1. Introduction

Text summarization is to generate a concise summary of documents to help a user quickly digest information and has become increasingly important due to the explosive growth of information. Most existing summarization methods can only generate an unstructured summary with a simple list of sentences. In many applications, however, we want to generate a (more structured) multi-faceted comparative summary, in which sentences are grouped into multiple facets and compared across different views. For example, a summary about laptop opinions may group sentences into facets such as "battery life" and "memory" and separate sentences with positive and negative opinions in each facet.

This project aims to systematically study this new summarization problem (called multi-faceted comparative summarization). We will develop general probabilistic approaches that can be applied to multiple instances of the problem in different domains. The basic idea is to use probabilistic mixture models to model and extract the multiple facets and multiple views of each facet in a set of text documents to be summarized. The extracted facets and views are then used to generate facet labels and select sentences for different facet-view combinations.

2. Team members

• ChengXiang Zhai (Professor)
• Hyun Duk Kim (Ph.D. student)
• Yue Lu (Ph.D. student)
• Qiaozhu Mei (Ph.D. student)
• Duo Zhang (Ph.D. student)

• Opinion integration as comparative summarization: Integrating and summarizing all the opinions expressed about an entity (e.g., a product or a person) are necessary to help people digest opinions. Opinions are generally expressed either in a well-written review article (e.g., a product review) or fragmentally in many other articles (e.g., many blog articles). We framed the problem of opinion integration as a comparative summarization problem and developed general probabilistic approaches to integrate scattered opinions with a well-written review article, showing promising results. [Lu & Zhai WWW 08].
• Combining social network analysis with topic modeling: Social networks can be leveraged to better infer topics in text. We have developed a regularized probabilistic latent semantic analysis model (NetPLSA) that can be applied to combine social network analysis with probabilistic topic modeling in a general way. Experiment results show that social networks can indeed improve the quality of the extracted topic models. [Mei et al. WWW 08].
• Rated aspect summarization: Web 2.0 applications have lead to the dramatic growth of user-generated content. We studied how to summarize a set of short comments on an entity and defined a novel rated aspect summarization problem where we generate a summary with multiple aspects (facets) together with a rating for each aspect. Experiment results on data from eBay show promising results. [Yue et al. WWW 09].
• Comparative summary of contradictory opinions: Opinions about a topic are often mixed with both positive and negative comments. In order to help users digest such mixed/contradictory opinions, we defined a novel comparative summarization problem where we aim at comparing opinions about the same but with different sentiment polarities. We proposed an optimization framework and heuristic algorithms for solving this problem. Experiment results on product review data show that the proposed algorithms are effective for generating a comparative summary of mixed opinions [Kim & Zhai 09].

4. Selected Publications (See all publications)

---

Text Information Management Group Home | About| People | Projects| Publications | Downloads| Links| Wiki