Enhancing Biomedical Entity Normalization Through Re-ranking Based on Convolutional Neural Networks

A significant number of researchers have devoted considerable effort to extracting knowledge from biomedical literature in order to study the interactions between biomedical entities, such as chemicals and diseases. One crucial task in biomedical text mining is linking entity mentions to its standardized concept identifiers. To address this issue, we propose a biomedical named entity normalization model that is based on a convolution neural network (CNN) for re-ranking. Our method integrates semantic and morphological information, combining abbreviation resolution, dictionary matching, and similarity-based candidate generation with CNN re-ranking. We evaluate our model on two benchmark datasets, i.e. the BioCreative V - Chemical Disease Relation (BC5 CDR) and NCBI Disease. Our approach achieves state-of-the-art disease normalization performance with F1 scores of 92.2% and 84.4% on these two datasets respectively, outperforming existing methods. This work contributes to more accurate and efficient extraction of biomedical knowledge from literature.