A Multi-Pass Sieve for Name Normalization

Jennifer D'SouzaAAAI 2015 · DOI

Abstract

Medical name normalization, the task of mapping entity mentions to unique ontology identifiers, is complicated by linguistic variations and domain-specific terminology. This paper introduces a multi-pass sieve framework consisting of eight deterministic, rule-based modules ordered from highest to lowest precision. By processing mentions through this "waterfall" architecture, the system achieves state-of-the-art performance on clinical and biomedical datasets. The approach offers a modular and transparent alternative to traditional vector-space models without sacrificing accuracy.

Introduction

Name normalization, or entity linking, is a critical step in clinical natural language processing. It involves mapping surface-level mentions (e.g., "swollen abdomen") to a canonical concept ID in a structured ontology (e.g., UMLS). In the medical domain, this task is particularly difficult due to the prevalence of abbreviations, complex syntactic shifts, and synonymous disorder terms.

While many existing systems rely on Term Frequency-Inverse Document Frequency (tf-idf) Vector Space Models (VSMs), these can be computationally intensive and lack the interpretability of rule-based systems. This work proposes a simpler, modular alternative based on the multi-pass sieve framework, originally successful in coreference resolution.

Methodology: The Multi-Pass Sieve Approach

The system utilizes a sequence of eight deterministic sieves. A mention is passed through the sieves in order; if a match is found in the target ontology, the process terminates. If no match is found, the mention proceeds to the next, lower-precision sieve.

The eight sieves are:

  1. Exact Match: Direct string matching with the ontology.
  2. Abbreviation Expansion: Resolving shorthand using algorithmic expansion and external resources.
  3. Subject ⇔ Object Conversion: Handling syntactic reordering (e.g., "changes on ekg" to "ekg changes").
  4. Numbers Replacement: Normalizing numerals and words.
  5. Disorder Synonyms Replacement: Swapping terms with known synonyms.
  6. Affixation: Applying morphological patterns learned from training data.
  7. Stemming: Using the Porter stemmer for base-form matching.
  8. Partial Match: A final "catch-all" using token overlap or tail substring matching.

Evaluation

The system was evaluated on two distinct corpora to test its robustness across different medical text styles:

  • ShARe/CLEF eHealth Challenge: 298 clinical notes mapped to the UMLS Metathesaurus.
  • NCBI Disease Corpus: 793 biomedical abstracts mapped to the MEDIC lexicon.

Results

The multi-pass sieve architecture demonstrated consistent, incremental improvements as each layer was added. The final system achieved an accuracy of 88.3% on the ShARe corpus and 83.3% on the NCBI Disease Corpus, outperforming previous state-of-the-art models like DNorm.

Incremental accuracy improvements across the eight sieves for both the ShARe and NCBI corpora.

As shown in the results, the initial "Exact Match" sieve captures the majority of mentions (69-83%). However, the subsequent sieves are vital for addressing the "long tail" of linguistic variation. Notably, the "Partial Match" sieve provides a significant final boost, demonstrating the efficacy of the waterfall approach in filtering out high-confidence matches before applying noisier rules.

Discussion and Conclusion

The success of the multi-pass sieve framework stems from its precision-first ordering. By prioritizing high-confidence matches, the system minimizes false positives that might otherwise be triggered by the more flexible, lower-precision sieves.

Furthermore, the modularity of the system allows for easy debugging and domain adaptation. Unlike "black box" machine learning models, individual sieves can be added or refined based on specific linguistic requirements without retraining the entire pipeline. While the system remains dependent on the quality of the underlying ontologies and some hand-crafted rules, its performance on both clinical notes and formal abstracts suggests a high degree of cross-domain robustness.

Summary updated

BibTeX
@inproceedings{dsouza2015multi,
  author = {Jennifer D'Souza},
  title = {{A Multi-Pass Sieve for Name Normalization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9729},
}
Figure