Harvesting Events from Multiple Sources: Towards a Cross-Document Event Extraction Paradigm

TL;DR

The paper defines cross-document event extraction (CDEE) to overcome the incomplete, biased event representations from single documents. It introduces the CLES dataset, a large-scale Wikipedia-derived Chinese corpus with nine event types and over 20k documents, and proposes a five-stage CDEE pipeline (event extraction, coreference, entity normalization, role normalization, and entity-role resolution) achieving around 72% end-to-end F1. Through extensive experiments, it demonstrates improvements in document-level extraction, cross-document integration, and coreference, while highlighting remaining challenges and the potential of LLM-based approaches. The work establishes CDEE as a new research direction, offering a substantial dataset and a solid benchmarking framework to foster future advancements in multi-document information extraction and knowledge construction.

Abstract

Document-level event extraction aims to extract structured event information from unstructured text. However, a single document often contains limited event information and the roles of different event arguments may be biased due to the influence of the information source. This paper addresses the limitations of traditional document-level event extraction by proposing the task of cross-document event extraction (CDEE) to integrate event information from multiple documents and provide a comprehensive perspective on events. We construct a novel cross-document event extraction dataset, namely CLES, which contains 20,059 documents and 37,688 mention-level events, where over 70% of them are cross-document. To build a benchmark, we propose a CDEE pipeline that includes 5 steps, namely event extraction, coreference resolution, entity normalization, role normalization and entity-role resolution. Our CDEE pipeline achieves about 72% F1 in end-to-end cross-document event extraction, suggesting the challenge of this task. Our work builds a new line of information extraction research and will attract new research attention.

Figures (3)

• Figure 1: An example of cross-document event extraction, where a comprehensive event is obtained from three event mentions in three documents.
• Figure 2: The process of dataset construction.
• Figure 3: Our CDEE pipeline framework consists of five main components. Event Extraction Module performs document-level event extraction. $d$ represents the document, and $s_i$ represents the $i$-th sentence in the document. Event Coreference Resolution Module clusters the event mentions of the same event. Entity Normalization Module links entities to a knowledge base. Role Normalization Module unifies the descriptions of event argument roles. Entity-Role Resolution Module performs deduplication and conflict resolution for cross-document events in the document collection.