A Two-Stage Label Rectification Framework for Noisy Event Extraction

Zijie Xu, Peng Wang, Ziyu Shang, Jiajun Liu

Abstract: Event extraction aims to identify event triggers and corresponding arguments. Existing methods are mainly devoted to well-designed deep neural networks based on the assumption that training data are high-quality. However, noisy labels are unavoidable in real-world scenarios, which is challenging for current event extraction applications. This paper proposes a novel Two-stage label rectification framework (Tolar) to tackle this problem from explicit and latent aspects. In the first stage, an event schema mapping module is designed to rectify the explicit label inconsistency. Then a self-adaptive iteration module addresses the latent semantic noise in the second stage. In order to cope with extremely noisy labels, we further design a Cooperative Global Pointer Network (CoGPN) to train two global pointer networks concurrently and let them filter possibly noisy labels mutually. Extensive experiments on ACE 2005 and MAVEN with synthetic noise demonstrate that our framework Tolar effectively enhances event extraction methods, and our CoGPN achieves state-of-the-art performance in extremely noisy settings.

External IDs:doi:10.1007/978-3-031-30675-4_39