IARNN-Based Semantic-Containing Double-Level Embedding Bi-LSTM for Question-and-Answer Matching

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Biomedical Informatics

Background:

• Traditional Question-Answering (Q&A) systems struggle with deep semantic understanding and accurate word segmentation in specialized domains like Chinese medicine.
• Existing methods often fail to fully leverage sentence semantics or are susceptible to errors introduced by Chinese word segmentation.
• Attention mechanisms can introduce feature deviation, impacting performance in complex Q&A tasks.

Purpose of the Study:

• To introduce a novel end-to-end framework, the semantic-containing double-level embedding Bi-LSTM model (SCDE-Bi-LSTM), for improved Chinese medical Q&A matching.
• To address limitations in current Q&A similarity calculations by incorporating deep semantic information.
• To mitigate errors from Chinese word segmentation and refine feature extraction in Q&A systems.

Main Methods:

• Developed a semantic-containing text similarity calculation method for the Q&A core module.
• Implemented a double-level embedding sentence representation to reduce word segmentation errors.
• Proposed an improved Bi-LSTM based algorithm for feature extraction to address attention mechanism deviations.
• Created extensive Chinese medical Q&A corpora for model validation.

Main Results:

• The SCDE-Bi-LSTM model significantly outperforms strong baselines on Chinese medical Q&A datasets.
• Achieved a top-1 accuracy improvement of up to 14%, reaching 79.15% on the medical corpora.
• Demonstrated the framework's applicability across different domains, including the insuranceQA dataset.

Conclusions:

• The proposed SCDE-Bi-LSTM framework effectively enhances Q&A matching accuracy in the Chinese medical domain by integrating semantic information and improving representation.
• The novel methods for similarity calculation, sentence embedding, and feature extraction contribute to superior performance over existing state-of-the-art approaches.
• The framework's robustness and adaptability suggest its potential for broader applications in specialized domain Q&A systems.