📄 Medical Term Extraction in an Arabic Medical Corpus

Summary

The paper tackles medical term extraction for Arabic, a language with almost no health-domain language resources. Doctors in most Arab countries teach, practise and publish in English or French, and there is no Arabic counterpart of UMLS or SNOMED. Within the Spanish-funded MULTIMEDICA project, the authors built an Arabic medical corpus from the Altibbi health portal and the health sections of three newspapers: Asharq Al-Awsat, Youm7 and El Khabar. They then compared two lexicon-driven extraction strategies:

• Full-term list: 3473 English terms taken from SNOMED and UMLS, machine-translated with Google and checked by a native linguist.
• Affix list: Arabic equivalents of medical Latin/Greek prefixes and suffixes. Arabic forms medical terms by combining whole words rather than by derivation, so each affix maps to a full word, for example -itis to the Arabic word for "inflammation".

The small affix list matched or slightly beat the large term list. After tokenization it reached about 39.4% recall at roughly 65% precision. This suggests that cheap, compositional lexicons can bootstrap terminology in under-resourced domains.

Key points

• Corpus size by source: Altibbi has 2,398,876 tokens in 43,278 documents, Aawsat 48,493 tokens in 68 documents, Youm7 18,948 tokens in 83 documents, and ElKhabar 21,032 tokens in 97 documents. The texts were crawled with wget and converted to XML with unified metadata.
• The full-term list was validated against Arabic Wikipedia, the Altibbi dictionary and the WHO-EMRO Unified Medical Dictionary. In multiword terms, mainly the nominal heads were translated.
• The 460 English affixes reduced to 410 Arabic equivalents. Overly generic affixes such as "re-" or "-ous" were dropped.
• The evaluation sub-corpus contains 389 term occurrences. Tokenizing the text before matching improved the affix approach, which then achieved 39.4% recall with precision of about 65%.
• The error analysis found these causes of missed terms:
  • dual forms ("two canals");
  • accusative indefinite endings;
  • dropped final weak letters;
  • spelling errors;
  • missing pharmaceutical and chemical terms;
  • borderline general words such as "surgery".

Interesting ideas

• I1. Rather than listing whole terms, the authors listed meaning-bearing components and flagged any word containing one, trading precision for recall and far lower upkeep.
• I2. Many exact-match failures came from inflection (duals, case endings, dropped letters), not from gaps in the lexicon itself.
• I3. A small validated seed list can be grown by querying dictionaries for every recorded combination that contains a seed term.
• I4. Evaluation was organised as rounds that change one pipeline factor at a time, such as adding tokenization, so the gain of each step is visible.
• I5. The corpus deliberately uses popularized health news, a register between specialist and everyday language, split by region to check for variety effects.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Unknown English words could get a provisional domain or semantic tag from affixes like cardio- or -itis, with the matched affix recorded as the explanation.
• I2 (tokens): Gazetteer and lexicon lookups should run on lemmas, and each miss should be logged with its morphological cause so lookup failures stay explainable.
• I3 (world): Wikidata or WordNet seed concepts could be expanded the same way into multiword entries for the ontology, keeping a provenance link to the seed.
• I4 (learn): Evaluation reports could follow the same one-change-per-round design to attribute gains to specific pipeline stages.
• I5 (learn): Test sets could include such mid-register texts as a separate slice to measure how much domain vocabulary hurts the tagger and parser.