Abstract
Relation extraction (RE) is a critical step in transforming unstructured clinical medical records into structured knowledge. While modern English-language clinical notes have been extensively studied, Traditional Chinese Medicine (TCM) records present unique challenges due to linguistic complexity and dense entity networks. This paper proposes MVG2RE, a novel multi-view heterogeneous graph representation learning method. By constructing multiple medical entity graphs based on co-occurrence, lexical patterns, and external domain knowledge, the model moves beyond local sentence-level extraction to a global inference approach. Experimental results on Chinese clinical records demonstrate that MVG2RE achieves a significant 12.6% improvement in F1-score over state-of-the-art methods, reaching a Precision of 86.38% and a Recall of 81.42%.
Introduction: The Challenge of TCM Clinical Records
Extracting structured information from Electronic Medical Records (EMRs) is essential for clinical decision support. However, Chinese clinical records, particularly those in TCM, are notoriously difficult to process. Unlike English, Chinese text lacks word delimiters, leading to segmentation errors. Furthermore, TCM narratives are often telegraphic and weakly ordered, making traditional sentence-level RE—which relies on clear grammatical structures—ineffective.
As shown above, a single clinical record involves an intricate web of entities including herbs, formulas, symptoms, syndromes, and diseases. A relationship between a herb and a symptom often cannot be understood in isolation; it requires context from the entire patient history and broader medical knowledge.
Proposed Method: MVG2RE
The authors introduce MVG2RE (Multi-View Graph model to extract RE), which redefines relation extraction as a link prediction problem on a heterogeneous graph.
Multi-View Graph Construction
To capture diverse evidence, the model constructs three distinct "views":
- Co-occurring View: Maps entities that appear together in the same record, capturing broad topological relationships.
- Lexical View: Analyzes the local context (surrounding words) of entity pairs to identify specific linguistic patterns indicative of a relationship.
- Semantic View: Leverages word embeddings and external medical Knowledge Graphs (KGs). By enriching entity embeddings with structured domain knowledge, the model gains a deeper "understanding" of medical concepts.
Representation Learning and Classification
The core architecture utilizes a Heterogeneous Graph Convolutional Network (GCN) with a dual-level attention mechanism. Node-level attention identifies relevant neighbors, while type-level attention weighs the importance of different entity categories (e.g., prioritizing "symptom" nodes when analyzing a "disease" node).
As illustrated in the architecture diagram, the model aggregates these views into a low-dimensional vector space. A softmax-normalized weighting system then produces the final representation used for relation classification.
Experimental Results
The model was evaluated using the TCMRelExtr dataset, which contains thousands of labeled instances across five relation types (e.g., herb-syndrome, syndrome-disease).
MVG2RE significantly outperformed existing baselines:
- Precision: 86.38%
- Recall: 81.42%
- F1-Score: 84.07%
This performance represents a 12.6% improvement in F1-score over previous state-of-the-art models. Ablation studies confirmed that the multi-view approach is far superior to any single view, proving that co-occurrence, context, and semantics provide complementary information.
Conclusion
MVG2RE demonstrates the power of global, graph-based inference for complex relation extraction tasks. By integrating multi-view evidence and external medical knowledge, the model successfully navigates the linguistic and structural hurdles of Chinese TCM records. Future work will explore incorporating edge weights to reflect evidence strength and developing joint models for simultaneous entity recognition and relation extraction.
Summary updated