Zero- and Few-Shot Event Detection via Prompt-Based Meta Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7928–7943 July 9-14, 2023 ©2023 Association for Computational Linguistics Zero- and Few-Shot Event Detection via Prompt-Based Meta Learning Zhenrui Yue Huimin Zeng Mengfei Lan Heng Ji Dong Wang University of Illinois Urbana-Champaign {zhenrui3, huiminz3, mlan3, hengji, dwang24}@illinois.edu Abstract With emerging online topics as a source for numerous new events, detecting unseen / rare event types presents an elusive challenge for existing event detection methods, where only limited data access is provided for training. To address the data scarcity problem in event detec- tion, we propose MetaEvent, a meta learning- based framework for zero- and few-shot event detection. Specifically, we sample training tasks from existing event types and perform meta training to search for optimal parame- ters that quickly adapt to unseen tasks. In our framework, we propose to use the cloze- based prompt and a trigger-aware soft verbal- izer to efficiently project output to unseen event types. Moreover, we design a contrastive meta objective based on maximum mean discrep- ancy (MMD) to learn class-separating features. As such, the proposed MetaEvent can perform zero-shot event detection by mapping features to event types without any prior knowledge. In our experiments, we demonstrate the effec- tiveness of MetaEvent in both zero-shot and few-shot scenarios, where the proposed method achieves state-of-the-art performance in exten- sive experiments on benchmark datasets Few- Event and MAVEN. 1 Introduction Event detection tasks have experienced significant improvements thanks to the recent efforts in de- veloping language-based methods (Lu et al., 2021; Pouran Ben Veyseh et al., 2021). One of such meth- ods is pretrained large language models, which can be fine-tuned for detecting events upon input con- text (Liu et al., 2019; Cong et al., 2021). However, the detection of unseen or rare events remains a challenge for existing methods, as large amounts of annotated data are required for training in a su- pervised fashion (Shang et al., 2022; Zhang et al., 2022c). For instance, existing models often fail to detect unseen event types due to the lack of domain knowledge, as demonstrated in Figure 1. Weights Training Input (e.g. Protesters held a demonstration.) Prediction → Conflict: Demonstration Inference Input (e.g. The hunter shoot and kill a deer.) Prediction → Life: Die Trained Event Detection Model Deployed Event Detection Model Seen Events Rare / Unseen Events Figure 1: Existing event detection model fails to detect rare / unseen events (green input) upon deployment. To detect unseen event types (i.e., zero-shot learning), existing methods leverage external knowledge or contrastive learning to build class- separating features (Lyu et al., 2021; Zhang et al., 2022a,b). Similarly, it is possible to leverage lim- ited annotated examples (i.e., few-shot learning) for event detection. For instance, prototypical features can be constructed to match event types upon infer- ence (Deng et al., 2020; Cong et al., 2021). Prompt- based methods align with the pretraining objective of language models to improve detection perfor- mance (Li et al., 2022a,b). Specifically, a …

Content truncated to 500 words