Knowledge Discovery in COVID-19 Research Literature

Proceedings of Recent Advances in Natural Language Processing, pages 402–410 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_046 402 Knowledge Discovery in COVID-19 Research Literature Ernesto L. Estevanell-Valladares1, Alejandro Piad-Morffis1, Suilan Estevez-Velarde1, Yoan Gutierrez2, Andres Montoyo2, Rafael Mu˜noz2, and Yudivi´an Almeida-Cruz1 1School of Math and Computer Science, University of Havana {estevanell,apiad,sestevez,yudivian}@matcom.uh.cu 2Department of Languages and Computing Systems, University of Alicante {ygutierrez,montoyo,rafael}@dlsi.ua.es Abstract This paper presents the preliminary results of an ongoing project that analyzes the growing body of scientific research published around the COVID-19 pandemic. In this research, a general-purpose semantic model is used to double annotate a batch of 500 sentences that were manually selected from the CORD-19 corpus. Afterwards, a baseline text-mining pipeline is designed and evaluated via a large batch of 100, 959 sentences. We present a qual- itative analysis of the most interesting facts automatically extracted and highlight possi- ble future lines of development. The prelimi- nary results show that general-purpose seman- tic models are a useful tool for discovering fine-grained knowledge in large corpora of sci- entific documents. 1 Introduction The COVID-19 pandemic that affected almost ev- ery country at the beginning of 2020 has provoked a massive increase in scientific papers related to biomedical sciences (Velavan and Meyer, 2020). The scientific community’s effort to fight the spread of the virus is evident in the record number of research papers about COVID-19 that have been submitted to conferences, journals and preprint ser- vices. Several academic publishers joined the effort by providing free access to research in related ar- eas that could be useful to scientists and academics. The amount of information produced during this period greatly surpassed the ability of human re- searchers to stay up-to-date, which in turn spawned an increased interest in the application of compu- tational techniques to automatically organize, nor- malize and link the existing information. A recent initiative is the COVID-19 Open Re- search Dataset (CORD-19), published by the Allen Institute for AI (Lo et al., 2020), which makes available a large corpus of scientific papers on COVID-19 and related topics. At the moment of writing, it contains 76, 674 scientific articles. The corpus has been used as part of a Kaggle chal- lenge1 , which focused mainly on unsupervised tasks related to organizing and categorizing the dif- ferent aspects of the whole COVID-19 situation. Based on these resources, computational tools, e.g., SciSight (Hope et al., 2020), SciFact (Wadden et al., 2020), and similar (Bras et al., 2020), have been created to enable the interactive visualization and exploration of the scientific literature and the dis- covery of connections between the available meth- ods, symptoms, interventions, etc. Unsupervised approaches are a natural strategy for dealing with large, unlabeled corpora, while supervised approaches have the caveat of requiring training examples to be manually annotated, but can provide precise answers to specific questions given enough supervised data. For example, identi- fying domain-specific entities such as symptoms, medication and treatments, and semantic relations between them. Learning to recognize this type of information in natural language, even academic lan- …

Content truncated to 500 words