Enhancing Text-to-Knowledge Graph Conversion for Biomedical Data Using BiLSTM-CRF Models

Named Entity Recognition (NER) and Relation Extraction (RE) in biomedical text data present inherent challenges, including overlapping entities, extended entity spans, ambiguity from similar entities, and the classification of complex, heterogeneous relationships between recognized entity pairs across diverse textual contexts. This paper proposes techniques employing a Bidirectional Long Short-Term Memory (BiLSTM) model enhanced with a Conditional Random Field (CRF). The study explores various design configurations for NER and RE, like joint learning, transfer learning, and task-specific learning. These models for NER and RE tasks are experimented on ChemProt dataset and during the testing phase, the models demonstrated high precision confirming effectiveness in knowledge graph construction. The proposed approach has the potential to be extended to other domains that require advanced text-to-graph conversion techniques and complex data handling requirements.