Structure and Co-occurrence aware for Document-level Event Argument Extraction

Anonymous

TL;DR: Our model incorporates the Co-occurrence-aware prefix to help the model capture event semantic boundaries, and the Structure-aware prefix to build structured information of the entire document.

Abstract: Document-level Event Argument Extraction (EAE) deals with longer texts, and more intricate relationships between events than sentence-level, which faced two problem: 1) semantic boundaries between events are difficult to distinguish; 2) redundant information distracts attention from events. To alleviate the aforementioned issues, we propose the Structure and Co-occurrence aware Event Argument Extraction model (SCEAE). SCEAE utilizes the PAIE architecture as the underlying framework. Building upon this framework, we incorporates two different knowledge-aware prefixes to tackle these problems. The Co-occurrence-aware prefix leverages knowledge of event co-occurrence to enhance the model's perception of semantic boundaries between events. The Structure-aware prefix helps the model establish structured relationships between the sentence. We tested our model on the RAMS, WikiEvents and MLEE datasets. The experiments showed that our model achieved gains of 2.1\%, 2.3\%, and 3.2\% in the Arg-C F1 metric compared to PAIE on RAMS, WikiEvents and MLEE respectively. Furthermore, our model achieved new state-of-the-art performance. We will make all the progress publicly available at https://github.com/---.

Paper Type: long

Research Area: Information Extraction

Contribution Types: Model analysis & interpretability

Languages Studied: English