Dynamic Prefix-Tuning for Generative Template-based Event Extraction

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5216 - 5228 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Dynamic Prefix-Tuning for Generative Template-based Event Extraction Xiao Liu1,2,3,4 and Heyan Huang1,2,3,4,∗and Ge Shi5 and Bo Wang1,2,3,4 1School of Computer Science and Technology, Beijing Institute of Technology 2Beijing Engineering Research Center of High Volume Language Information Processing and Cloud Computing Applications 3Key Laboratory of Intelligent Information Processing and Information Security, Ministry of Industry and Information Technology, China 4Southeast Academy of Information Technology, Beijing Institute of Technology 5Faculty of Information Technology, Beijing University of Technology {xiaoliu,hhy63,bwang}@bit.edu.cn, shige@bjut.edu.cn Abstract We consider event extraction in a generative manner with template-based conditional gen- eration. Although there is a rising trend of casting the task of event extraction as a se- quence generation problem with prompts, these generation-based methods have two signifi- cant challenges, including using suboptimal prompts and static event type information. In this paper, we propose a generative template- based event extraction method with dynamic prefix (GTEE-DYNPREF) by integrating con- text information with type-specific prefixes to learn a context-specific prefix for each con- text. Experimental results show that our model achieves competitive results with the state-of- the-art classification-based model ONEIE on ACE 2005 and achieves the best performances on ERE. Additionally, our model is proven to be portable to new types of events effectively. 1 Introduction Event extraction is an essential yet challenging task for natural language understanding. Given a piece of text, event extraction systems need to recognize event triggers with specific types and the event arguments with the correct roles in each event record according to an event ontology, which defines the event types and argument roles (Dod- dington et al., 2004; Ahn, 2006). As an example, the context in Figure 1 contains two event records, a Transport event triggered by “returned” and an Arrest-Jail event triggered by “capture”. In the Transport event, the Artifact is “the man”, the Destination is “Los Angeles” and the Origin is “Mexico”. In the Arrest-Jail event, the Person is “the man”, the Time is “Tuesday” and the Agent is “bounty hunters”. In this work, we focus on the task setting of extracting ∗Corresponding author. events without gold entity annotations, which is more practical in real-world applications. Most of the event extraction work treats the extraction of event triggers and event arguments as several classification tasks, either learned in a pipelined framework (Ji and Grishman, 2008; Liu et al., 2020; Du and Cardie, 2020; Li et al., 2020) or a joint formulation (Li et al., 2013; Yang and Mitchell, 2016; Nguyen et al., 2016; Liu et al., 2018; Wadden et al., 2019; Lin et al., 2020). There is a rising trend of casting the task of event extraction as a sequence generation problem by ap- plying special decoding strategies (Paolini et al., 2021; Lu et al., 2021) or steering pretrained lan- guage models to output conditional generation se- quences with discrete prompts (Li et al., 2021; Hsu et …

Content truncated to 500 words