Exploring unanswerability in machine reading comprehension: approaches, benchmarks, and open challenges

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Machine Reading Comprehension (MRC) systems typically assume questions have answers in the provided text.
• Real-world applications frequently present questions lacking answers within the given context.
• This limitation hinders the practical deployment of current MRC technologies.

Purpose of the Study:

• To comprehensively review existing methods for addressing unanswerable questions in MRC.
• To categorize these methods into model-agnostic and model-specific approaches.
• To identify critical gaps and challenges in current research.

Main Methods:

• Systematic literature review of techniques for handling unanswerable questions in MRC.
• Categorization of methods based on their approach (model-agnostic vs. model-specific).
• Analysis of relevant datasets and evaluation metrics used in the field.

Main Results:

• Existing methods are broadly classified into model-agnostic and model-specific strategies.
• Key techniques, datasets, and evaluation metrics are explored.
• Identified gaps highlight areas for future research and development.

Conclusions:

• Developing MRC systems that can reliably identify and handle unanswerable questions is crucial.
• Further research is needed to improve robustness and generalizability.
• Addressing unanswerable questions will enhance the practical utility of MRC systems.