Overview of BioASQ 2021-MESINESP track. Evaluation of advance hierarchical classification techniques for scientific literature, patents and clinical trials.
Abstract
There is a pressing need to exploit recent advances in natural language processing technologies, in particular language models and deep learning approaches, to enable improved retrieval, classification and ultimately access to information contained in multiple, heterogeneous types of documents. This is particularly true for the field of biomedicine and clinical research, where medical experts and scientists need to carry out complex search queries against a variety of document collections, including literature, patents, clinical trials or other kind of content like EHRs. Indexing documents with structured controlled vocabularies used for semantic search engines and query expansion purposes is a critical task for enabling sophisticated user queries and even cross-language retrieval. Due to the complexity of the medical domain and the use of very large hierarchical indexing terminologies, implementing efficient automatic systems to aid manual indexing is extremely difficult. This paper provides a summary of the MESINESP task results on medical semantic indexing in Spanish (BioASQ/ CLEF 2021 Challenge). MESINESP was carried out in direct collaboration with literature content databases and medical indexing experts using the DeCS vocabulary, a similar resource as MeSH terms. Seven participating teams used advanced technologies including extreme multilabel classification and deep language models to solve this challenge which can be viewed as a multi-label classification problem. MESINESP resources, we have released a Gold Standard collection of 243,000 documents with a total of 2179 manual annotations divided in train, development and test subsets covering literature, patents as well as clinical trial summaries, under a cross-genre training and data labeling scenario. Manual indexing of the evaluation subsets was carried out by three independent experts using a specially developed indexing interface called ASIT. Additionally, we have published a collection of large-scale automatic semantic annotations based on NER systems of these documents with mentions of drugs/medications (170,000), symptoms (137,000), diseases (840,000) and clinical procedures (415,000). In addition to a summary of the used technologies by the teams, this paper shows that there was a clear improvement in terms of the best scoring systems when compared to previous efforts and there was also a clear time saving of up to 67% when using pre-indexing with these systems compared to manual indexing of documents.
