Improving Cascade Decoding with Syntax-aware Aggregator and Contrastive Learning for Event Extraction

Improving Cascade Decoding with Syntax-aware Aggregator and Contrastive Learning for Event Extraction Zeyu Sheng , Yuanyuan Liang , Yunshi Lan∗ School of Data Science & Engineering, East China Normal University, Shanghai, China {51205903051,leonyuany}@stu.ecnu.edu.cn, yslan@dase.ecnu.edu.cn Abstract Cascade decoding framework has shown superior performance on event extraction tasks. How- ever, it treats a sentence as a sequence and neglects the potential benefits of the syntactic struc- ture of sentences. In this paper, we improve cascade decoding with a novel module and a self- supervised task. Specifically, we propose a syntax-aware aggregator module to model the syntax of a sentence based on cascade decoding framework such that it captures event dependencies as well as syntactic information. Moreover, we design a type discrimination task to learn better syn- tactic representations of different event types, which could further boost the performance of event extraction. Experimental results on two widely used event extraction datasets demonstrate that our method could improve the original cascade decoding framework by up to 2.2% percentage points of F1 score and outperform a number of competitive baseline methods. 1 Introduction As an important yet challenging task in natural language processing, event extraction has attracted much attention for decades (Chen et al., 2015; Nguyen and Grishman, 2018; Zheng et al., 2019; Lai et al., 2021; Wang et al., 2021; Li et al., 2022; Ma et al., 2022; Zhou et al., 2022). This task aims at predicting event types, triggers and arguments from a given sentence. We display three examples in Figure 1. Given an example sentence (a) “In 2018, Chuangwei Tech acquired equity of Qianhong Electronics for 1.5 billion ...”, an event extraction system is able to recognize the trigger “acquired”, that corresponds to the event type “invest”, and the argument “Chuangwei Tech”, that plays the subject role of “sub” in the event. A great number of methods have been developed for event extraction. Early methods formulate the event extraction as a sequence labeling task, where each token is considered as a candidate for labeling. They perform trigger extraction and argument extraction with joint learning (Li et al., 2013; Nguyen et al., 2016; Nguyen and Nguyen, 2019), which easily causes the label conflict issue. Considering the precedence relationship between the components in an event, pipeline methods are explored to perform trigger and argument extraction in separate stages (Chen et al., 2015; Du and Cardie, 2020; Liu et al., 2020; Ma et al., 2022). But the error is accumulated along with the pipeline. Recently, a cascade decoding framework (Xu et al., 2020; Sheng et al., 2021) is proposed to extract events with a cascade tagging strategy, which could not only handle the label conflict issue, but also avoid error propagation. In above methods, a sentence is treated as a sequence, and methods suffer from the low efficiency problem in capturing long-range dependency. We take sentence (a) in Figure 1 as an example. The argu- ment “1.5 billion” is far from the trigger “acquired” based on the sequential order while they are closely connected via …

Content truncated to 500 words