A Robust Joint-Training Graph Neural Networks Model for Event Detection with Symmetry and Asymmetry Noisylabels

Abstract

Events are the core element of information in descriptive corpus. Although many progresses have beenmade in Event Detection (ED), it is still a challenge in Natural Language Processing (NLP) to detect event information from data with unavoidable noisy labels. A robust Joint-training Graph ConvolutionNetworks (JT-GCN) model is proposed to meet the challenge of ED tasks with noisy labels in this paper. Specifically, we first employ two Graph Convolution Networks with Edge Enhancement (EE-GCN) tomake predictions simultaneously. A joint loss combining the detection loss and the contrast loss fromtwonetworks is then calculated for training. Meanwhile, a small-loss selection mechanism is introduced tomitigate the impact of mislabeled samples in networks training process. These two networks gradually reach an agreement on the ED tasks as joint-training progresses. Corrupted data with label noise are generated from the benchmark dataset ACE2005. Experiments on ED tasks has been conducted with bothsymmetry and asymmetry label noise on dif erent level. The experimental results show that the proposedmodel is robust to the impact of label noise and superior to the state-of-the-art models for EDtasks.