An Unsupervised Approach to Biography Production Using Wikipedia

Proceedings of ACL-08: HLT, pages 807–815, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics An Unsupervised Approach to Biography Production using Wikipedia Fadi Biadsy,† Julia Hirschberg† and Elena Filatova* †Department of Computer Science Columbia University, New York, NY 10027, USA {fadi,julia}@cs.columbia.edu *InforSense LLC Cambridge, MA 02141, USA efilatova@inforsense.com Abstract We describe an unsupervised approach to multi-document sentence-extraction based summarization for the task of producing biographies. We utilize Wikipedia to auto- matically construct a corpus of biographical sentences and TDT4 to construct a corpus of non-biographical sentences. We build a biographical-sentence classifier from these corpora and an SVM regression model for sentence ordering from the Wikipedia corpus. We evaluate our work on the DUC2004 evaluation data and with human judges. Overall, our system significantly outperforms all systems that participated in DUC2004, according to the ROUGE-L metric, and is preferred by human subjects. 1 Introduction Producing biographies by hand is a labor-intensive task, generally done only for famous individuals. The process is particularly difficult when persons of interest are not well known and when information must be gathered from a wide variety of sources. We present an automatic, unsupervised, multi-document summarization (MDS) approach based on extractive techniques to producing biographies, answering the question “Who is X?” There is growing interest in automatic MDS in general due in part to the explosion of multilingual and multimedia data available online. The goal of MDS is to automatically produce a concise, well- organized, and fluent summary of a set of docu- ments on the same topic. MDS strategies have been employed to produce both generic summaries and query-focused summaries. Due to the complexity of text generation, most summarization systems em- ploy sentence-extraction techniques, in which the most relevant sentences from one or more docu- ments are selected to represent the summary. This approach is guaranteed to produce grammatical sen- tences, although they must subsequently be ordered appropriately to produce a coherent summary. In this paper we describe a sentence-extraction based MDS procedure to produce biographies from online resources automatically. We make use of Wikipedia, the largest free multilingual encyclope- dia on the internet, to build a biographical-sentence classifier and a component for ordering sentences in the output summary. Section 2 presents an overview of our system. In Section 3 we describe our cor- pus and in Section 4 we discuss the components of our system in more detail. In Section 5, we present an evaluation of our work on the Document Under- standing Conference of 2004 (DUC2004), the biog- raphy task (task 5) test set. In Section 6 we com- pare our research with previous work on biography generation. We conclude in Section 7 and identify directions for future research. 2 System Overview In this section, we present an overview of our biog- raphy extraction system. We assume as input a set of documents retrieved by an information retrieval en- gine from a query consisting of the name of the per- son for whom the biography is desired. We further assume that these …

Content truncated to 500 words