107 lines
3.5 KiB
Python
107 lines
3.5 KiB
Python
from __future__ import annotations
|
|
|
|
from dataclasses import dataclass
|
|
|
|
import numpy as np
|
|
|
|
from generate.proposition import Proposition
|
|
from vocab.manifold import VocabManifold
|
|
|
|
|
|
@dataclass(frozen=True, slots=True)
|
|
class ArticulationPlan:
|
|
subject: str
|
|
predicate: str
|
|
object: str | None
|
|
surface: str
|
|
output_language: str
|
|
frame_id: str
|
|
|
|
|
|
def _candidate_indices(vocab: VocabManifold, output_language: str) -> np.ndarray:
|
|
indices = vocab.indices_for_language(output_language)
|
|
if len(indices) == 0:
|
|
raise ValueError(f"No articulation candidates for output language {output_language!r}.")
|
|
return indices
|
|
|
|
|
|
def _surface_for_word(vocab: VocabManifold, word: str) -> str:
|
|
morphology = vocab.morphology_for_word(word)
|
|
if morphology is None:
|
|
return word
|
|
return morphology.lemma or morphology.surface
|
|
|
|
|
|
def _resolve_slot(
|
|
versor: np.ndarray | None,
|
|
vocab: VocabManifold,
|
|
output_language: str,
|
|
) -> str | None:
|
|
if versor is None:
|
|
return None
|
|
word, _ = vocab.nearest(
|
|
versor,
|
|
candidate_indices=_candidate_indices(vocab, output_language),
|
|
)
|
|
return _surface_for_word(vocab, word)
|
|
|
|
|
|
def _assemble(subject: str, predicate: str, object_: str | None, output_language: str) -> str:
|
|
if output_language == "he":
|
|
parts = [predicate, subject]
|
|
if object_ is not None:
|
|
parts.append(object_)
|
|
return " ".join(part for part in parts if part)
|
|
if output_language == "grc":
|
|
parts = [subject]
|
|
if object_ is not None:
|
|
parts.append(object_)
|
|
parts.append(predicate)
|
|
return " ".join(part for part in parts if part)
|
|
parts = [subject, predicate]
|
|
if object_ is not None:
|
|
parts.append(object_)
|
|
return " ".join(part for part in parts if part)
|
|
|
|
|
|
def realize(
|
|
proposition: Proposition,
|
|
vocab: VocabManifold,
|
|
output_language: str = "en",
|
|
) -> ArticulationPlan:
|
|
"""
|
|
Map proposition frame slots to morphology-backed target-language surface forms.
|
|
|
|
Slot resolution is purely geometric:
|
|
1. take the slot versor from Proposition
|
|
2. restrict candidates to the configured output language
|
|
3. choose nearest manifold point by CGA inner product through VocabManifold
|
|
4. return MorphologyEntry.lemma when available, else surface
|
|
|
|
Falls back to proposition.surface when subject or predicate slot versors are
|
|
absent (e.g. OOV-heavy input where grounding partially failed), rather than
|
|
raising and crashing ChatRuntime.chat().
|
|
"""
|
|
subject = _resolve_slot(proposition.subject_versor, vocab, output_language)
|
|
predicate = _resolve_slot(proposition.predicate_versor, vocab, output_language)
|
|
object_ = _resolve_slot(proposition.object_versor, vocab, output_language)
|
|
|
|
if subject is None or predicate is None:
|
|
# Graceful fallback: slot versors unavailable, use proposition surface directly.
|
|
return ArticulationPlan(
|
|
subject=proposition.subject or "",
|
|
predicate=proposition.predicate or "",
|
|
object=proposition.object_ if hasattr(proposition, "object_") else None,
|
|
surface=proposition.surface,
|
|
output_language=output_language,
|
|
frame_id=proposition.frame_id,
|
|
)
|
|
|
|
return ArticulationPlan(
|
|
subject=subject,
|
|
predicate=predicate,
|
|
object=object_,
|
|
surface=_assemble(subject, predicate, object_, output_language),
|
|
output_language=output_language,
|
|
frame_id=proposition.frame_id,
|
|
)
|