites that focus on such content like Amazon.com's user reviews or subjective questions on Yahoo Answers. While sentiment analysis has recently gained much attention and review mining became an important topic of research, less emphasis has been put on efficient text summarization of opinion sentiments.

Rather, opinion mining has been narrowly defined as estimating the average polarity of sentiments expressed about various facets of the opinions' target. For products like consumer electronics with a finite set of features, this model enables relative comparison of targets within a given category. However, it is not well-suited to targets with less obvious facets (people, brands, etc.) and does not address the problem of textually summarizing the sentiments for an individual target.  For example, reviews of a band's latest album might overwhelmingly say it is "not as good as their first one", or that it is "uninspired". These explain the motivation behind the negative ratings for a particular target where the sentiments don't relate to obvious facets. I will describe an algorithm for textual summarization of review sentiments that scales to very large collections and detail its performance on a dataset comprising several million product reviews. In doing so, I will characterize the consensus building in this massive review pool, and discuss the potential for scaling this algorithm to all of the opinions on the web.

Comments

[sarah]

Need a close-up of the tie!