MGRC: An End-to-End Multigranularity Reading Comprehension Model for Question Answering

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks excel in extractive question answering.
• Multistage matching models retrieve relevant text before extracting answers.
• Pipeline approaches suffer from error propagation, especially in sentence retrieval.

Purpose of the Study:

• To propose a novel end-to-end multigranularity reading comprehension model.
• To unify paragraph identification, sentence selection, and answer extraction.
• To address error propagation and improve feature learning in question answering.

Main Methods:

• Developed a unified framework for multigranularity reading comprehension.
• Implemented an end-to-end approach to mitigate error propagation.
• Utilized shared features across matching granularities for improved representation learning.

Main Results:

• The proposed model outperforms vanilla BERT and existing multistage methods on four large-scale datasets.
• Demonstrated significant improvements in extractive question answering.
• Ablation studies confirmed the effectiveness of individual model components.

Conclusions:

• The end-to-end multigranularity model effectively addresses limitations of pipeline approaches.
• Unified modeling enhances performance by leveraging shared features.
• The approach represents a significant advancement in reading comprehension models.