Bi-Directional Iterative Prompt-Tuning for Event Argument Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6251 - 6263 December 7-11, 2022 ©2022 Association for Computational Linguistics Bi-Directional Iterative Prompt-Tuning for Event Argument Extraction Lu Dai1, Bang Wang1, Wei Xiang 1, Yijun Mo2 ∗ 1School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, China 2School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China {dailu18, wangbang, xiangwei, moyj}@hust.edu.cn Abstract Recently, prompt-tuning has attracted growing interests in event argument extraction (EAE). However, the existing prompt-tuning methods have not achieved satisfactory performance due to the lack of consideration of entity in- formation. In this paper, we propose a bi- directional iterative prompt-tuning method for EAE, where the EAE task is treated as a cloze- style task to take full advantage of entity in- formation and pre-trained language models (PLMs). Furthermore, our method explores event argument interactions by introducing the argument roles of contextual entities into prompt construction. Since template and ver- balizer are two crucial components in a cloze- style prompt, we propose to utilize the role label semantic knowledge to construct a seman- tic verbalizer and design three kinds of tem- plates for the EAE task. Experiments on the ACE 2005 English dataset with standard and low-resource settings show that the proposed method significantly outperforms the peer state- of-the-art methods. Our code is available at https://github.com/HustMinsLab/BIP. 1 Introduction As a key step of event extraction, event argument extraction refers to identifying event arguments with predefined roles. For example, for an "At- tack" event triggered by the word "fired" in the sen- tence "Iraqis have fired sand missiles and AAA at aircraft", EAE aims to identify that "Iraqis", "mis- siles", "AAA" and "aircraft" are event arguments with the "Attacker", "Instrument", "Instrument" and "Target" roles, respectively. In order to exploit the rich linguistic knowledge contained in pre-trained language models, fine- tuning methods have been proposed for EAE. The paradigm of these methods is to use a pre-trained language model to obtain semantic representations, ∗ Corresponding author: Yijun Mo Iraqis have fired sand missiles and AAA at aircraft Pre-trained Language Model neural network Role Attacker entity mention representation event trigger representation (a) Fine-Tuning for EAE Masked Language Model Verbalizer Iraqis have fired sand missiles and AAA at aircraft [SEP] T(fired, Iraqis, [MASK]) Role Attacker Word attacker (b) Prompt-Tuning for EAE. Figure 1: Illustration of fine-tuning and prompt-tuning methods for predicting the argument role of the entity mention "Iraqis" in the event triggered by the word "fired". and then feed these representations into a well- designed neural network to extract event arguments. For example in Figure 1(a), an event trigger rep- resentation and an entity mention representation are first obtained through a pre-trained language model, and then input to a designed neural net- work, such as hierarchical modular network (Wang et al., 2019) and syntax-attending transformer net- work (Ma et al., 2020), to determine the argument role that the entity mention plays in the event trig- gered by the trigger. …

Content truncated to 500 words