CorefPrompt: Prompt-based Event Coreference Resolution by Measuring Event Type and Argument Compatibilities

Sheng Xu, Peifeng Li, Qiaoming Zhu

Abstract

Event coreference resolution (ECR) aims to group event mentions referring to the same real-world event into clusters. Most previous studies adopt the “encoding first, then scoring” framework, making the coreference judgment rely on event encoding. Furthermore, current methods struggle to leverage human-summarized ECR rules, e.g., coreferential events should have the same event type, to guide the model. To address these two issues, we propose a prompt-based approach, CorefPrompt, to transform ECR into a cloze-style MLM (masked language model) task. This allows for simultaneous event modeling and coreference discrimination within a single template, with a fully shared context. In addition, we introduce two auxiliary prompt tasks, event-type compatibility and argument compatibility, to explicitly demonstrate the reasoning process of ECR, which helps the model make final predictions. Experimental results show that our method CorefPrompt performs well in a state-of-the-art (SOTA) benchmark.

Introduction

Event Coreference Resolution (ECR) is the task of identifying when different mentions in a text (e.g., "the explosion" and "the blast") refer to the same real-world occurrence. This is notoriously difficult because event triggers and their arguments (who, where, when) are often scattered across a document.

Most existing systems follow an "encode first, then score" architecture. First, a transformer encodes the text to create embeddings; then, a separate classifier scores pairs of these embeddings. The authors identify two major flaws in this traditional approach:

  1. Information Blocking: The encoder doesn't "know" it's being used for coreference during the encoding phase, potentially losing crucial contextual nuances.
  2. Rule Integration: It is difficult to force deep neural networks to follow human-summarized logic, such as the rule that coreferential events must share the same event type and compatible arguments.

Methods

CorefPrompt transforms ECR into a cloze-style Masked Language Model (MLM) task. Instead of a binary classifier, it uses a sophisticated prompt template that integrates modeling and judgment into a single pass with a shared context.

Architecture of CorefPrompt

The template consists of three main parts:

  • Prefix Templates: These guide the model to focus on specific event mentions within the text.
  • Anchor Templates: Inserted near event mentions to explicitly mark their types and arguments (e.g., "Event A expresses a [MASK] event with [Person] as participants").
  • Inference Templates: Inspired by "Chain of Thought" prompting, these explicitly demonstrate the reasoning process: "Events A and B have [MASK] event type and [MASK] participants... refer to [MASK] event."

Mask Token Embedding Updating

Standard prompting relies purely on the model's internal attention. CorefPrompt enhances this via tensor matching. It calculates element-wise products and multi-perspective cosine similarities between event mentions and injects these "interactive features" directly into the mask token embeddings before they reach the final prediction head. This ensures the model explicitly compares the two events rather than just looking at them in isolation.

Experimental Setup

The model was evaluated on the standard KBP (Knowledge Base Population) benchmarks, using KBP 2015 and 2016 for training and KBP 2017 for testing. To handle the massive number of non-coreferent event pairs, the authors employed a CorefNM undersampling strategy to select the most "difficult" negative samples (those with high similarity), training a more robust model.

Results and Discussion

CorefPrompt achieved state-of-the-art performance on the KBP 2017 dataset, slightly edging out previous benchmarks.

Comparison with State-of-the-Art Models

Key Insights

  • Prompting as a Bridge: By using MLM, the "modeling" of the event and the "judgment" of coreference happen simultaneously in a shared context, solving the information blocking problem.
  • Explicit Reasoning Works: The auxiliary tasks (predicting event type and compatibility) guide the model to focus on features that matter for coreference, mimicking human logic.
  • Efficiency over Brute Force: Unlike previous SOTA models requiring document-level encoding and 44GB+ of GPU memory, CorefPrompt uses segment-level encoding, fitting on consumer-grade 24GB GPUs (e.g., RTX 3090) while maintaining performance.
  • Ablation Impact: Removing "Anchor" and "Auxiliary" tasks caused the event-pair classification F1 to drop by over 2 points, confirming that explicitly marking arguments and types is critical.

Conclusion

CorefPrompt reimagines ECR as a prompt-based MLM task, bridging the gap between event modeling and coreference judgment. While it achieves SOTA results with lower memory requirements, future work is needed to address pipeline dependencies (pre-identified triggers) and the computational overhead of generating samples for every possible event pair. Jointly identifying triggers and resolving coreference in a single pass remains a promising future direction.

Summary updated

Abstract

Event coreference resolution (ECR) aims to group event mentions referring to the same real-world event into clusters. Most previous studies adopt the “encoding first, then scoring” framework, making the coreference judgment rely on event encoding. Furthermore, current methods struggle to leverage human-summarized ECR rules, e.g., coreferential events should have the same event type, to guide the model. To address these two issues, we propose a prompt-based approach, CorefPrompt, to transform ECR into a cloze-style MLM (masked language model) task. This allows for simultaneous event modeling and coreference discrimination within a single template, with a fully shared context. In addition, we introduce two auxiliary prompt tasks, event-type compatibility and argument compatibility, to explicitly demonstrate the reasoning process of ECR, which helps the model make final predictions. Experimental results show that our method CorefPrompt performs well in a state-of-the-art (SOTA) benchmark.
BibTeX
@inproceedings{xu2023corefprompt,
  author = {Sheng Xu and Peifeng Li and Qiaoming Zhu},
  title = {{CorefPrompt: Prompt-based Event Coreference Resolution by Measuring Event Type and Argument Compatibilities}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=mDrx5kowNd},
}
Figure