Double-target self-supervised clustering with multi-feature fusion for medical question texts

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Medical question-and-answer data presents challenges due to unstructured text, short character limits, and irregular language.
• Single feature extraction models struggle to fully capture the content of medical texts.

Purpose of the Study:

• To develop an end-to-end text clustering model for medical questions.
• To enhance information representation in medical texts through multi-feature fusion.

Main Methods:

• Proposed a multi-feature fusion double-target self-supervised clustering (MF-DSC) model.
• Integrated term frequency and lexical semantics to generate weighted word vectors for deep learning.
• Employed a self-attention mechanism and two target functions (L cluster, L topic) for unified clustering.

Main Results:

• MF-DSC demonstrated superior performance over five other models.
• Achieved high scores in normalized mutual information (0.4346), adjusted Rand index (0.4934), average clustering accuracy (0.8649), and F1 (0.5737).

Conclusions:

• The MF-DSC model effectively clusters medical texts.
• Multi-feature fusion and self-supervised learning enhance the representation and clustering of medical question data.