COBERT: COVID-19 Question Answering System Using BERT

Area of Science:

• Computational Biology
• Infectious Disease Epidemiology
• Information Retrieval

Background:

• The COVID-19 pandemic created an overwhelming volume of scientific literature, making it challenging to access timely information.
• Frontline workers, policymakers, and researchers require efficient methods to stay updated on COVID-19 research findings.

Purpose of the Study:

• To develop an automated system for answering complex queries using the Coronavirus Open Research Dataset Challenge (CORD-19) literature.
• To provide a reliable method for extracting precise and up-to-date scientific knowledge from a large corpus of COVID-19 research.

Main Methods:

• Proposed COBERT, a retriever-reader dual algorithmic system.
• Utilized a TF-IDF vectorizer to retrieve the top 500 relevant documents.
• Employed a Bidirectional Encoder Representations from Transformers (BERT) reader, pre-trained on SQuAD 1.1, to refine sentences and extract answers.

Main Results:

• The COBERT system successfully processed 59K coronavirus-related documents.
• The DistilBERT version of the system achieved high performance metrics.
• Achieved an Exact Match (EM) score of 80.6% and an F1 score of 87.3%.

Conclusions:

• COBERT offers an effective solution for navigating and extracting information from extensive scientific literature.
• The system aids in disseminating critical research findings to relevant stakeholders during public health crises.
• The retriever-reader approach demonstrates significant potential for scientific knowledge discovery in rapidly evolving fields.