Cross-document Event Extraction and Tracking: Task, Evaluation, Techniques and Challenges

International Conference RANLP 2009 - Borovets, Bulgaria, pages 166–172 Cross-document Event Extraction and Tracking: Task, Evaluation, Techniques and Challenges Heng Ji* Ralph Grishman† Zheng Chen* Prashant Gupta‡ *Computer Science Department, Queens College and The Graduate Center, The City University of New York †Computer Science Department, New York University ‡Computer Science Department, Indian Institute of Information Technology Allahabad hengji@cs.qc.cuny.edu Abstract This paper proposes a new task of cross-document event extraction and tracking and its evaluation metrics. We identify important person entities which are frequently involved in events as ‘centroid entities’. Then we link the events involving the same centroid entity along a time line. We also present a system performing this task and our current approaches to address the main research challenges. We demonstrate that global inference from background knowledge and cross-document event aggregation are crucial to enhance the performance. This new task defines several extensions to the traditional single-document Information Extraction paradigm beyond ‘slot filling’. Keywords Information Extraction, Cross-document Extraction, Event Extraction 1. Introduction Consider a user monitoring or browsing a multi-source news feed, with assistance from an Information Extraction (IE) system. Various events are evolving, updated, repeated and corrected in different documents; later information may override earlier more tentative or incomplete facts. In this environment, traditional single- document IE would be of little value; a user would be confronted by thousands of unconnected events with tens of thousands of arguments. Add to this the fact that the extracted results contain unranked, redundant and erroneous facts and some crucial facts are missing, and it’s not clear whether these IE results are really beneficial. How can we take proper advantage of the power of extraction to aid news analysis? In this paper we define a new cross-document IE task beyond ‘slot filling’ to generate more coherent, salient, complete and concise facts. A high-coherence text has fewer conceptual gaps and thus requires fewer inferences and less prior knowledge, rendering the text easier to understand [1]. In our task, coherence is the extent to which the relationships between the events in the documents can be made explicit. We aim to provide a more coherent presentation by linking events based on shared arguments. In the news from a certain period some entities are more central than others; we propose to identify these centroid entities, and then link the events involving the same centroid entity on a time line. In this way we provide coherent event chains so that users can more efficiently review and analyze events, such as tracking a person’s movement activities and trends. This will offer a richer set of views than is possible with document clustering for summarization or with topic tracking. • To sum up, the specific goals of this paper are to: Formulate a tractable but challenging task of cross- document IE, producing a product useful for browsing, analysis, and search; • Propose a set of metrics for this task; • Present a first cut at a system for performing this task; • Lay out the potential research challenges …

Content truncated to 500 words