Joint Extraction of Entities, Relations, and Events via Modeling Inter-Instance and Inter-Label Dependencies

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4363 - 4374 July 10-15, 2022 ©2022 Association for Computational Linguistics Joint Extraction of Entities, Relations, and Events via Modeling Inter-Instance and Inter-Label Dependencies Minh Van Nguyen1, Bonan Min2, Franck Dernoncourt3, and Thien Huu Nguyen1,4 1 Dept. of Computer and Information Science, University of Oregon, Eugene, OR, USA 2 Raytheon BBN Technologies, USA 3 Adobe Research, San Jose, CA, USA 4 VinAI Research, Vietnam {minhnv,thien}@cs.uoregon.edu, bonan.min@raytheon.com, dernonco@adobe.com Abstract Event trigger detection, entity mention recog- nition, event argument extraction, and relation extraction are the four important tasks in in- formation extraction that have been performed jointly (Joint Information Extraction - JointIE) to avoid error propagation and leverage depen- dencies between the task instances (i.e., event triggers, entity mentions, relations, and event arguments). However, previous JointIE models often assume heuristic manually-designed de- pendency between the task instances and mean- field factorization for the joint distribution of instance labels, thus unable to capture optimal dependencies among instances and labels to improve representation learning and IE per- formance. To overcome these limitations, we propose to induce a dependency graph among task instances from data to boost representation learning. To better capture dependencies be- tween instance labels, we propose to directly estimate their joint distribution via Conditional Random Fields. Noise Contrastive Estimation is introduced to address the maximization of the intractable joint likelihood for model training. Finally, to improve the decoding with greedy or beam search in prior work, we present Simu- lated Annealing to better find the globally opti- mal assignment for instance labels at decoding time. Experimental results show that our pro- posed model outperforms previous models on multiple IE tasks across 5 datasets and 2 lan- guages. 1 Introduction To extract structured information from unstructured text, a typical information extraction (IE) pipeline involves four major tasks: event trigger detection (ETD), event argument extraction (EAE), entity mention recognition (EMR), and relation extraction (RE). Previous work has performed such IE tasks via pipelined approaches (Li et al., 2013; Chen et al., 2015; Du and Cardie, 2020; Li et al., 2020), where a model for one task uses output predic- tions from other models performing other tasks. Consequently, errors from the predictions can be propagated between the models in the pipeline. Recently, ETD, EMR, EAE, and RE have been solved jointly in a single model, i.e., Joint Infor- mation Extraction - JointIE (Wadden et al., 2019; Lin et al., 2020; Nguyen et al., 2021a; Zhang and Ji, 2021), to avoid error propagation and leverage dependency between prediction instances of the four IE tasks (i.e., event trigger, entity mention, relation, and event argument candidates in a sen- tence). For example, if a Person entity mention is a Victim argument for a Die event, it is likely that the same entity mention is also a Target argument for an Attack event in the same sentence. To implic- itly exploit instance dependency for representation learning, Wadden et al. …

Content truncated to 500 words