AMPERE: AMR-Aware Prefix for Generation-Based Event Argument Extraction Model

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10976–10993 July 9-14, 2023 ©2023 Association for Computational Linguistics AMPERE: AMR-Aware Prefix for Generation-Based Event Argument Extraction Model I-Hung Hsu∗1 Zhiyu Xie∗3 Kuan-Hao Huang2 Premkumar Natarajan1 Nanyun Peng2 1Information Science Institute, University of Southern California 2Computer Science Department, University of California, Los Angeles 3Computer Science Department, Tsinghua University {ihunghsu, pnataraj}@isi.edu {khhuang, violetpeng}@cs.ucla.edu xiezy19@mails.tsinghua.edu.cn Abstract Event argument extraction (EAE) identifies event arguments and their specific roles for a given event. Recent advancement in generation- based EAE models has shown great perfor- mance and generalizability over classification- based models. However, existing generation- based EAE models mostly focus on problem re- formulation and prompt design, without incor- porating additional information that has been shown to be effective for classification-based models, such as the abstract meaning represen- tation (AMR) of the input passages. Incorpo- rating such information into generation-based models is challenging due to the heterogeneous nature of the natural language form prevalently used in generation-based models and the struc- tured form of AMRs. In this work, we study strategies to incorporate AMR into generation- based EAE models. We propose AMPERE, which generates AMR-aware prefixes for every layer of the generation model. Thus, the prefix introduces AMR information to the generation- based EAE model and then improves the gen- eration. We also introduce an adjusted copy mechanism to AMPERE to help overcome po- tential noises brought by the AMR graph. Comprehensive experiments and analyses on ACE2005 and ERE datasets show that AMPERE can get 4% − 10% absolute F1 score improve- ments with reduced training data and it is in general powerful across different training sizes. 1 Introduction Event argument extraction (EAE) aims to recog- nize event arguments and their roles in an event. For example, in Figure 1, EAE models need to ex- tract districts, u.s. supreme court, and washington and the corresponding roles — Plaintiff, Adjudi- cator, and Place for the Justice:Appeal event with trigger appeal. EAE has long been a challenging task in NLP, especially when training data is lim- ited (Wang et al., 2019; Ma et al., 2022). It is ∗The authors contribute equally. an important task for various downstream appli- cations (Zhang et al., 2020; Berant et al., 2014; Hogenboom et al., 2016; Wen et al., 2021; Wu et al., 2022). Recently, generation-based EAE models (Hsu et al., 2022a; Lu et al., 2021; Li et al., 2021; Paolini et al., 2021; Parekh et al., 2022) are proposed and have shown great generalizability and competitive performance compared to traditional classification- based methods (Chen et al., 2015; Ma et al., 2020; Hsu et al., 2022b; Fincke et al., 2022). However, existing generation-based EAE models mostly fo- cus on problem reformulation and prompt design without incorporating auxiliary syntactic and se- mantic information that is shown to be effective in classification-based methods (Huang et al., 2016; Xu and Huang, 2022; Huang et al., 2018; Ahmad et al., 2021; Veyseh et al., 2020). In this work, we …

Content truncated to 500 words