Multi-dimensional Logical Reasoning for Event Relation Extraction

Event Relation Extraction (ERE) aims to reason and categorize multiple relations among events, including temporal, causal, and subevent relations, to enhance natural language understanding. Many existing approaches, while effective in simple relation extraction, face two main challenges: ambiguity in event understanding and the neglect of implicit relational constraints. To address these two issues, we propose a Multi-dimensional Logical Reasoning Model for Event Relation Extraction (MLERE), which is designed to enrich event representations and enable interpretable relational inference through structured relational constraints. The model consists of two key modules: the Multi-dimensional Clues Discovery module, which refines event representations by constructing an event-centric heterogeneous graph. In this graph, each event is represented as a node, and various event-specific indicators-such as timing, causality, and subevent indicators-are encoded as edges between nodes. The Constraint-guided Reasoning module leverages the constructed heterogeneous graph to capture and integrate both inter- and intra-relational constraints, facilitating the interpretation of complex event dependencies and the identification of implicit event relations. Experimental results on three ERE datasets demonstrate that MLERE significantly outperforms baselines in terms of average F1 score. We provide extensive ablation studies to confirm the contribution of each module, and further analysis explores the properties and impact of learned relational constraints. Code is available at https://github.com/mira-ai-lab/MLERE.