Identifying while Learning for Document Event Causality Identification

TL;DR

This work addresses document-level event causality identification by introducing an iterative identifying while learning framework (iLIF) that jointly learns contextual event representations and directed causal structures. iLIF constructs a directed event causality graph (ECG) from high-confidence predictions and uses a causal graph encoder to update per-event representations, enabling improved causal direction and existence identification through successive iterations. The model integrates a contextual text encoder, a heterogeneously connected ECG, and an iteration-aware training strategy with a termination criterion, achieving state-of-the-art results on EventStoryLine and MAVEN-ERE under both direction and existence settings. Experiments show that incorporating directionality and iterative graph-based refinement yields tangible gains, while ablations and parameter studies highlight the importance of edge heterogeneity, iteration, and thresholding. The approach advances practical ECI by leveraging structured causal reasoning and disciplined propagation control, with potential impacts on knowledge graph construction and cross-document reasoning.

Abstract

Event Causality Identification (ECI) aims to detect whether there exists a causal relation between two events in a document. Existing studies adopt a kind of identifying after learning paradigm, where events' representations are first learned and then used for the identification. Furthermore, they mainly focus on the causality existence, but ignoring causal direction. In this paper, we take care of the causal direction and propose a new identifying while learning mode for the ECI task. We argue that a few causal relations can be easily identified with high confidence, and the directionality and structure of these identified causalities can be utilized to update events' representations for boosting next round of causality identification. To this end, this paper designs an *iterative learning and identifying framework*: In each iteration, we construct an event causality graph, on which events' causal structure representations are updated for boosting causal identification. Experiments on two public datasets show that our approach outperforms the state-of-the-art algorithms in both evaluations for causality existence identification and direction identification.

Figures (7)

• Figure 1: An example of the event causality graph and event structures in the EventStoryLine corpus.
• Figure 2: Illustration of the iterative learning and identifying framework (iLIF). Causality identification is based on events' contextual text and causal graph representations. The event causality graph (ECG) is iteratively constructed to update events' causal graph representations. The final ECG contains all identified causal relations as the output.
• Figure 3: F1 scores on EventStoryLine when using different edge thresholds in the direction setting.
• Figure 4: Results of different maximum number of iterations on EventStoryLine in the direction setting.
• Figure 5: Comparison of causality identification results in the initial ECG and final ECG.