Pandemic Literature Search: Finding Information on COVID-19

Pandemic Literature Search: Finding Information on COVID-19 Vincent Nguyen1,2 Maciej Rybinski1 Sarvnaz Karimi1 Zhenchang Xing2 1CSIRO Data61, Sydney, Australia 2The Australian National University, Canberra, Australia {firstname.lastname}@csiro.au {zhenchang.xing}@anu.edu.au Abstract Finding information related to a pandemic of a novel disease raises new challenges for infor- mation seeking and retrieval, as the new infor- mation becomes available gradually. We inves- tigate how to better rank information for pan- demic information retrieval. We experiment with different ranking algorithms and propose a novel end-to-end method for neural retrieval, and demonstrate its effectiveness on the TREC COVID search.1 This work could lead to a search system that aids scientists, clinicians, policymakers and others in finding reliable an- swers from the scientific literature. 1 Introduction As COVID-19—an infectious disease caused by a coronavirus—led the world to a pandemic, a large number of scientific articles appeared in journals and other venues. In a span of five months, PubMed alone indexed over 60,000 articles matching coron- avirus related search terms such as SARS-CoV-2 or COVID-19. This volume of published mate- rial can be overwhelming. There is a need for effective search algorithms and question answer- ing systems to find relevant information and an- swers. In response to this need, an international challenge—TREC COVID Search (Roberts et al., 2020; Voorhees et al., 2020)—was organised by several institutions, such as NIST and Allen Insti- tute for AI, where research groups and tech com- panies developed systems that searched over scien- tific literature on coronavirus. Through an iterative setup organised in different rounds, participants are presented with several topics. The evaluations measure the effectiveness of these systems in find- ing the relevant articles containing answers to the questions in the topics. We propose a method that improves the sys- tems developed for the TREC-COVID challenge by 1We release our code for the neural index in GitHub: https://git.io/JkZ7I adopting a novel hybrid neural end-to-end approach for ranking of search results. Our method combines a traditional inverted index and word-matching re- trieval with a neural indexing component based on BERT architecture (Devlin et al., 2019). Our neural indexer leverages the Siamese network train- ing framework (Reimers and Gurevych, 2019) fine- tuned on an auxiliary task (unrelated to literature re- trieval) to produce universal sentence embeddings. This means that neural indexing can be performed offline for the entire document collection and does not need to be retrained on additional queries. This allows for incorporating the neural component for the entire retrieval process, contrasting with the typ- ical multi-stage neural re-ranking approaches (Li et al., 2020; Zhang et al., 2020; Liu et al., 2017; Wang et al., 2011). Our method is competitive with the top systems presented in TREC COVID 2. It improves as cor- pus size increases despite not being trained on addi- tional data which is a useful property in pandemic information retrieval. 2 Related Work The use of neural networks in search has mostly been limited to reranking top results retrieved by a ‘traditional’ ranking mechanism, such as Okapi BM25 (Robertson et al., …

Content truncated to 500 words