A Hybrid Detection and Generation Framework with Separate Encoders for Event Extraction

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3163–3180 May 2-6, 2023 ©2023 Association for Computational Linguistics A Hybrid Detection and Generation Framework with Separate Encoders for Event Extraction Ge Shi∗⋆ Beijing University of Technology shige@bjut.edu.cn Yunyue Su∗ Computer Network Information Center, CAS suyunyue@cnic.cn Yongliang Ma Langboat mayongliang@langboat.com Ming Zhou Langboat zhouming@langboat.com Abstract The event extraction task typically consists of event detection and event argument extraction. Most previous work models these two subtasks with shared representation by multiple classi- fication tasks or a unified generative approach. In this paper, we revisit this pattern and propose to use independent encoders to model event de- tection and event argument extraction, respec- tively, and use the output of event detection to construct the input of event argument extrac- tion. In addition, we use token-level features to precisely control the fusion between two en- coders to achieve joint bridging training rather than directly reusing representations between different tasks. Through a series of careful ex- periments, we demonstrate the importance of avoiding feature interference of different tasks and the importance of joint bridging training. We achieved competitive results on standard benchmarks (ACE05-E, ACE05-E+, and ERE- EN) and established a solid baseline. 1 Introduction Event extraction has always been an important and challenging task in Natural Language Processing (NLP) (Sundheim, 1992). It aims to extract event triggers with specific types and event arguments with correct roles from unstructured plain texts into a structured form, which mostly describes “who, when, where, what, why” and “how” of real-world events that happened (Li et al., 2021a). For ex- ample, Figure 1 shows a Meet event, triggered by "met", which describes the Entity "Kelly" meet with another Entity "officials" in Place "Seoul". Previous studies can be roughly classified into classification-based and generation-based methods depending on the decoder used. Classification- based method usually divides EE into two sub- tasks: (1) Event Detection (ED), which identifies ∗Equal contribution. Yunyue Su contributes during her internship at Langboat. ⋆Corresponding author. event triggers and their types. (2) Event Argument Extraction (EAE) extracts the arguments and their corresponding roles for given event triggers and then models them as classification tasks, either learned in a pipeline framework or a joint formu- lation. Recently, generation-based event extrac- tion methods have emerged as an alternative to tra- ditional classification-based methods due to their better data-efficient and flexibility to include ad- ditional guidance. These methods take a sentence with discrete or continuous prompts as input and use BART-style backbone learning to summarize the sentence into a natural sentence based on a manually designed template. The template is com- posed of natural utterances describing argument role labels, which can provide rich label semantics, leading to great success in generation-based event extraction. However, most of these methods simultaneously learn shared representations for ED and EAE. As shown in previous works (Nguyen and Grishman, 2015; Lu et al., 2019), ED relies more on lexical (e.g., lemma, synonyms) and shallow syntactic fea- tures (e.g., pos tags, …

Content truncated to 500 words