Adversarial Training for Weakly Supervised Event Detection

Proceedings of NAACL-HLT 2019, pages 998–1008 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 998 Adversarial Training for Weakly Supervised Event Detection Xiaozhi Wang1∗, Xu Han1∗, Zhiyuan Liu1† , Maosong Sun1, Peng Li2 1Department of Computer Science and Technology, Tsinghua University, Beijing, China Institute for Artificial Intelligence, Tsinghua University, Beijing, China State Key Lab on Intelligent Technology and Systems, Tsinghua University, Beijing, China 2Pattern Recognition Center, WeChat, Tencent Inc, China {xz-wang16, hanxu17}@mails.tsinghua.edu.cn {liuzy, sms}@tsinghua.edu.cn patrickpli@tencent.com Abstract Modern weakly supervised methods for event detection (ED) avoid time-consuming human annotation and achieve promising results by learning from auto-labeled data. However, these methods typically rely on sophisticated pre-defined rules as well as existing instances in knowledge bases for automatic annotation and thus suffer from low coverage, topic bias, and data noise. To address these issues, we build a large event-related candidate set with good coverage and then apply an adversar- ial training mechanism to iteratively identify those informative instances from the candi- date set and filter out those noisy ones. The experiments on two real-world datasets show that our candidate selection and adversarial training can cooperate together to obtain more diverse and accurate training data for ED, and significantly outperform the state-of-the- art methods in various weakly supervised sce- narios. The datasets and source code can be obtained from https://github.com/ thunlp/Adv-ED. 1 Introduction Event detection (ED) aims at detecting event trig- gers, which are often words or phrases evoking events in instances, and then identifying their spe- cific event types. For example, we can extract the trigger “married” of the event “Marry” from the text “Mark Twain and Olivia Langdon married in 1870”. Detecting and identifying events is an important subtask of event extraction and also beneficial for various downstream NLP applications, such as question answering (Yang et al., 2003), information retrieval (Basile et al., 2014), and reading comprehension (Cheng and Erk, 2018). Hence, many efforts have been de- voted to detecting event triggers and types. ∗ indicates equal contribution † Corresponding author: Z.Liu(liuzy@tsinghua.edu.cn) Most prior methods for ED are based on feature engineering, such as the token-level features (Ahn, 2006; Ji and Grishman, 2008) and the structured features (Li et al., 2013; Araki and Mitamura, 2015). As the rapid development of neural net- works, various neural models have been proposed to directly embed textual semantic information into a low-dimensional space and then detect event triggers based on those feature vectors (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016; Ghaeini et al., 2016; Feng et al., 2016). These methods follow a supervised learning ap- proach to train models on human-annotated data, and their requirement of human-annotated data is a bottleneck in practice. Considering weak su- pervision is widely adopted to take full advan- tages of large-scale raw data, especially some spe- cific work for information extraction (Mintz et al., 2009; Riedel et al., 2010; Zeng et al., 2015; Cao et al., 2018), weak supervision has been explored to automatically label training data for ED …

Content truncated to 500 words