Event Causality Identification Model Based on External Vocabulary and Hypergraph Denoising

LIAO Tao; HU Haiqian; NIU Bingyu

doi:10.13501/j.cnki.42-1908/n.2025.12.019

Abstract：To address the problem that existing event causality identification methods failed to consider the noise interference generated after the introduction of external knowledge,leading to increase ambiguity in event representation and thereby affecting the identification performance,an event causality identification model based on external vocabulary and hypergraph denoising(EHDM)was proposed.Firstly,contextual knowledge was retrieved from an external vocabulary to enrich the semantic information of events,encoding event descriptions with this contextual knowledge.Secondly,a hypergraph was constructed based on knowledge features corresponding to multiple relationships within the event background knowledge.Features were further processed through a hypergraph convolutional neural network and a multi-head attention mechanism to obtain denoised event feature representations.Thirdly,context-based feature representations were encoded from the events and their contexts,which were then fused with the denoised event feature representations via a gate unit.Finally,the fused feature representations were input into a multi-layer perceptron to obtain prediction values,thereby achieving causality identification.The results demonstrated that EHDM achieved an F1 score improvement of 1.5 percentage points over relation graph convolutional networks(RGCN)on the intra-sentential aspect of the causal-timebank(CTB)dataset.On the intra-sentential aspect of the event story line(ESL)dataset,EHDM achieved an F1 score improvement of 2.4 percentage points over RGCN and also recorded increases of 2.1 and 3.0 percentage points in cross-sentence and overall F1 scores,respectively,compared to the event relation graph Transformer model.This research confirmed EHDM′s effective application in the field of event causal relationship identification.