Extracting Social Networks and Biographical Facts From Conversational Speech Transcripts

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 1040–1047, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Extracting Social Networks and Biographical Facts From Conversational Speech Transcripts Hongyan Jing IBM T.J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598 hjing@us.ibm.com Nanda Kambhatla IBM India Research Lab EGL, Domlur Ring Road Bangalore - 560071, India kambhatla@in.ibm.com Salim Roukos IBM T.J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598 roukos@us.ibm.com Abstract We present a general framework for automatically extracting social networks and biographical facts from conversational speech. Our approach relies on fusing the output produced by multiple informa- tion extraction modules, including entity recognition and detection, relation detec- tion, and event detection modules. We describe the specific features and algo- rithmic refinements effective for conver- sational speech. These cumulatively in- crease the performance of social network extraction from 0.06 to 0.30 for the devel- opment set, and from 0.06 to 0.28 for the test set, as measured by f-measure on the ties within a network. The same frame- work can be applied to other genres of text — we have built an automatic biography generation system for general domain text using the same approach. 1 Introduction A social network represents social relationships between individuals or organizations. It consists of nodes and ties. Nodes are individual actors within the networks, generally a person or an or- ganization. Ties are the relationships between the nodes. Social network analysis has become a key technique in many disciplines, including modern sociology and information science. In this paper, we present our system for au- tomatically extracting social networks and bio- graphical facts from conversational speech tran- scripts by integrating the output of different IE modules. The IE modules are the building blocks; the fusing module depicts the ways of assembling these building blocks. The final output depends on which fundamental IE modules are used and how their results are integrated. The contributions of this work are two fold. We propose a general framework for extracting social networks and biographies from text that applies to conversational speech as well as other genres, in- cluding general newswire stories. Secondly, we present specific methods that proved effective for us for improving the performance of IE systems on conversational speech transcripts. These improve- ments include feature engineering and algorithmic revisions that led to a nearly five-fold performance increase for both development and test sets. In the next section, we present our framework for extracting social networks and other biograph- ical facts from text. In Section 3, we discuss the refinements we made to our IE modules in order to reliably extract information from conversational speech transcripts. In Section 4, we describe the experiments, evaluation metrics, and the results of social network and biography extraction. In Sec- tion 5, we show the results of applying the frame- work to other genres of text. Finally, we discuss related work and conclude with lessons learned and future work. 2 The General Framework For …

Content truncated to 500 words