Knowledge Extraction in the Biomedical Domain using Deep Learning Techniques

Abstract

Knowledge Extraction (KE) transforms raw data into structured, actionable informa tion. Relation Extraction (RE), a key component of KE, identifies interactions between entities in text. In the biomedical field, RE is crucial due to the rapid growth of scientific publications, health records, and clinical databases. Structured data from these sources can enhance diagnostics, optimize treatments, and advance disease understanding. How ever, manual extraction is labor-intensive and error-prone, while traditional methods lack precision and scalability. Deep learning models, particularly Transformer-based archi tectures, offer promising solutions by enabling fast and accurate processing of complex biomedical data, supporting better decision-making and research advancements. This research aims to explore the application of Transformer-based techniques, with a particular focus on the PubMedBERT model, for extracting relations from biomedical texts. The primary objective is to design a high-performance system capable of identifying and structuring relationships between biomedical entities with great precision. The proposed Enriched-PubMedBERT model achieved robust performance through f ine-tuning and an entity enrichment strategy that integrates external biomedical knowl edge. By combining PubMedBERT’s textual representations generated by PubMedBERT with embeddings derived from a drug interaction graph, the model captures both global context and entity-specific details, achieving an F1 score of 95.33%, surpassing existing models.