KenMeSH: Knowledge-enhanced End-to-end Biomedical Text Labelling

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2941 - 2951 May 22-27, 2022 c⃝2022 Association for Computational Linguistics KenMeSH: Knowledge-enhanced End-to-end Biomedical Text Labelling Xindi Wang1,3, Robert E. Mercer1,3, and Frank Rudzicz2,3,4 1Department of Computer Science, University of Western Ontario, London, Ontario, Canada 2Department of Computer Science, University of Toronto, Toronto, Ontario, Canada 3Vector Institute for Artificial Intelligence, Toronto, Ontario, Canada 4 Unity Health Toronto, Toronto, Ontario, Canada xwang842@uwo.ca, mercer@csd.uwo.ca, frank@cs.toronto.edu Abstract Currently, Medical Subject Headings (MeSH) are manually assigned to every biomedical ar- ticle published and subsequently recorded in the PubMed database to facilitate retrieving relevant information. With the rapid growth of the PubMed database, large-scale biomedi- cal document indexing becomes increasingly important. MeSH indexing is a challenging task for machine learning, as it needs to as- sign multiple labels to each article from an extremely large hierachically organized col- lection. To address this challenge, we pro- pose KenMeSH, an end-to-end model that combines new text features and a dynamic Knowledge-enhanced mask attention that inte- grates document features with MeSH label hi- erarchy and journal correlation features to in- dex MeSH terms. Experimental results show the proposed method achieves state-of-the-art performance on a number of measures. 1 Introduction The PubMed1 database is a resource that provides access to the MEDLINE bibliographic database of references and abstracts together with the full text articles of some of these citations which are available in the PubMed Central2 (PMC) repository. MEDLINE3 contains more than 28 million refer- ences (as of Feb. 2021) to journal articles in the biomedical, health, and related disciplines. Jour- nal articles in MEDLINE are indexed according to Medical Subject Headings (MeSH)4, an hier- archically organized vocabulary that has been de- veloped and maintained by the National Library of Medicine (NLM)5. Currently, there are 29,369 main MeSH headings, and each MEDLINE citation 1https://pubmed.ncbi.nlm.nih.gov/about/ 2https://en.wikipedia.org/wiki/PubMed_Central 3https://www.nlm.nih.gov/medline/medline_overview. html 4https://www.nlm.nih.gov/mesh/meshhome.html 5https://www.nlm.nih.gov has 13 MeSH indices, on average. MeSH terms are distinctive features of MEDLINE and can be used in many applications in biomedical text mining and information retrieval (Lu et al., 2008; Huang et al., 2011; Gu et al., 2013), being recognized as impor- tant tools for research (e.g., knowledge discovery and hypothesis generation). Currently, MeSH indexing is done by human annotators who examine full articles and assign MeSH terms to each article according to rules set by NLM6. Human annotation is time consuming and costly – the average cost of annotating one article in MEDLINE is about $9.40 (Mork et al., 2013). Nearly 1 million citations were added to MEDLINE in 2020 (approximately 2,600 on a daily basis)7. The rate of articles being added to the MEDLINE database is constantly increasing, so there is a huge financial and time-consuming cost for the status quo. Therefore, it is imperative to develop an automatic annotation system that can assist MeSH indexing of large-scale biomedical articles efficiently and accurately. Automatic MeSH indexing can be regarded as an extreme multi-label text classification (XMC) prob- lem, where …

Content truncated to 500 words