COVID-19 information retrieval with deep-learning based semantic search, question answering, and abstractive...

Area of Science:

• Information Science
• Computer Science
• Medical Informatics

Background:

• The COVID-19 pandemic generated a vast amount of scientific literature.
• Over 400,000 coronavirus-related publications were collected in 2020 alone.
• Health workers require efficient tools to navigate this literature and find accurate information.

Purpose of the Study:

• To develop CO-Search, a semantic, multi-stage search engine for complex queries on COVID-19 literature.
• To assist health workers in finding scientific answers and avoiding misinformation.
• To improve information retrieval for scientific research during a crisis.

Main Methods:

• CO-Search employs a two-part system: a hybrid semantic-keyword retriever and a re-ranker.
• The retriever combines Siamese-BERT (deep learning) with BM25 and TF-IDF (keyword-based) models.
• The re-ranker utilizes question-answering and abstractive summarization modules for relevance scoring.
• A text augmentation technique was developed to enhance training data for the retriever.

Main Results:

• The system demonstrated strong performance on the TREC-COVID information retrieval challenge.
• CO-Search effectively handles complex queries over a large corpus of scientific documents.
• The hybrid approach balances semantic understanding with keyword importance.

Conclusions:

• CO-Search offers a robust solution for navigating the COVID-19 research landscape.
• The tool has the potential to significantly aid healthcare professionals in evidence-based decision-making.
• Semantic search engines can be valuable assets during public health emergencies.