Multistage BiCross encoder for multilingual access to COVID-19 health information

Area of Science:

• Information Science
• Computer Science
• Public Health

Background:

• The COVID-19 pandemic created an "infodemic" of online health information.
• A need arose for accurate semantic search to find reliable COVID-19 data.

Purpose of the Study:

• To develop a high-precision, high-recall system for multilingual COVID-19 information retrieval.
• To address the challenge of semantic search across millions of documents.

Main Methods:

• Proposed a novel neural Multistage BiCross encoder.
• Implemented a three-stage ranking pipeline using Okapi BM25, bi-encoder, and cross-encoder.
• Participated in the Multilingual Information Access (MLIA) shared task.

Main Results:

• The approach achieved high precision and recall in document ranking.
• Demonstrated superior performance compared to state-of-the-art methods.
• Validated through independent evaluation in both monolingual and bilingual runs.

Conclusions:

• The Multistage BiCross encoder is effective for multilingual semantic search of COVID-19 information.
• The system successfully navigates the complexities of the COVID-19 infodemic.
• This method offers a reliable solution for accessing critical health information.