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.

What problem does this paper attempt to address?

This paper mainly focuses on the task of document-level Event Causality Identification (ECI). Existing methods typically adopt a "learn first, recognize later" approach, where event representation is learned first and then causal relationships are identified. However, these methods mainly focus on the existence of causal relationships and ignore the directionality. The paper proposes a new "simultaneous recognition and learning" approach that emphasizes the importance of causality direction. The authors believe that some causal relationships can be easily identified with high confidence, and the directionality and structure of these relationships can be used to update the event representation, facilitating subsequent rounds of causal relationship identification. The paper designs an iterative learning and recognition framework (iLIF) that constructs an event causality graph in each iteration and updates the causal structure representation based on this graph to enhance causal recognition. Experimental results show that the iLIF method outperforms existing algorithms on two public datasets, both in terms of causality existence recognition and direction recognition. Furthermore, the paper points out that existing solutions ignore causality direction and adopt the "learn first, recognize later" paradigm. In contrast, the newly proposed iLIF method considers causality direction, which improves the effectiveness of event causality identification. By iteratively updating event representations, iLIF can utilize the recognized causal relationships to improve subsequent recognition processes, reducing the problem of error propagation.

Identifying while Learning for Document Event Causality Identification

EiGC: an Event-Induced Graph with Constraints for Event Causality Identification

Explicit and implicit knowledge-enhanced model for event causality identification

Dual Graph Convolutional Networks for Document-Level Event Causality Identification.

Chinese Event Causality Identification Based on Retrieval Enhancement

A Survey of Event Causality Identification: Principles, Taxonomy, Challenges, and Assessment

Improving Event Causality Identification via Self-Supervised Representation Learning on External Causal Statement

Modeling Correlated Causal-Effect Structure with a Hypergraph for Document-Level Event Causality Identification

In-context Contrastive Learning for Event Causality Identification

Dual Semantic Enhanced Event Causality Identification with Derivative Temporal Prompt

Advancing Event Causality Identification via Heuristic Semantic Dependency Inquiry Network

Enhancing Event Causality Identification with Rationale and Structure-Aware Causal Question Answering

Semantic Structure Enhanced Event Causality Identification

Document-level event causality identification via graph inference mechanism

Event Causality Identification Via Structure Optimization and Reinforcement Learning

PPAT: Progressive Graph Pairwise Attention Network for Event Causality Identification

Towards Event-level Causal Relation Identification

Separate and Integrate Different Level Reasoning for Event Causality Identification

Event Causality Extraction with Event Argument Correlations

Modeling Document-level Causal Structures for Event Causal Relation Identification

LearnDA: Learnable Knowledge-Guided Data Augmentation for Event Causality Identification