Integrate Candidate Answer Extraction with Re-Ranking for Chinese Machine Reading Comprehension.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Machine Reading Comprehension (MRC) is a challenging NLP task requiring machines to understand text and answer questions.
• Existing methods often use pipeline systems, leading to re-encoding issues and distribution shifts.

Purpose of the Study:

• To propose an efficient Chinese MRC method, IERC-AR, addressing limitations of traditional pipeline approaches.
• To improve the accuracy and efficiency of machine reading comprehension for Chinese text.

Main Methods:

• Developed an end-to-end model integrating candidate answer extraction and re-ranking modules.
• Utilized RoBERTa-WWM for precise word representations to handle polysemy and capture Chinese word-level features.
• Employed a self-attention mechanism in the re-ranking module to refine answer predictions.

Main Results:

• The proposed IERC-AR method demonstrated superior performance on the Les MMRC dataset.
• Achieved state-of-the-art results compared to existing MRC approaches.
• The end-to-end architecture mitigated re-encoding problems and ensured consistent data distribution.

Conclusions:

• IERC-AR effectively enhances machine reading comprehension for Chinese.
• The integrated approach and improved modules contribute to higher accuracy in MRC tasks.
• The study presents a robust solution for Chinese MRC challenges.