Handling polysemous triggers and arguments in event extraction: an adaptive semantics learning strategy with reward–penalty mechanism

Abstract

Event extraction (EE) is a complex natural language processing (NLP) task that aims at identifying and classifying triggers and arguments in raw text. The polysemy of triggers and arguments stands out as one of the key challenges affecting the precise extraction of events. Existing approaches commonly consider the semantic distribution of triggers and arguments to be balanced. However, the sample quantities of different semantics in the same trigger or argument vary in real-world scenarios, leading to a biased semantic distribution. The bias introduces two challenges: (1) low-frequency semantics is difficult to identify; (2) high-frequency semantics is often mistakenly identified. To tackle these challenges, we propose an adaptive learning method with the reward–penalty mechanism for balancing the semantic distribution in polysemous triggers and arguments. The reward–penalty mechanism balances the semantic distribution by enlarging the gap between the target and nontarget semantics by rewarding correct classifications and penalizing incorrect classifications. Additionally, we propose a sentence-level event situation awareness (SA) mechanism to guide the encoder to accurately learn the knowledge of events mentioned in the sentence, thereby enhancing target event semantics in the distribution of polysemous triggers and arguments. Finally, for various semantics in different tasks, we propose task-specific semantic decoders to precisely identify the boundaries of the predicted triggers and arguments for the semantics. Our experimental results on ACE2005 and its variants, along with the rich Entities, Relations, and Events (ERE), demonstrate the superiority of our approach over single-task and multi-task EE baselines.

1. Introduction

Events, serving as carriers of information, possess significant research value due to their elevated information content and rich semantic details. The accelerated evolution of the Internet and the emergence of numerous Internet applications have brought a large number of unstructured and fragmented text resources. How to quickly and accurately obtain structured target event information from these resources has always been a key and challenging problem for scholars engaged in the field of event extraction (EE). EE (Ahn, 2006; Chen et al., 2015; Lu D et al., 2023) is the task of identifying and classifying triggers and arguments from unstructured text based on the predefined event schema, as shown in Fig. 1. EE enables users to obtain information in a timely and intuitive manner on who (doer), when (time), where (place), how (artifact), whom (recipient), and what (event) occurred.

The extracted event can be widely used in downstream applications, such as event graph construction (Shu et al., 2021; Xu TY et al., 2022), recommendation systems (Cui ZJ et al., 2023; Xia et al., 2023), and decision aids (Anelli et al., 2022; You MS et al., 2023).

Many efforts have been devoted to EE. Earlier EE methodologies relied mainly on manually crafted multi-granularity features (Ji and Grishman, 2008; Hong et al., 2011; McClosky et al., 2011), which were labor-intensive. The emergence of deep learning techniques (Chen et al., 2015; Nguyen et al., 2016; Sha et al., 2018), capable of automatically learning features of tasks from extensive annotated data, has overcome the limitations of manual feature design. Recently, pre-trained language models (PLMs) (Yang S et al., 2019; Lin et al., 2020; Lu YJ et al., 2021; Liu X et al., 2022) with rich general language representations, such as BERT and RoBERTa, have become the backbone of EE systems, reducing the need for extensive annotated data.