An Open-Set Semi-Supervised Multi-Task Learning Framework for Context Classification in Biomedical Texts

Abstract

ObjectiveIn biomedical research, knowledge about the relationships between entities, including genes, proteins, and drugs, is vital for unraveling the complexities of biological processes and intracellular pathway mechanisms. Natural language processing (NLP) and text mining methods have shown great success in biomedical relation extraction (RE). However, extracted relations often lack contextual information like cell type, cell line, and intracellular location, which are crucial components of biological knowledge. Previous studies have treated this problem as a post hoc context-relation association task, which is limited by the absence of a golden standard corpus, leading to error propagation and decreased model performance. To address these challenges, we created CELESTA (Context Extraction through LEarning with Semi-supervised multi-Task Architecture), a framework for biomedical context classification, applicable to both open-set and close-set scenarios. MethodsTo capture the inherent relationships between biomedical relations and their associated contexts, we designed a multi-task learning (MTL) architecture that seamlessly integrates with the semi-supervised learning (SSL) strategies during training. Our framework addresses the challenges caused by the lack of labeled data by assuming that the unlabeled data contain both in-distribution (ID) and out-of-distribution (OOD) data points. Further, we created a large-scale dataset consisting of five context classification tasks by curating two large Biological Expression Language (BEL) corpora and annotating them with our new entity span annotation method. We developed an OOD detector to distinguish between ID and OOD instances within the unlabeled data. Additionally, we utilized the data augmentation method combined with an external database to enrich our dataset, providing exclusive features for models during training process. ResultsWe conducted extensive experiments on the dataset, demonstrating the effectiveness of the proposed framework in significantly improving context classification and extracting contextual information with high accuracy. The newly created dataset and code used for this work are publicly available on GitHub (https://github.com/pitt-miskov-zivanov-lab/CELESTA).