RAAT: Relation-Augmented Attention Transformer for Relation Modeling in Document-Level Event Extraction

Abstract

In document-level event extraction (DEE) task, event arguments always scatter across sentences (across-sentence issue) and multiple events may lie in one document (multi-event issue). In this paper, we argue that the relation information of event arguments is of great significance for addressing the above two issues, and propose a new DEE framework which can model the relation dependencies, called Relation-augmented Document-level Event Extraction (ReDEE). More specifically, this framework features a novel and tailored transformer, named as Relation-augmented Attention Transformer (RAAT). RAAT is scalable to capture multi-scale and multi-amount argument relations. To further leverage relation information, we introduce a separate event relation prediction task and adopt multi-task learning method to explicitly enhance event extraction performance. Extensive experiments demonstrate the effectiveness of the proposed method, which can achieve state-ofthe-art performance on two public datasets. Our code is available at https://github. com/TencentYoutuResearch/RAAT.

HyperED: A hierarchy‐aware network based on hyperbolic geometry for event detection

“…Despite these limitations, feature-based methods also leave a lot of directions for future studies, such as cross-document or cross-sentence event extraction, graph-based event extraction, which is still prevalent in the past two years. 21,[22][23][24] As neural networks become mainstream, learning an excellent representation emerges as a new goal for event detection. Conventional methods utilize CNN, RNN, 25 BERT 26 and their variant models to encode the input into high-dimensional embeddings.…”

“…On the other hand, the patterns are always designed for specific scenarios or expressions, which limits their reusability. Despite these limitations, feature‐based methods also leave a lot of directions for future studies, such as cross‐document or cross‐sentence event extraction, graph‐based event extraction, which is still prevalent in the past two years 21,22–24 …”

An Iteratively Parallel Generation Method with the Pre-Filling Strategy for Document-level Event Extraction

“…It first selects pseudo triggers and then make a beam extraction based on the graph. GIT (Xu et al, 2021a) and RAAT (Liang et al, 2022) follow the same auto-regressive generation with Doc2EDAG, but add a heterogeneous graph network and entity relation extraction network to enhance the entity representation, respectively.…”

“…Then a widely used entity-based directed acyclic graph (EDAG) generation method is proposed to better deal with multiple events extraction (Zheng et al, 2019). Several variant methods based on EDAG generation are presented by utilizing more meticulous feature engineering, such as heterogeneous graph feature (Xu et al, 2021a; and entity relation feature (Liang et al, 2022). Additionally, a parallel method is proposed to avoid the error broadcast in EDAG generation (Yang et al, 2021), and an efficient model is designed to lighten the model and accelerates the decoding speed (Zhu et al, 2022).…”

“…Most trigger-free DEE methods (Zheng et al, 2019;Xu et al, 2021a;Liang et al, 2022) build entity-based directed acyclic graph (EDAG) to autoregressively generate event arguments with their roles under a predefined order. However, it is important to note that determining the optimal order from the permutation space of possible orders has an NP-hard complexity of O(n!).…”