Introduction
In medical informatics, integrating external knowledge triples (e.g., <Amiodarone, Treats, Arrhythmia>) into natural language processing tasks is a common strategy to enhance model performance. However, medical literature is often "entangled," containing numerous mentions of drugs and conditions that may not be the primary focus of the study. Standard models frequently struggle with this "noise," treating all mentioned knowledge with equal weight. To address this, researchers have developed HiCapsRKL, a framework designed to determine the hierarchical association between a paper and a piece of knowledge, ranging from "Highly Relevant" to "Irrelevant."
Methodology: The HiCapsRKL Framework
The HiCapsRKL (Hierarchical Capsule Routing for Knowledge-Literature association) architecture addresses the challenge of feature entanglement through a multi-stage process.
1. Fragment Extraction and Encoding
The model first identifies high-signal text fragments:
- Relationship Correlation ($RCor$): Sentences containing both entities of a knowledge triple to determine if their relationship is actually described.
- Knowledge Importance ($KImp$): High-signal sections like the title, keywords, and the first sentence of the abstract.
These fragments, along with the full literature and the knowledge triple, are processed using a BERT language encoder to generate dense representations.
2. Capsule Routing: Decoupling Entangled Features
The core innovation of HiCapsRKL is the application of Capsule Routing to the $RCor$ and $KImp$ representations. Unlike standard pooling layers that compress data and lose structural information, capsule networks use iterative dynamic routing to "decouple" overlapping features. This allows the model to isolate specific signals indicating a strong relationship or high importance even in dense, multi-topic medical text.
3. Multi-task Learning
The framework is trained using a multi-task objective, simultaneously predicting $RCor$ labels, $KImp$ labels, and a final four-level relevance grade. This joint training ensures that the model learns to prioritize relationship-level evidence alongside general importance.
Datasets and Evaluation
The model was evaluated using data from the Chinese Medical Association (2019) and the Chinese Medical Knowledge Graph (CMeKG).
- Training: A weakly supervised set of nearly 50,000 pairs was generated using KNN clustering and Latent Dirichlet Allocation (LDA).
- Testing: Two manually labeled benchmarks were used: a Relevance Prediction set (1,050 pairs) and a Medical Literature Retrieval set (3,469 pairs).
Key Results and Insights
HiCapsRKL demonstrated significant improvements over baselines such as standard BERT, SiameseBERT, and various Knowledge Graph embedding methods (e.g., TransE, RotatE).
- Classification Performance: The model achieved a Micro-F1 of 64.9% and a Macro-F1 of 54.8%, outperforming the standard BERT baseline by over 5 percentage points.
- Retrieval Accuracy: In ranking tasks, it reached an NDCG@10 of 81.2%, effectively prioritizing the most relevant literature in search results.
- Ablation Findings: Studies showed that $RCor$ (relationship) fragments provide a stronger signal for relevance than $KImp$ (importance) fragments, highlighting the necessity of capturing how entities interact within the text rather than just their presence in key sections.
Conclusion
By leveraging capsule routing to filter noise and identify hierarchical associations, HiCapsRKL provides a robust solution for integrating structured knowledge into medical NLP. While capsule routing introduces higher computational complexity, its ability to handle feature entanglement makes it a powerful tool for processing dense medical literature. Future work will focus on applying these relevance filters to downstream tasks like automated diagnosis and medical summarization.
Summary updated