A deep learning relation extraction approach to support a biomedical semi-automatic curation task: the case of the gluten bibliome

Abstract

Discover relevant biomedical interactions in the literature is crucial for enhancing biology research. This curation process has an essential role in studying the different processes and interactions reported that affect the biological process (e.g., genome, metabolome, and transcriptome). In this sense, the objective of this work is twofold: reduce the manual effort required to curate and review the existing biochemical interactions reported in the gluten-related bibliome, while proposing a novel vector-space integrated into a deep learning model to assists manual curators in a real curation task by learning from their previous decisions. With this objective, the present work proposes a novel vector-space that combine (i) high-level lexical and syntactic inference features as Wordnets and Health-related domain ontologies, (ii) unsupervised semantic resources as word embedding, (iii) semantic and syntactic sentence knowledge, (iv) abbreviation resolution support, (v) several state-of-the-art Named-entity recognition methods, and, finally, (vi) different feature construction and optimization techniques to support a semi-automatic curation workflow. Therefore, the application of the proposed workflow over a classified set of 2,451 relevant gluten-related documents produces a total of 8,349 relevant and 471,813 irrelevant relations distributed in thirteen domain health-related categories. Experimental results showed that the proposed workflow is a valuable approach for a semi-automatic relation extraction task. It was able to obtain satisfactory results in the early stages of a real-world curation task and saved manual annotation efforts by learning from the decisions made by manual curators in iterative annotation rounds. The average F.score for the proposed relation categories was 0.731, being the lowest F.score at 0.47 and the highest F.score at 0.929. The different resources used in this work as well as the manually curated corpus are public available on our GitHub repository.