Skip to content

Language Planning

LanguagePlanner converts a language-neutral CommunicationPlan into a target-language LanguagePlan.

No English pivot

SA does not generate English and then translate it.

wrong: semantic meaning → English sentence → target language
right: semantic meaning → target-language planning → GF target language

LanguagePlan

A LanguagePlan contains:

  • target language and locale;
  • output block topology;
  • RealizationUnits;
  • SA↔GF operation IDs;
  • semantic role bindings;
  • lexical slots;
  • grammatical/communicative features;
  • source obligation IDs.

RealizationUnit

A RealizationUnit is the smallest planned grammatical unit passed to the realizer boundary.

Conceptually:

RealizationUnit
  unit_id
  operation_id
  role_bindings
  lexical_slots
  feature_bindings
  obligation_ids
  semantic_refs
  output_slot

There is no extra general-purpose construction IR between LanguagePlan and the realizer.

Language-specific variation

Different languages may legitimately:

  • split the same obligations into different sentence counts;
  • realize possession differently;
  • use different argument packaging;
  • express politeness differently;
  • choose different information structures.

The obligation set and semantic roles must remain faithful.

No language branches in shared core

Shared planning logic must not become:

if language == "sq": ...
elif language == "fr": ...

Language-specific behavior belongs in validated grammar/runtime artifacts or explicit language-planning strategies behind stable interfaces.

Extensible predicate planning hints

For domain predicates that cannot be interpreted from generic roles alone, ACLs may provide supporting hints:

  • sa:operation
  • sa:role-map
  • sa:slot-map
  • sa:feature-map

These hints select a canonical SA operation without adding domain branches to the core. They are realization metadata, not new facts.