Summarization of Multiple User Reviews in the Restaurant Domain

We discuss an automatic review summarization system whose goal is to provide a set of useful fragments, called snippets, from a collection of user reviews about a restaurant. The goal is to produce a small summary per restaurant for use in a spoken dialog system. The system comprises three stages: 1) judging overall relevance of snippets, 2) categorizing snippets and, 3) selecting the most relevant snippets per category to eliminate redundancy. We provide about 6 to 12 snippets per restaurant. There are over 200k restaurants in our database, containing about 24M words of reviews written about them. In a user study, our snippet selection was substantially preferred over the baseline system of random snippet selection. The system is featured in the United States as a toll-free number 1-877-456-DATA.