Beyond Single-Event Extraction: Towards Efficient Document-Level Multi-Event Argument Extraction

TL;DR

This work tackles document-level event argument extraction for documents with multiple events, where traditional single-event EAE processes are inefficient and fail to exploit cross-event correlations. The authors introduce the DEEIA framework, comprising a multi-event prompt mechanism, a Dependency-guided Encoding (DE) module, and an Event-specific Information Aggregation (EIA) module to extract arguments for all events in a document simultaneously. They demonstrate state-of-the-art performance on RAMS, WikiEvents, MLEE, and ACE05, with significant inference-time reductions compared to single-event baselines and multi-event baselines. The results indicate that explicit modeling of event dependencies and event-oriented context improves both accuracy and efficiency, making DEEIA practical for document-scale IE tasks.

Abstract

Recent mainstream event argument extraction methods process each event in isolation, resulting in inefficient inference and ignoring the correlations among multiple events. To address these limitations, here we propose a multiple-event argument extraction model DEEIA (Dependency-guided Encoding and Event-specific Information Aggregation), capable of extracting arguments from all events within a document simultaneouslyThe proposed DEEIA model employs a multi-event prompt mechanism, comprising DE and EIA modules. The DE module is designed to improve the correlation between prompts and their corresponding event contexts, whereas the EIA module provides event-specific information to improve contextual understanding. Extensive experiments show that our method achieves new state-of-the-art performance on four public datasets (RAMS, WikiEvents, MLEE, and ACE05), while significantly saving the inference time compared to the baselines. Further analyses demonstrate the effectiveness of the proposed modules.

Figures (10)

• Figure 1: Subfigure (a) explains the prompt-based EAE task with one of the events in context $X$. The prompt is manually designed for the specified event type, with mentions of roles as slots, such as $\left< killer \right>$. (b) shows the difference between traditional Single-EAE method and our Multi-EAE method, the latter is more difficult.
• Figure 2: We select document samples containing different numbers of events and calculate the inference time on one sample for a Single-EAE method PAIE ma2022prompt and our Multi-EAE model DEEIA. The results are averaged on 100 repeated experiments. With the increase of event numbers within a document, the efficiency advantage of our Multi-EAE model becomes increasingly apparent.
• Figure 3: The architecture of the proposed DEEIA model. For an input document, $\text{P}_1$, $\text{P}_2$, and $\text{P}_3$ represent simplified prompts.
• Figure 4: The averaged performance of the PAIE-multi, TabEAE-multi, and DEEIA models on samples with different event numbers in MLEE dataset. Our model achieves better results on samples with multiple events.
• Figure 5: Visualization of attentive weights in EIA module from an example in RAMS. We calculate the attentive weight $\textbf{p}_{k}$ based on the representations of argument slot "government" and the trigger "bombarding".