MTQA: Text-Based Multitype Question and Answer Reading Comprehension Model.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Text-based multitype question answering (MTQA) is a challenging area in reading comprehension.
• Existing MTQA models face difficulties in construction and performance due to complex corpora.
• There is a significant need for improved performance in MTQA models.

Purpose of the Study:

• To propose a novel text-based multitype question and answer reading comprehension model (MTQA).
• To enhance the model's ability to handle diverse question types and complex textual data.
• To improve the overall performance and generalization capabilities of MTQA systems.

Main Methods:

• The proposed MTQA model employs a multilayer transformer encoding and decoding structure.
• Specialized decoding headers were introduced for answer type prediction, fragment extraction, arithmetic operations, counting, and negation.
• High-performance ELECTRA checkpoints were utilized with secondary pretraining and an absolute loss function.

Main Results:

• The MTQA model demonstrated superior performance on the DROP and QUOREF corpora compared to existing state-of-the-art models.
• Experimental results indicate significant improvements in feature extraction capabilities.
• The model exhibited strong generalization capabilities across different datasets.

Conclusions:

• The proposed MTQA model effectively addresses the complexities of text-based multitype question answering.
• The novel architecture and training strategy lead to enhanced performance and generalization.
• This work contributes a significant advancement to the field of reading comprehension models.