Modeling Task Interactions in Document-Level Joint Entity and Relation Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5409 - 5416 July 10-15, 2022 ©2022 Association for Computational Linguistics Modeling Task Interactions in Document-Level Joint Entity and Relation Extraction Liyan Xu Jinho D. Choi Department of Computer Science Emory University, Atlanta, USA {liyan.xu,jinho.choi}@emory.edu Abstract We target on the document-level relation extrac- tion in an end-to-end setting, where the model needs to jointly perform mention extraction, coreference resolution (COREF) and relation extraction (RE) at once, and gets evaluated in an entity-centric way. Especially, we address the two-way interaction between COREF and RE that has not been the focus by previous work, and propose to introduce explicit inter- action namely Graph Compatibility (GC) that is specifically designed to leverage task char- acteristics, bridging decisions of two tasks for direct task interference. Our experiments are conducted on DocRED and DWIE; in addition to GC, we implement and compare different multi-task settings commonly adopted in previ- ous work, including pipeline, shared encoders, graph propagation, to examine the effective- ness of different interactions. The result shows that GC achieves the best performance by up to 2.3/5.1 F1 improvement over the baseline. 1 Introduction There has been a growing interest in document- level relation extraction recently since the introduc- tion of several large-scale datasets such as DocRED (Yao et al., 2019), which requires inter-sentence reasoning over the global entities and classifies relation instances on the entity-level, with each en- tity being a cluster of coreferent mentions across a document. In this line of entity-centric research, recent work has made great advancement on the global reasoning while regarding the entities as given (Nan et al., 2020; Zhou et al., 2021; Xu et al., 2021; Ru et al., 2021). Nevertheless, the more prac- tical end-to-end setting that extracts global entities and relations jointly has not drawn much attention, which poses extra burden to the model that needs to resolve mentions, coreference and relations at once. In this work, we specifically address this end-to-end setting such that given a document, the model targets to extract all gold triples (eh, et, r), where an instance is evaluated as correct only if the head/tail entity clusters (eh/et) as well as the relation r are all correct. To leverage the potentials that different tasks could benefit from each other, two popular methods have been taken by recent span-extraction-based models. One is to simply share the encoder (hence sharing mention representation) in the multi-task learning while decoding separately in a pipeline manner (Luan et al., 2018; Sanh et al., 2019). The other is to add graph propagation that enriches men- tion representation with task-specific decisions, e.g. DYGIE (Luan et al., 2019). However, the task interactions above only hap- pen on the representation level, and still employ the pipeline-like decoding, thus no explicit interactions have been made that directly interfere the decisions of different tasks. Meanwhile, the improvement from graph propagation has been diminished under strong encoders like BERT (Joshi et al., …

Content truncated to 500 words