Analysis of Retrieval Performance for Methods Fine-Tuned with ColBERT Architecture

Nowadays, knowledge-based retrieval from large datasets in healthcare is emerging as an important field of study. The growth and diversification of datasets increases the importance of artificial intelligence solutions that work without the need for human intervention. For healthcare professionals, fast and accurate presentation of information relevant to the user’s queries affects the success of information retrieval processes. This study aims to retrieve healthcare-specific information accurately and efficiently with a coding model-based architecture. ColBERT (Contextualized Late Interaction over BERT) architecture is preferred for speed and performance, and the models are fine-tuned with this architecture and adapted to the medical domain. The success of the fine-tuned models is evaluated. The findings show that model performance can be improved by using hard negative instance mining and hybrid score weighting techniques in medical information retrieval. In addition to offering the potential to improve the speed and accuracy of health information retrieval systems, the study is considered to be an important reference for future research in this field.