Learning Constraints and Descriptive Segmentation for Subevent Detection

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5216–5226 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5216 Learning Constraints and Descriptive Segmentation for Subevent Detection Haoyu Wang1, Hongming Zhang2∗, Muhao Chen3 & Dan Roth1 1Department of Computer and Information Science, UPenn 2Department of Computer Science and Engineering, HKUST 3Department of Computer Science, USC {why16gzl,danroth}@seas.upenn.edu; hzhangal@cse.ust.hk; muhaoche@usc.edu Abstract Event mentions in text correspond to real- world events of varying degrees of granular- ity. The task of subevent detection aims to resolve this granularity issue, recognizing the membership of multi-granular events in event complexes. Since knowing the span of descrip- tive contexts of event complexes helps infer the membership of events, we propose the task of event-based text segmentation (EVENTSEG) as an auxiliary task to improve the learning for subevent detection. To bridge the two tasks together, we propose an approach to learning and enforcing constraints that cap- ture dependencies between subevent detection and EVENTSEG prediction, as well as guid- ing the model to make globally consistent in- ference. Specifically, we adopt Rectifier Net- works for constraint learning and then con- vert the learned constraints to a regulariza- tion term in the loss function of the neural model. Experimental results show that the proposed method outperforms baseline meth- ods by 2.3% and 2.5% on benchmark datasets for subevent detection, HiEve and IC, respec- tively, while achieving a decent performance on EVENTSEG prediction1. 1 Introduction Since real-world events are frequently conveyed in human languages, understanding their linguistic counterparts, i.e. event mentions in text, is of vi- tal importance to natural language understanding (NLU). One key challenge to understanding event mentions is that they refer to real-world events with varied granularity (Glavaš et al., 2014) and form event complexes (Wang et al., 2020). For ex- ample, when speaking of a coarse-grained event “publishing a paper”, it can involve a complex of ∗ This work was done when the author was visiting the University of Pennsylvania. 1Our code is publicly available at http://cogcomp. org/page/publication_view/950. Former Penn State football coach Jerry Sandusky posted (e1) bail Thursday after spending a night in jail following a new round of sex-abuse charges (e2) filed against him. Sandusky secured his release using (e3) $200,000 in real estate holdings and a $50,000 certified check provided (e4) by his wife, Dorothy, according to online court record … He was also charged (e5) last month with abusing eight boys, some on campus, over 15 years, allegations that were not immediately brought to the attention of authorities even though high-level people at Penn State apparently knew about them. In all, he faces more than 50 charges (e6). The scandal (e7) has resulted in the ousting (e8) of school President Graham Spanier and longtime coach Joe Paterno. Figure 1: An example of PARENT-CHILD relations and EVENTSEGs from the HiEve dataset (Glavaš et al., 2014). The blue and yellow segments denote the tex- tual spans of event complexes “posted” and “scandal” respectively. Curved arrows denote PARENT-CHILD relations within a text …

Content truncated to 500 words