A Neural Text Ranking Approach for Automatic MeSH Indexing

The U.S. National Library of Medicine (NLM) has been indexing the biomedical literature with MeSH terms since the mid-1960s, and in recent years the library has increasingly relied on AI assistance and automation to curate the biomedical literature more efficiently. Since 2002, the NLM has been using natural language processing algorithms to assist indexers by providing MeSH term recommendations, and we are continually working to improve the quality of these recommendations. This work presents a new neural text ranking approach for automatic MeSH indexing. The domain-specific pretrained transformer model, PubMedBERT, was fine-tuned on MEDLINE data and used to rank candidate main headings obtained from a Convolutional Neural Network (CNN). Pointwise, listwise, and multi-stage ranking approaches are demonstrated, and the algorithm performance was evaluated by participating in the BioASQ challenge task 9a on semantic indexing. The neural text ranking approach was found to have very competitive performance in the final batch of the challenge, and the multi-stage ranking method typically boosted the CNN model performance by about 5% points in terms of micro F1-score.