First Steps towards Building a Medical Lexicon for Spanish with Linguistic and Semantic Information

Leonardo Campillos-LlanosACL 2019

TL;DR

This paper introduces MedLexSp, a comprehensive medical lexicon for Spanish that bridges the gap between medical terminology and computational linguistics. By integrating linguistic information (lemmas, inflected forms, part-of-speech) with semantic data (UMLS CUIs, types, and groups), MedLexSp aims to improve natural language processing tasks in the medical domain, particularly for Spanish.

What Problem Does This Paper Address?

In the medical field, especially for languages other than English, there's a significant challenge in developing robust computational tools for processing medical text. This is largely due to a lack of comprehensive, well-structured medical lexicons that go beyond simple term lists. Existing resources often lack detailed linguistic information (like inflected word forms or verb conjugations) and semantic annotations, which are crucial for tasks like named entity recognition (NER), concept normalization, and relation extraction. This "data bottleneck" hinders the development of advanced NLP applications for Spanish medical data, which is often sensitive and less readily available than in English.

Methods

The creation of MedLexSp involved a multi-pronged approach to gather and enrich medical terms:

  • Leveraging Existing Resources: The project started with a pre-existing list of Spanish medical terms and their inflected forms. These were then mapped to Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs) using exact string matching.
  • Enrichment through Thesauri and Corpora: To expand coverage, the researchers utilized various medical thesauri and specialized corpora. This included:
  • Thesauri: Spanish versions of MeSH, SNOMED CT, WHO ART, National Cancer Institute (NCI) Dictionary of Cancer Terms, OrphaData, and the Nomenclator de prescripción (for drug names).
  • Corpora: Summaries of Product Characteristics (SPCs) and MedlinePlus Spanish were used to extract terms and variants, particularly those used in patient language.
  • NLP Techniques for Variant Generation: Several methods were employed to generate linguistic variants of terms:
  • String Distance Metrics: To match terms with minor variations (e.g., misspellings, hyphenation).
  • Derivational Variants: Identifying and mapping related terms (e.g., "aneurysm" to "aneurysmatic").
  • Conjugated Verbs: Mapping verb lemmas to their conjugated forms and linking them to CUIs.
  • Syntactic Variants: Generating variations by reordering words in multi-word terms.
  • Affixes and Roots: Incorporating common medical prefixes and roots.
  • Abbreviations and Acronyms: Mapping full terms to their common abbreviations.
  • Semantic Annotation: Each lexical entry was enriched with UMLS semantic types and groups. The authors specifically filtered out overly general or ambiguous semantic types to improve precision for medical NLP tasks.
  • Part-of-Speech (POS) Tagging: The inclusion of POS information is ongoing, with initial testing of automated methods.

Datasets

  • Base List: A pre-existing Spanish medical lexicon, derived from a corpus of 4.2 million tokens.
  • UMLS Metathesaurus (2018AB): Used as the primary source for CUIs, semantic types, and groups.
  • Domain-Specific Thesauri and Databases:
    • Spanish Medical Subject Headings (MeSH)
    • SNOMED CT (Spanish version)
    • World Health Organization Adverse Drug Reactions Terminology (WHO ART)
    • National Cancer Institute (NCI) Dictionary of Cancer Terms
    • OrphaData (INSERM)
    • Nomenclator de prescripción (AEMPS)
    • Spanish Drug Effect database (SDEdb)
  • Corpora:
    • Summaries of Product Characteristics (SPCs) (via Easy Drug Package Leaflets - EasyDPL)
    • MedlinePlus Spanish
  • Evaluation Dataset: Spanish texts from the MANTRA corpus, annotated with UMLS CUIs, semantic types, and groups.

Key Insights and Ideas

  • The Importance of Linguistic Richness: Simply mapping terms to CUIs isn't enough. Including inflected forms, verb conjugations, and syntactic variants is crucial for capturing the full spectrum of medical language used in free text.
  • Bridging the Gap Between Thesauri and Real-World Usage: Many terms found in comprehensive thesauri (like MeSH) appear with low frequency in actual medical texts. This highlights the need for lexicons that are grounded in corpus data and reflect actual usage patterns.
  • Strategic Semantic Filtering: Discarding overly general or ambiguous UMLS semantic types (like "Concepts and Ideas") can significantly improve the performance of medical NLP tasks by reducing noise.
  • The Challenge of Spanish Medical NLP: The paper underscores the relative scarcity of high-quality, readily available medical NLP resources for Spanish compared to English, emphasizing the value of projects like MedLexSp.

Key Results

  • The current version of MedLexSp contains 76,454 lemmas and 203,043 inflected forms, mapped to 30,647 UMLS CUIs.
  • The lexicon was evaluated on the MANTRA corpus, achieving an overall F-measure of 0.83 (exact match) and 0.87 (approximate match) for specific UMLS semantic groups (DISO, CHEM, PROC, ANAT).
  • The evaluation revealed that unmatched entities were often due to misspellings, discontinuous entities, or incorrect scope annotation.

Limitations and Future Work

  • Limited Evaluation Scope: The current evaluation was performed on a small set of texts, limiting generalizability. The authors aim to evaluate MedLexSp on a larger, annotated Spanish medical corpus.
  • Task-Dependence: The lexicon is not entirely task-independent; terms might need filtering based on specific semantic types for different NLP tasks (e.g., removing "Occupation" terms for general medical NER).
  • Contextual Variation: The lexicon doesn't fully capture complex contextual variations, such as coordinated terms or adjective modifiers in multi-word expressions (e.g., "breast cancer" vs. "right breast cancer").
  • Spanish Dialectal Variation: The lexicon primarily reflects Peninsular Spanish, with some terms from other varieties included. Further work is needed to improve coverage of dialectal differences.
  • Future Directions: The authors are interested in exploring embedding-based methods for term expansion and evaluating the lexicon with a broader set of domain texts. They also plan to improve the POS tagging automation.

Summary updated

BibTeX
@inproceedings{campillosllanos2019first,
  author = {Leonardo Campillos-Llanos},
  title = {{First Steps towards Building a Medical Lexicon for Spanish with Linguistic and Semantic Information}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure