Robustness Analysis on Graph Neural Networks Model for Event Detection

Abstract

Event Detection (ED), which aims to identify trigger words from the given text and classify them into corresponding event types, is an important task in Natural Language Processing (NLP); it contributes to several downstream tasks and is beneficial for many real-world applications. Most of the current SOTA (state-of-the-art) models for ED are based on Graph Neural Networks (GNN). However, a few studies focus on the issue of GNN-based ED models’ robustness towards text adversarial attacks, which is a challenge in practical applications of EDs that needs to be solved urgently. In this paper, we first propose a robustness analysis framework for an ED model. Using this framework, we can evaluate the robustness of the ED model with various adversarial data. To improve the robustness of the GNN-based ED model, we propose a new multi-order distance representation method and an edge representation update method based on attention weights, then design an innovative model named A-MDL-EEGCN. Extensive experiments illustrate that the proposed model can achieve better performance than other models both on original data and various adversarial data. The comprehensive robustness analysis according to experimental results in this paper brings new insights into the evaluation and design of a robust ED model.