Neural Gibbs Sampling for Joint Event Argument Extraction

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 169–180 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 169 Neural Gibbs Sampling for Joint Event Argument Extraction Xiaozhi Wang1∗, Shengyu Jia3∗, Xu Han1, Zhiyuan Liu1,2† , Juanzi Li1,2, Peng Li4, Jie Zhou4 1Department of Computer Science and Technology, BNRist; 2KIRC, Institute for Artificial Intelligence; 3Department of Electrical Engineering, Tsinghua University, Beijing, 100084, China 4Pattern Recognition Center, WeChat AI, Tencent Inc, China {wangxz20, jsy20, hanxu17}@mails.tsinghua.edu.cn Abstract Event Argument Extraction (EAE) aims at pre- dicting event argument roles of entities in text, which is a crucial subtask and bottleneck of event extraction. Existing EAE methods ei- ther extract each event argument roles inde- pendently or sequentially, which cannot ade- quately model the joint probability distribution among event arguments and their roles. In this paper, we propose a Bayesian model named Neural Gibbs Sampling (NGS) to jointly ex- tract event arguments. Specifically, we train two neural networks to model the prior distri- bution and conditional distribution over event arguments respectively and then use Gibbs sampling to approximate the joint distribution with the learned distributions. For overcom- ing the shortcoming of the high complexity of the original Gibbs sampling algorithm, we fur- ther apply simulated annealing to efficiently estimate the joint probability distribution over event arguments and make predictions. We conduct experiments on the two widely-used benchmark datasets ACE 2005 and TAC KBP 2016. The Experimental results show that our NGS model can achieve comparable results to existing state-of-the-art EAE methods. The source code can be obtained from https:// github.com/THU-KEG/NGS. 1 Introduction Event argument extraction (EAE) is a crucial sub- task of Event Extraction, which aims at predicting entities and their event argument roles in event mentions. For instance, given the sentence “Fox’s stock price rises after the acquisition of its entertain- ment businesses by Disney”, the event detection (ED) model will first identify the trigger word “ac- quisition” triggering a Transfer-Ownership event. Then, with the trigger word and event type, ∗ indicates equal contribution † Corresponding author: Z.Liu (liuzy@tsinghua.edu.cn) 's stock price rises after the acquisition of its by . Transfer- Ownership Fox entertainment businesses Disney Buyer Seller Artifact Text Event Argument Role Event Type acquisition Fox Trigger Word Argument Figure 1: An example of event extraction, including event detection and event argument extraction. the EAE model is required to identify that “Fox” and “Disney” are event arguments whose roles are “Seller” and “Buyer” respectively. As ED is well-studied in recent years (Liu et al., 2018a; Nguyen and Grishman, 2018; Zhao et al., 2018; Wang et al., 2019a), EAE becomes the bottleneck and has drawn growing attention. As EAE is the bottleneck of event extraction, especially is also important for various NLP ap- plications (Yang et al., 2003; Basile et al., 2014; Cheng and Erk, 2018), intensive efforts have al- ready been devoted to designing effective EAE sys- tems. The early feature-based methods (Patward- …

Content truncated to 500 words