CapsTM: capsule network for Chinese medical text matching

Area of Science:

• Natural Language Processing (NLP)
• Deep Learning
• Artificial Intelligence

Background:

• Text Matching (TM) is a core NLP task crucial for information retrieval, question answering, and more.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), have advanced TM but struggle with small datasets and feature structure preservation.
• Capsule networks offer a promising alternative to address CNN limitations in TM.

Purpose of the Study:

• To propose and evaluate CapsTM, a novel deep learning architecture for Text Matching.
• To leverage capsule networks to overcome the limitations of CNNs in handling small samples and preserving feature structures.
• To assess the effectiveness of CapsTM on a specialized dataset of Chinese medical questions.

Main Methods:

• Developed CapsTM, a five-layer neural network incorporating capsule networks.
• Employed Bidirectional Long Short-Term Memory (BiLSTM) and attention mechanisms for text representation and interaction.
• Extended the ESIM model by integrating a capsule layer prior to the prediction layer for enhanced feature extraction.

Main Results:

• A new corpus of 36,360 Chinese medical question pairs was created for evaluation.
• CapsTM achieved the highest F-score of 0.8666 on the test set, outperforming state-of-the-art methods.
• Experimental results validate the efficacy of CapsTM on the Chinese medical question matching task.

Conclusions:

• CapsTM demonstrates superior performance in Chinese medical question matching compared to existing approaches.
• The integration of capsule networks enhances the model's ability to capture complex textual relationships.
• CapsTM represents a significant advancement in applying deep learning to specialized text matching domains.