Abstract
Most existing event causality identification (ECI) methods rarely consider the event causal label information and the interaction information between event pairs. In this paper, we propose a framework to enrich the representation of event pairs by introducing the event causal label information and the event pair interaction information. In particular, 1) we design an event-causal-label-aware module to model the event causal label information, in which we design the event causal label prediction task as an auxiliary task of ECI, aiming to predict which events are involved in the causal relationship (we call them causality-related events) by mining the dependencies between events. 2) We further design an event pair interaction graph module to model the interaction information between event pairs, in which we construct the interaction graph with event pairs as nodes and leverage graph attention mechanism to model the degree of dependency between event pairs. The experimental results show that our approach outperforms previous state-of-the-art methods on two benchmark datasets EventStoryLine and Causal-TimeBank.
Introduction
Event Causality Identification (ECI) is the task of determining if one event in a text (e.g., "scrambling") was caused by another (e.g., "use bandwidth"). Despite its importance, ECI remains challenging due to two primary hurdles. First, the "Needle in the Haystack" problem: in a typical sentence, only a small fraction of event pairs are actually causal, while the majority are unrelated noise. Second, most existing methods treat event pairs in isolation, failing to capture the transitive nature of causality—where knowing the relationship between pairs $(e_1, e_2)$ and $(e_2, e_3)$ can provide critical evidence for the relationship between $(e_1, e_3)$.
Methodology
We propose the ECLEP framework to address these challenges through a multi-stage approach:
- Encoding Layer: Utilizes a pre-trained BERT model to generate contextual embeddings for every word and event within the input sentence.
- Event-Causal-Label-Aware Module: This module performs an auxiliary task of predicting whether an individual event is "causality-related" (involved in any causal relationship). By identifying these key events, the model can prune the search space and reduce interference from unrelated noise.
- Event Pair Interaction Graph Module: Instead of independent classification, this module constructs a graph where each node represents an event pair. Edges are established between pairs that share a common event. A Graph Attention Network (GAT) is then employed to model the dependencies and information flow between these nodes, effectively capturing causal transitivity.
- Prediction and Training: The final pair representations are fed into a classifier, with the model trained using a joint loss function that combines the ECI task and the auxiliary event-labeling task.
Results and Discussion
The ECLEP framework was evaluated on two standard benchmarks: EventStoryLine (ESL) and Causal-TimeBank (CTB).
- Performance: ECLEP achieved new state-of-the-art results, with an F1-score of 57.1 on ESL (a 1.9% improvement) and 56.3 on CTB (a 3.1% improvement).
- Ablation Studies: Removing either the causal label module or the interaction graph module led to significant performance drops, confirming that both filtering noise and structural reasoning are essential.
- Efficiency: By identifying causality-related events first, the model theoretically reduces the number of candidate pairs in ESL from an average of 13.12 down to 3.68, significantly cleaning the data for the final classifier.
Conclusion
By integrating event causal label information and modeling inter-pair dependencies through an interaction graph, ECLEP significantly improves the accuracy of causality identification. Future work will explore extending this framework to document-level causality and other relational tasks like temporal relation extraction.
Summary updated