A pattern-first pipeline approach for entity and relation extraction

Entity-relation extraction is the task of extracting entities and their semantic relations from a piece of unstructured text. In recent studies, Machine Reading Comprehension (MRC) based methods have been applied to this task and achieved significant results. As a pipelined approach, these methods always extract head entities first, and then identify related tail entities by enumerating each relationship. These entity-first methods will lead to the entity redundancy problem. They also suffer from the error propagation issue, which is an inherent issue of the multi-step inference process. Moreover, most existing MRC-based models, which use tagging-based methods for entity recognition, could not deal with overlapping entities. To address these, we propose Patti, a Pattern-First Pipeline Approach for Entity and Relation Extraction. Firstly, Patti leverages a novel MRC-based pattern classifier to identify relation patterns. Next, a span-based method was introduced to extract entities under the guidance of questions parameterized by the patterns yield in the first step. Finally, to alleviate the error propagation issue, Patti employs an additional MRC-based classifier to remove falsely extracted candidate entity-relation triples. Experiment results show that our approach significantly outperforms the entity-first baseline models on CoNLL04 and ACE05 datasets. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

Entity-relation extraction is the task of extracting entities and their semantic relations from unstructured text. Recent studies have shown that Machine Reading Comprehension (MRC) based methods achieve significant results in this task. However, traditional entity-first methods suffer from entity redundancy and error propagation. To address these issues, we propose a pattern-first pipeline approach.