Answering List-Type Questions in Health Domain with Pretrained Large Language Model: A Case for COVID-19 Symptoms

Area of Science:

• Natural Language Processing
• Health Informatics

Background:

• List-type questions are common in health information seeking.
• Identifying COVID-19 symptoms from social media presents annotation challenges for supervised learning.

Purpose of the Study:

• To investigate the effectiveness of GPT-3 and few-shot learning for detecting symptom mentions in Twitter posts.
• To address the challenge of limited annotated data for health-related question answering.

Main Methods:

• Utilized GPT-3, a pre-trained large language model.
• Employed few-shot learning techniques (5-shot and 10-shot).
• Evaluated on a corpus of 655 annotated tweets for symptom detection.

Main Results:

• Few-shot learning with GPT-3 demonstrated effectiveness in identifying symptom mentions.
• Achieved promising results on the annotated tweet corpus.
• Indicated the viability of the approach with minimal annotation effort.

Conclusions:

• Few-shot learning combined with large language models like GPT-3 is a promising strategy for list-type question answering in the health domain.
• This approach can significantly reduce the effort required for data annotation.
• Facilitates automatic identification of health information, such as COVID-19 symptoms, from social media.