MUSIED: A Benchmark for Event Detection from Multi-Source Heterogeneous Informal Texts

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2947 - 2964 December 7-11, 2022 ©2022 Association for Computational Linguistics MUSIED: A Benchmark for Event Detection from Multi-Source Heterogeneous Informal Texts Xiangyu Xi1∗, Jianwei Lv1∗, Shuaipeng Liu1∗†, Wei Ye2†, Fan Yang1, Guanglu Wan1 1 Meituan Group, Beijing, China 2 National Engineering Research Center for Software Engineering, Peking University, Beijing, China liushuaipeng@meituan.com,wye@pku.edu.cn Abstract Event detection (ED) identifies and classifies event triggers from unstructured texts, serving as a fundamental task for information extrac- tion. Despite the remarkable progress achieved in the past several years, most research efforts focus on detecting events from formal texts (e.g., news articles, Wikipedia documents, fi- nancial announcements). Moreover, the texts in each dataset are either from a single source or multiple yet relatively homogeneous sources. With massive amounts of user-generated text accumulating on the Web and inside enterprises, identifying meaningful events in these infor- mal texts, usually from multiple heterogeneous sources, has become a problem of significant practical value. As a pioneering exploration that expands event detection to the scenarios involving informal and heterogeneous texts, we propose a new large-scale Chinese event detec- tion dataset based on user reviews, text con- versations, and phone conversations in a lead- ing e-commerce platform for food service. We carefully investigate the proposed dataset’s tex- tual informality and multi-source heterogene- ity characteristics by inspecting data samples quantitatively and qualitatively. Extensive ex- periments with state-of-the-art event detection methods verify the unique challenges posed by these characteristics, indicating that multi- source informal event detection remains an open problem and requires further efforts. Our benchmark and code are released at https: //github.com/myeclipse/MUSIED. 1 Introduction Event detection (ED), which aims to identify event triggers and classify them into specific types from unstructured texts, has been widely researched and applied in various downstream tasks (Basile et al., 2014; Cheng and Erk, 2018; Kuhnle et al., 2021). Advanced models have been continuously pro- posed, ranging from feature-based models (Shasha *The first three authors contributed equally. †Corresponding authors. et al., 2010; Hong et al., 2011; Li et al., 2013) to recent neural-based models (Chen et al., 2015; Nguyen et al., 2016; Chen et al., 2018; Xi et al., 2021; Xiangyu et al., 2021). Despite the signifi- cant progress, we find that previous works have the following two limitations in practical scenarios. 1. Current efforts mainly focused on event detection from formal texts. For example, a pop- ular line of works (Li et al., 2013; Chen et al., 2015; Nguyen et al., 2016; Chen et al., 2018; Lou et al., 2021) aim to detect general domain events from news articles (e.g., ACE 2005 (Doddington et al., 2004)) and Wikipedia documents (e.g., MAVEN (Wang et al., 2020b)). Some other explorations involve extracting events from the financial an- nouncements (Yang et al., 2018; Zheng et al., 2019; Liang et al., 2021) or cybersecurity articles(Trong et al., 2020), which are also written in a relatively official style. In practical scenarios, however, we usually face the bottleneck of identifying …

Content truncated to 500 words