Towards a Unified Approach to Simultaneous Single-Document and Multi-Document Summarizations

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1137–1145, Beijing, August 2010 Towards a Unified Approach to Simultaneous Single-Document and Multi-Document Summarizations Xiaojun Wan Institute of Compute Science and Technology The MOE Key Laboratory of Computational Linguistics Peking University wanxiaojun@icst.pku.edu.cn Abstract Single-document summarization and multi- document summarization are very closely re- lated tasks and they have been widely investi- gated independently. This paper examines the mutual influences between the two tasks and proposes a novel unified approach to si- multaneous single-document and multi- document summarizations. The mutual influ- ences between the two tasks are incorporated into a graph model and the ranking scores of a sentence for the two tasks can be obtained in a unified ranking process. Experimental re- sults on the benchmark DUC datasets demon- strate the effectiveness of the proposed ap- proach for both single-document and multi- document summarizations. 1 Introduction Single-document summarization aims to pro- duce a concise and fluent summary for a single document, and multi-document summarization aims to produce a concise and fluent summary for a document set consisting of multiple related documents. The two tasks are very closely re- lated in both task definition and solution method. Moreover, both of them are very important in many information systems and applications. For example, given a cluster of news articles, a multi-document summary can be used to help users to understand the whole cluster, and a sin- gle summary for each article can be used to help users to know the content of the specified article. To date, single-document and multi-document summarizations have been investigated exten- sively and independently in the NLP and IR fields. A series of special conferences or work- shops on automatic text summarization (e.g. SUMMAC, DUC, NTCIR and TAC) have ad- vanced the technology and produced a couple of experimental online systems. However, the two summarization tasks have not yet been simulta- neously investigated in a unified framework. Inspired by the fact that the two tasks are very closely related and they can be used simultane- ously in many applications, we believe that the two tasks may have mutual influences on each other. In this study, we propose a unified ap- proach to simultaneous single-document and multi-document summarizations. The mutual influences between the two tasks are incorpo- rated into a graph-based model. The ranking scores of sentences for single-document summa- rization and the ranking scores of sentences for multi-document summarization can boost each other, and they can be obtained simultaneously in a unified graph-based ranking process. To the best of our knowledge, this study is the first at- tempt for simultaneously addressing the two summarization tasks in a unified graph-based framework. Moreover, the proposed approach can be easily adapted for topic-focused summa- rizations. Experiments have been performed on both the single-document and multi-document summari- zation tasks of DUC2001 and DUC2002. The results demonstrate that the proposed approach can outperform baseline independent methods for both the two summarization tasks. The two tasks are validated to have mutual influences …

Content truncated to 500 words