TL;DR
SurfCon is a novel framework designed to discover medical synonyms (like "vitamin C" and "ascorbic acid") using only "privacy-aware" clinical data—term lists and co-occurrence graphs—rather than raw patient text. By combining a bi-level surface form encoder with a dynamic context-matching mechanism, it significantly outperforms existing methods and can even find synonyms for terms it has never seen before (Out-of-Vocabulary).
What Problem Does This Paper Address?
Unstructured clinical notes are a goldmine for medical knowledge, but they are also a privacy minefield. Due to HIPAA and other regulations, researchers rarely get access to raw patient records. Instead, institutions often release "privacy-aware" datasets: lists of extracted medical terms and their aggregated co-occurrence counts (how often two terms appear near each other).
The challenge is that traditional synonym discovery relies heavily on "local context"—the specific sentences where a word appears. Without raw text, we lose that context. Furthermore, medical synonyms are tricky:
- Surface Similarity: "Vit C" and "Vitamin C" look alike but are spelled differently.
- Semantic Similarity: "Vitamin C" and "Ascorbic acid" look nothing alike but mean the same thing.
- The OOV Problem: Users often search for terms (Out-of-Vocabulary) that aren't even in the provided co-occurrence graph.
Methods
The authors propose SurfCon, which tackles these issues through two main modules:
- Bi-level Surface Form Encoding: This module looks at the "look" of a word. It uses both character-level (to catch misspellings and abbreviations) and word-level embeddings. These are combined into a single vector that represents the term's appearance.
- Context Matching: This is the "secret sauce" for handling semantic similarity and OOV terms.
- Inductive Context Prediction: Instead of relying on the noisy, raw co-occurrence graph, SurfCon uses the surface form vector to predict what a term's global context (its neighbors) should be. This allows the model to "hallucinate" a neighborhood for OOV terms.
- Dynamic Matching Mechanism: When comparing two terms, the model doesn't just average their neighbors. It uses an attention-like mechanism to weight neighbors that are relevant to both terms. For example, if comparing "vitamin C" and "ascorbic acid," it might prioritize neighbors like "iron absorption" or "anemia."
Finally, these two scores (surface and context) are combined and optimized using ListNet, a popular ranking algorithm, to produce a final recommended list of synonyms.
Datasets
The researchers used publicly available medical term co-occurrence data extracted from 20 million clinical notes from Stanford Hospitals and Clinics.
- 1-day and All-day datasets: These represent different "temporal bins" for co-occurrence (terms appearing in the same day vs. the same patient record).
- Ground Truth: Synonyms were labeled using the UMLS (Unified Medical Language System), where different strings mapping to the same Concept Unique Identifier (CUI) are considered synonyms.
- OOV Test Set: A specific set of 2,000 terms not present in the co-occurrence graph was created to test the model's ability to handle unseen queries.
Key Insights and Ideas
- Prediction as Denoising: Raw co-occurrence graphs are often pruned for privacy (e.g., removing edges with <100 counts), which creates "holes" in the data. By training a model to predict context from surface forms, SurfCon effectively "denoises" the graph.
- Surface Form is a Strong Prior: In medicine, many synonyms are just variations of the same string (abbreviations, typos). The bi-level encoder captures this better than standard word-level models.
- Dynamic vs. Static: The authors found that dynamic matching (weighting neighbors based on the specific pair being compared) provided a ~4% performance boost over static averaging, proving that context relevance is pair-dependent.
Key Results
SurfCon consistently outperformed strong baselines (including Word2vec, LINE, and specialized synonym discovery models like DPE) across all settings.
- MAP Scores: On the "In-Vocabulary" task, SurfCon achieved a Mean Average Precision (MAP) of 0.9176, compared to 0.8565 for the best surface-only baseline.
- OOV Performance: Even for terms never seen in the graph, SurfCon maintained high accuracy (MAP of 0.8301), whereas context-only models like Word2vec cannot function at all on OOV terms.
- Qualitative Success: In case studies, the model successfully identified "aphonia" (the inability to produce voice) as a synonym for "unable to vocalize," despite having zero character overlap.
Limitations and Future Work
- Simple Encoders: The current model uses relatively simple look-up tables for embeddings. The authors suggest that integrating more complex encoders like BERT or Transformers could further improve results.
- Beyond Synonymy: SurfCon is currently tuned for synonyms. Future work could extend the framework to discover other relationships, such as hypernyms (is-a) or drug-side effect relationships.
- Domain Specificity: While tested on clinical data, the "privacy-aware" constraint is common in finance and legal fields, suggesting SurfCon could be adapted for those domains as well.
Summary updated