Incremental event detection via an improved knowledge distillation based model

doi:10.1016/j.neucom.2023.126519

2023-09-01

PRE

DOI：10.1016/j.neucom.2023.126519

原文链接

原文求助

摘要

En 中文

Event detection (ED) is of fundamental importance for many information extraction tasks. But conventional ED models require a fixed set of pre-defined event types. When an event trigger that cannot be categorized as any of these types occurs, they have to be re-trained from scratch with data from both old and new classes, so they are impractical in real life. In this paper, we propose a more realistic incremental event detection (IED) model that incrementally learns new event types at different times while not catastrophically forgetting the learned old classes. Although a knowledge distillation-based (KD-based) approach can be used to solve catastrophic forgetting, we find that conflicts arise when applying a general KD method to the IED task. Specifically, such conflicts mainly occur in two common scenarios: the first is that the triggers are labeled as a non-trigger word class in the current step, but at the same time, predicted as an old event type by the previous model (None-old class label confusion). The second is that the triggers are labeled as a new event type in the current step, but at the same time, predicted as a non-trigger word class by the previous model (None-new class label confusion). To solve the conflicts in each scenario, we generate pseudo labels and modify the distillation loss to improve the prediction accuracy on old and new classes, respectively. Comparative experiments demonstrate the effectiveness of our method, outperforming the state-of-the-art model.

Keyword:

Event detection

Incremental learning

Knowledge distillation

Pseudo labels

Catastrophic forgetting