Utilizing Contextual Clues and Role Correlations for Enhancing Document-level Event Argument Extraction

TL;DR

This work tackles document-level event argument extraction by addressing two core gaps: the lack of contextual clue integration and the omission of semantic role correlations. It introduces CARLG, a lightweight framework consisting of Contextual Clues Aggregation (CCA) and Role-based Latent Information Guidance (RLIG), which leverage pre-trained encoder attentions and latent role embeddings to enhance candidate representations. CARLG is instantiated in two forms, CARLG_span and CARLG_prompt, and consistently improves span-based and prompt-based EAE methods across RAMS, WikiEvents, and MLEE with minimal parameter overhead. The approach achieves state-of-the-art results while maintaining efficiency, and analyses confirm the effectiveness of CCA and RLIG in enriching context interaction and inter-role guidance, with potential for extending to trigger-agnostic EAE and other information extraction tasks.

Abstract

Document-level event argument extraction is a crucial yet challenging task within the field of information extraction. Current mainstream approaches primarily focus on the information interaction between event triggers and their arguments, facing two limitations: insufficient context interaction and the ignorance of event correlations. Here, we introduce a novel framework named CARLG (Contextual Aggregation of clues and Role-based Latent Guidance), comprising two innovative components: the Contextual Clues Aggregation (CCA) and the Role-based Latent Information Guidance (RLIG). The CCA module leverages the attention weights derived from a pre-trained encoder to adaptively assimilates broader contextual information, while the RLIG module aims to capture the semantic correlations among event roles. We then instantiate the CARLG framework into two variants based on two types of current mainstream EAE approaches. Notably, our CARLG framework introduces less than 1% new parameters yet significantly improving the performance. Comprehensive experiments across the RAMS, WikiEvents, and MLEE datasets confirm the superiority of CARLG, showing significant superiority in terms of both performance and inference speed compared to major benchmarks. Further analyses demonstrate the effectiveness of the proposed modules.

Figures (12)

• Figure 1: A document from RAMS ebner2020multi dataset. The event violence has the trigger of Homicide and there are three distinct arguments, each assuming different roles. We use the color red to emphasize specific words, representing contextual clues that are significant for the extraction of arguments.
• Figure 2: The visualization represents the co-occurrence frequencies of the top 15 common roles within the RAMS test set, which is defined as the ratio between the co-occurrence times and the sum of the total occurrence times of two roles, and the self-co-occurrence count is set to zero.
• Figure 3: Contextual Clues Aggregation Module. The core process of the Role-based Latent Information Guidance algorithm is also similar to this.
• Figure 4: The main structure of $\text{CARLG}_\text{span}$. The input sequence, encompassing roles and event types, passes through a transformer-based encoder. This encoder generates output in the form of context representations, role representations, and attention heads. The Contextual Clue Aggregation (CCA) module dynamically consolidates contextual clues, while the Role-based Latent Information Generation (RLIG) module generates latent role embeddings and captures role correlations. Subsequently, the context vector is combined with the latent role vector to create the ultimate span representation. Finally, the classification module leverages this representation to forecast argument roles for all potential spans.
• Figure 5: Detailed RLIG module for $\text{CARLG}_\text{span}$. Sub-figure (a) illustrates the information fusion and (b) illustrates the Latent Role Information Guidance operation.