📄 Adapting a Medical speech to speech translation system (MedSLT) to Arabic

Summary

The paper extends MedSLT, a grammar-based medical speech translation system for doctor–patient diagnosis dialogues, to Arabic. Commercial MT tools often fail on this kind of input. They lose the question form, copy English structure into ungrammatical Arabic, or mishandle ellipsis. The authors therefore reuse MedSLT's Regulus architecture. A general unification grammar is written once, specialized automatically to a domain using a corpus, and compiled into a Nuance CFG for recognition and into a grammar for semantic head-driven generation. Translation passes through a flat interlingua. The Arabic grammar, modelled on the existing French, English, Spanish and Catalan grammars, covers headache questions with 322 word forms.

The evaluation tests only the specialized Arabic generation grammar. Four translators judged output for 522 English diagnostic questions. Among sentences the English side analyzed correctly, about 90–93% were rated Good or OK. Systran's output on a comparable sample was mostly judged Bad. The main lesson is that a new language can be added cheaply when its grammar is derived from grammars that already exist.

Key points

• The architecture combines Regulus unification grammars, corpus-based grammar specialization, compilation to CFG for the Nuance recognizer, and semantic head-driven generation.
• Translation runs in five steps:
  • source analysis,
  • ellipsis resolution,
  • source-to-interlingua mapping,
  • interlingua-to-target mapping,
  • target generation.
• The representation at every step is a flat list of attribute–value pairs.
• The Arabic grammar has 322 word forms and 38 rules for hal yes-no questions, plus some wh-questions. It adds a "dual" value to the agreement feature. It handles pro-drop and verb–subject inversion through a vbar constituent rather than movement.
• On the 522-question corpus:
  • 114 sentences (21.8%) were not analyzed by the English grammar.
  • 21 (4%) were not translated, mostly pain terms still awaiting validation by Arabic medical specialists.
  • Only 3 (0.6%) failed at generation.
  • Good+OK on the 408 analyzed sentences ranged from 89.9% to 93% across the four judges.
• On 124 well-analyzed sentences, Systran produced 98 Bad, 20 OK and 6 Good translations.

Interesting ideas

• I1. A broad general grammar is pruned automatically into a domain-specific one using a corpus of typical sentences, which yields flatter trees with fewer empty optional-modifier nodes.
• I2. Nouns carry semantic types, and verbs state which types their subject and object require. Dropping these restrictions noticeably hurt recognition, so they served as filters rather than decoration.
• I3. Meaning is a flat bag of typed attribute–value pairs, and the interlingua is simply the most explicit, standardized English version of it. Translation then reduces to mapping between simple structures, which makes language divergences easier to handle.
• I4. Attached particles such as the article and prepositional clitics are split off as separate tokens inside the grammar and glued back by orthographic rules only after generation.
• I5. The evaluation reports failures separately for each pipeline stage (not analyzed, not translated, not generated). It measures quality only on inputs the earlier stages handled, and it adds a commercial-system baseline.

For Mova

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

• I1 (tree): Domain-specialized parser profiles, for example for school tests or fables, could be derived from the general parser using in-domain silver trees, while the general model stays as a fallback.
• I2 (frames): VerbNet selectional restrictions could be used actively to rank competing parses and role assignments, rejecting frames whose arguments violate the verb's type demands.
• I3 (back): A normalized, maximally explicit English frame could serve as the meta-world pivot for the text → graph → text compressor, with per-language mapping rules on each side.
• I4 (back): Generation could work on uncontracted tokens and apply a final surface pass for contractions, a/an selection and possessive clitics, keeping grammar rules smaller.
• I5 (learn): Answer-level evaluation could attribute each error to a stage (tokens, tree, frames, world) and report conditional accuracy, so regressions are located rather than averaged away.