📄 Automated Processing of Medical English

Summary

The authors, working at the Division of Computer Research and Technology of the National Institutes of Health, describe APME (Automated Processing of Medical English). It is a system under development that aims to turn free-text pathology diagnoses into a structured semantic representation suitable for storage and retrieval. The authors argue that general machine translation and language understanding remain out of reach. A restricted domain lets limited goals be built and tested. Pathology diagnoses are such a domain, and they touch every other medical specialty. This sublanguage has several hard features: it largely lacks verbs, commas stand in for "to be," modifiers can be separated from their nouns, and Latin, Greek and English forms are mixed. The lexicon base is SNOP (Systematized Nomenclature of Pathology), a hierarchical code system with four semantic axes. The grammar layer combines three parts: morphological normalization, morphosemantic splitting of compound words, and paraphrase transformations of noun phrases into equivalent variants. Relations between SNOP-coded operands are expressed by relational predicates in a formal intermediate language. The paper reports no quantitative evaluation. Its value lies in an early, explicit design for domain semantics that is driven by a lexicon and mapped onto an interlingua.

Key points

• SNOP has four independent semantic fields: Topography (T, body site), Morphology (M, structural change), Etiology (E, causative agent) and Function (F, physiological manifestations and some infectious diseases). Each term belongs to exactly one field.
• Each SNOP term gets a four-digit code. The first digit gives the section and the remaining digits refine it step by step, so moving along the code list yields broader or narrower terms. Semantic linkage pointers cross-reference terms within or across fields.
• Morphological analysis handles the pathologist's unedited text with three kinds of rewriting rules:
  • adjective suffixes become noun suffixes;
  • plurals become singulars;
  • words are mapped to synonyms or near-synonyms, e.g. DYSENTERY → DISENTERIA.
• Compound words are split into semantically labelled parts with a fixed order. For example, BRONCHITIS is written as W(T, M) → (C1-T BRONCH(US)) + (C2-M -ITIS), where -ITIS marks inflammation.
• Relations between concepts are carried by relational predicates such as "in," "due to," or a bare comma. Semantic rules act as axioms, and the theorems derived from them form the grammar of the intermediate language.

Interesting ideas

• I1. Lemmatize toward the lexicon's canonical form instead of a generic base, for example by mapping "muscular" to "muscle" when the dictionary lists only the noun, so that lookups hit more often.
• I2. Treat a compound word as a small semantic structure: each morpheme carries a category, and the parts have a fixed order (site + process), so one word can supply two concept slots.
• I3. Group surface variants of a noun phrase into one paraphrase class: "muscle atrophy," "atrophy of muscle," "atrophy, muscle," "muscular atrophy" and "atrophic muscle" all reduce to the same head-modifier relation.
• I4. Paraphrase rules should be licensed by semantic categories, not only by syntax, because purely formal transformations can produce phrases that do not fit the meaning.
• I5. Function words and even punctuation (a comma, "due to") can act as relational predicates between typed operands in a verbless sublanguage.

For Mova

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

• I1 (tokens): A lemma normalizer could rewrite derived adjectives and plurals into the head entry used by our ontology, which would improve linking of domain terms.
• I2 (world): When building entities, "bronchitis" or "pneumonia" could be expanded into a condition plus an affected body part instead of one opaque node.
• I3 (frames): A canonicalization step could map compound, "of", comma and adjectival variants of an NP to one relation, so frames and coreference treat them as the same thing.
• I4 (learn): Self-induced repair or rewrite rules could carry type constraints from WordNet or the ontology, with a negative control checking that a rule does not fire on mismatched categories.
• I5 (tree): For headline-like or list-like text, UD attachments through commas and prepositions could be mapped to explicit relations (location, cause) instead of being left as loose appositions.