Topic-Driven Multi-Document Summarization with Encyclopedic Knowledge and Spreading Activation

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 763–772, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Topic-Driven Multi-Document Summarization with Encyclopedic Knowledge and Spreading Activation Vivi Nastase EML Research gGmbH Heidelberg, Germany nastase@eml-research.de Abstract Information of interest to users is often dis- tributed over a set of documents. Users can specify their request for information as a query/topic – a set of one or more sentences or questions. Producing a good summary of the relevant information relies on understand- ing the query and linking it with the associ- ated set of documents. To “understand” the query we expand it using encyclopedic knowl- edge in Wikipedia. The expanded query is linked with its associated documents through spreading activation in a graph that represents words and their grammatical connections in these documents. The topic expanded words and activated nodes in the graph are used to produce an extractive summary. The method proposed is tested on the DUC summariza- tion data. The system implemented ranks high compared to the participating systems in the DUC competitions, confirming our hypothesis that encyclopedic knowledge is a useful addi- tion to a summarization system. 1 Introduction Topic-driven summarization reflects a user-based summarization task: from a set of documents de- rive a summary that contains information on a spe- cific topic of interest to a user. Producing a good summary relies on “understanding” the user’s infor- mation request, and the documents to be summa- rized. It is commonly agreed that the verbal part of a text provides pointers to a much larger body of knowledge we assume the listener has. An Amer- ican citizen, for example, when told There will be fireworks on July 4th, understands that there will be a celebration involving fireworks on the occasion of the U.S. Independence Day. Understanding an utterance implies lexical, common-sense and ency- clopedic knowledge. Lexical knowledge is usually incorporated in systems through machine readable dictionaries, wordnets or thesauri. Common-sense and encyclopedic knowledge were harder to capture, but recently Wikipedia has opened the possibility of accessing such knowledge on a large scale, and in numerous languages. To “understand” a user’s information request – one or more sentences or questions (the topic of the summary) – summarization systems try to ex- pand it. This will provide later stages of process- ing with more keywords/keyphrases for retrieving from the documents relevant fragments. In this pa- per we experiment with Wikipedia for topic expan- sion. The body of research involving Wikipedia as a source of knowledge is growing fast, as the NLP community finds more and more applications of this useful resource: it is used to acquire knowl- edge (Suchanek et al., 2007; Auer et al., 2007); to induce taxonomies and compute semantic relat- edness (Ponzetto & Strube, 2007b; 2007a); as a source of features for text classification (Gabrilovich & Markovitch, 2006) and for answering questions (Ahn et al., 2004; Katz et al., 2005). The work pre- sented here uses hyperlinks in Wikipedia articles to expand keywords and …

Content truncated to 500 words