A study on pharmaceutical text relationship extraction based on heterogeneous graph neural networks.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Effective information extraction from pharmaceutical texts is crucial for clinical research.
• Traditional Chinese Medicine (TCM) texts present unique challenges due to streamlined sentences and complex, heterogeneous entity relationships.
• Current relationship extraction models often fail to capture the intricate associations between entities and relationships, leading to incomplete structured representations.

Purpose of the Study:

• To develop an advanced relationship extraction model tailored for the complexities of Traditional Chinese Medicine (TCM) text.
• To address the limitations of mainstream models in handling heterogeneous entities and complex semantic relationships within TCM literature.

Main Methods:

• Proposed a heterogeneous graph neural network (HGNN) model specifically adapted for TCM text relationship extraction.
• Utilized bidirectional encoder representation from transformers (BERT) for fine-tuned word embedding as model input.
• Constructed a heterogeneous graph network to link words, phrases, and relationship nodes, capturing hidden layer representations.
• Implemented a two-stage subject-object entity identification method with a binary classifier to pinpoint TCM entities and form relationship groups.

Main Results:

• The proposed HGNN model embedded with BERT achieved a precision of 86.99% and an F1-score of 87.40% on the TCM relationship extraction dataset.
• Demonstrated significant improvements of 8.83% in precision and 10.21% in F1-score compared to existing models like CNN, Bert-CNN, and Graph LSTM.

Conclusions:

• The developed heterogeneous graph neural network model effectively extracts complex relationships from Traditional Chinese Medicine texts.
• The model's superior performance highlights the potential of graph neural networks and BERT embeddings for advancing information extraction in specialized biomedical domains.