Efficient Machine Reading Comprehension for Health Care Applications: Algorithm Development and Validation of a...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Extractive machine reading comprehension (MRC) models excel in open domains but struggle with complex, large-context areas like healthcare.
• Longer contexts lead to decreased accuracy and slower predictions in MRC models.
• Reducing input context by extracting only relevant information is a potential solution.

Purpose of the Study:

• To develop an effective context extraction method for MRC tasks.
• To enable MRC models to process long articles more efficiently and accurately.
• To enhance question-answering capabilities in specialized domains.

Main Methods:

• Developed a novel method to estimate sentence utility for answering questions within a given context.
• Trained two models to predict sentence utility based on empirical studies and MRC model confidence scores.
• Extracted a shorter, more precise context for the MRC model based on sentence utility estimations.

Main Results:

• Demonstrated effectiveness on COVID-19 and biomedical QA datasets.
• Reduced inference time by 6-7 times.
• Improved MRC model accuracy, with F1-scores increasing from 0.724 to 0.744 (COVID-19) and 0.651 to 0.704 (biomedical).

Conclusions:

• The proposed context extraction method enhances MRC prediction accuracy and significantly reduces inference time.
• This technique is compatible with any MRC model and applicable to tasks involving extensive text processing.
• Potential challenges exist where extractive transformer MRC models may still underperform even with precise contexts.