Enhancing Cross-Document Event Coreference Resolution by Discourse Structure and Semantic Information

TL;DR

This work tackles cross-document event coreference resolution by addressing the missing document-level information in prior approaches. It introduces DIE-EC, a model that fuses document-level RST-based discourse structure with cross-document lexical chains into a unified graph processed by Graph Attention Networks, followed by an MLP-based pair scoring and agglomerative clustering. The authors also present WEC-Zh, a large-scale Chinese CDECR dataset, and demonstrate state-of-the-art results on both English (WEC-Eng) and Chinese data, including strong ablations showing the value of discourse information and lexical coherence. The approach has a meaningful impact on handling long-distance dependencies and cross-lingual coreference, offering a practical framework for improved event understanding across documents.

Abstract

Existing cross-document event coreference resolution models, which either compute mention similarity directly or enhance mention representation by extracting event arguments (such as location, time, agent, and patient), lacking the ability to utilize document-level information. As a result, they struggle to capture long-distance dependencies. This shortcoming leads to their underwhelming performance in determining coreference for the events where their argument information relies on long-distance dependencies. In light of these limitations, we propose the construction of document-level Rhetorical Structure Theory (RST) trees and cross-document Lexical Chains to model the structural and semantic information of documents. Subsequently, cross-document heterogeneous graphs are constructed and GAT is utilized to learn the representations of events. Finally, a pair scorer calculates the similarity between each pair of events and co-referred events can be recognized using standard clustering algorithm. Additionally, as the existing cross-document event coreference datasets are limited to English, we have developed a large-scale Chinese cross-document event coreference dataset to fill this gap, which comprises 53,066 event mentions and 4,476 clusters. After applying our model on the English and Chinese datasets respectively, it outperforms all baselines by large margins.

Figures (5)

• Figure 1: An example to show cross-document event coreference resolution and our main idea of building document and cross-document information. the first Chief-Executive election and the head of Hong Kong Special Administrative Region Election refer to the same event "1996 Hong Kong Special Administrative Region Chief Executive election".
• Figure 2: The architecture of DIE-EC. In the document-level RST trees, N denotes nucleus and S represents satellite. Lexical chains are across documents. After merging RST trees and lexical chains, Graph Attention Networks (GAT) is applied for representaton learning. Then an MLP is utilized to compute the coreference score.
• Figure 3: An example in WEC-Zh. Upheaval of the Five Barbarians and the Xiongnu, Xianbei, Jie, Qiang, and Di tribes, established non-Han political power refer to the event "During the Western Jin Dynasty, there was Upheaval of the Five Barbarians".
• Figure 4: The impact of lexical chains with different lexical overlap rates.
• Figure 5: The impact of RST with different document lengths.