Knowledge guided distance supervision for biomedical relation extraction in Chinese electronic medical records. (15th October 2022)

• Record Type: Journal Article
• Authors: Zhao, Qing; Xu, Dezhong; Li, Jianqiang; Zhao, Linna; Akhtar Rajput, Faheem
• Is Part Of: Expert systems with applications. Volume 204(2022)
• Publication Date: 2022-10-15
• Subjects: Distant supervision -- Biomedical relation extraction -- Entity-type alignment -- Relation embedding; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117606
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: The entity-type alignment is first proposed for coarse-grained relation extraction. The relation embeddings are learned and used as knowledge-enhanced features. A novel word attention mechanism derived from the relation translation is given. A bootstrap learning process is presented for fine-grained relation disambiguation. Abstract: The goal of biomedical relation extraction is to obtain structured information from electronic medical records by identifying relations among clinical entities. By integrating the advantages of unsupervised and semi-supervised learning, the distant supervision approach has achieved significant success for a relation extraction task without a large amount of labeled corpora. However, in many cases, the recognized entities from the Chinese clinical text are not defined in semantic knowledge base, which limits the application of distant supervision for biomedical relation extraction. This work proposes a Knowledge Guided Distance Supervision (KGDS) model for handling the biomedical relation extraction task in Chinese electronic medical records. To handle the unknown entities, entity-type alignment (instead of entity alignment in traditional distant supervision) is employed for extracting coarse-grained relations. Then, by learning the relation embeddings both from semantic knowledge base and electronic medical record dataset as knowledge-enhanced features, this work presents a knowledge-enhanced bootstrapping learning process for fine-grained relation disambiguation. The empirical experiments on the real-world dataset of electronic medical records illustrate that our KGDS model achieves the best performance comparing to other state-of-the-art models, thereby advancing the field of biomedical relation extraction from Chinese electronic medical records.