A Multi-graph Representation for Event Extraction

Background Introduction: Event extraction is a popular task in the field of natural language processing, aiming to identify event trigger words and their related arguments from a given text. This task is typically divided into two subtasks: event detection (extracting event trigger words) and argument extraction. The traditional pipeline method performs these two subtasks separately, but suffers from the problem of error propagation. In recent years, joint models have emerged that can learn these two subtasks together, avoiding error propagation, but still neglecting the argument multiplexing problem.

Paper Summary: This paper proposes a multigraph-based event extraction framework. A multigraph allows multiple parallel edges between two nodes, which can effectively represent the semantic structure of events and solve the argument multiplexing problem. Based on this framework, the authors designed an end-to-end multigraph event extraction model (MGREE) that can simultaneously extract event trigger words, related arguments, and their semantic roles.

Research Institution and Authors: This research was supported by the National Key R&D Program of China and was a collaborative effort between researchers from the Key Laboratory of Text Computing and Cognitive Intelligence Engineering, Guizhou University, and the Department of Computer Science and Technology, Xi’an Jiaotong University. The first author is Hui Huang from Guizhou University, and the corresponding authors are Yanping Chen and Yongbin Qin from Guizhou University.

Research Process: (a) Tokenize the input sentence, and use pre-trained language models like BERT and Skip-gram word embeddings to generate node representations; (b) Construct an N×N×K three-dimensional tensor as the multigraph representation, where N is the number of tokens in the sentence, and K is the number of annotation types (including event types, argument types, and semantic role types); © Learn the confidence score of each edge based on an attention mechanism to obtain the multigraph representation; (d) Design a rule-based event decoding algorithm to extract event trigger words and their argument roles from the multigraph.

Main Results: Experiments were conducted on four public datasets, including ACE05, and the MGREE model achieved the latest state-of-the-art performance in the event extraction task, with an F1 score improvement of approximately 4% over the existing best model. Analysis experiments showed that the multigraph representation effectively solved the argument multiplexing problem and improved the discriminative ability of neural networks in event extraction.

Research Significance: (1) Introduced the multigraph representation, which can effectively represent event semantic structures and solve the argument multiplexing problem; (2) Designed the end-to-end event extraction model MGREE, achieving the latest state-of-the-art performance; (3) Analyzed the advantages of the multigraph representation, such as representational capacity and computational efficiency.

Innovations: (1) Proposed the multigraph representation, which is the first to solve the argument multiplexing problem; (2) MGREE is the first end-to-end event extraction model based on the multigraph representation; (3) Analytical experiments verified the advantages of the multigraph representation in terms of representational capacity and computational efficiency.

This research proposed an innovative multigraph event representation method that can effectively solve the argument multiplexing problem and designed an end-to-end event extraction model that achieved the best performance on public datasets, which is of great significance for the development of the event extraction task.