From Intuition to Verification: Cognitive Neuro-Symbolic Reasoning for Document-level Event Causality Identification

ACL ARR 2026 January Submission4663 Authors

Keywords: Document-level Event Causality Identification;Neuro-Symbolic Reasoning;Cognitive Modeling;Evidential Deep Learning

Abstract: Document-level Event Causality Identification (DECI) aims to infer causal relations between events distributed across long documents, where causality is often implicit and evidence is fragmented. Existing approaches typically follow two paradigms: structure-based models that emphasize predefined graphs but struggle to capture implicit semantic relations, and generative large language models (LLMs) that flexibly propose causal hypotheses yet lack reliable global verification. Inspired by the cognitive transition from intuition to verification, we propose COgnitive Verification for Event Reasoning (COVER), a cognitive neuro-symbolic framework that explicitly integrates intuitive hypothesis generation with structured verification for DECI. COVER treats causal reasoning as a closed-loop process. In the intuition stage, an LLM serves as a variational prior to generate causal hypotheses and retrieve supportive commonsense knowledge, which is filtered via entropy-aware knowledge anchoring. In the verification stage, these hypotheses are embedded into a document-level neuro-symbolic causal graph and evaluated under global structural constraints with uncertainty-aware reasoning, enabling unreliable hypotheses to be refined rather than directly accepted. Experiments on CEC 2.0 and MAVEN-ERE demonstrate that COVER consistently outperforms strong baselines, with notable gains on implicit and long-range causal relations.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Information Extraction,Information Retrieval and Text Mining,NLP Applications

Languages Studied: English

Submission Number: 4663