Patient address parsing via KG-aware contrastive learning and constrained on-prem LLM inference

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Address parsing is crucial for large-scale information systems but faces challenges with ambiguity and privacy.
• Existing methods struggle with semantic and structural issues, hallucination, and deployment constraints.

Purpose of the Study:

• To introduce AddrKG-LLM, a novel framework for accurate and controllable address parsing.
• To address limitations of current approaches in handling noisy, abbreviated free-text addresses.

Main Methods:

• A two-stage framework combining knowledge-graph (KG)-aware retrieval with schema-restricted large language model (LLM) decoding.
• Contrastive learning over multi-view administrative graphs for candidate retrieval and re-ranking (Recall@K).
• On-premises, candidate-restricted LLM decoding for JSON-compliant outputs, ensuring field consistency and KG alignment.

Main Results:

• AddrKG-LLM demonstrated consistent gains in micro-level ([Formula: see text]) and macro-level ([Formula: see text]) accuracy compared to baselines.
• The framework achieved favorable Recall@K, indicating high gold coverage.
• Evaluations used de-identified real-world records, assessing structural consistency and system properties like latency.

Conclusions:

• AddrKG-LLM offers an accurate, controllable, and deployable solution for structuring noisy addresses.
• Coupling KG-aware retrieval with constrained on-prem LLM decoding effectively overcomes existing challenges.
• The method enhances data standardization for large-scale information systems.