| from __future__ import annotations |
|
|
| from .benchmark import ( |
| JAPANESE_RE, |
| BenchmarkExample, |
| CorpusToken, |
| PreparedSplit, |
| SplitCoverage, |
| ) |
| from .domain import RerankRequest |
| from .mozc import MozcDictionaryIndex |
|
|
|
|
| def prepare_mozc_examples( |
| sentences: tuple[tuple[CorpusToken, ...], ...], |
| index: MozcDictionaryIndex, |
| *, |
| pool_size: int, |
| context_mode: str, |
| ) -> PreparedSplit: |
| if pool_size < 2: |
| raise ValueError("pool_size must be at least 2") |
| if context_mode not in {"left_only", "bidirectional"}: |
| raise ValueError("context_mode must be left_only or bidirectional") |
| eligible_tokens = 0 |
| ambiguous = 0 |
| oracle_in_pool = 0 |
| oracle_miss = 0 |
| examples: list[BenchmarkExample] = [] |
| for sentence in sentences: |
| surfaces = tuple(token.surface for token in sentence) |
| for target_index, token in enumerate(sentence): |
| if not token.reading or not JAPANESE_RE.search(token.surface): |
| continue |
| eligible_tokens += 1 |
| candidates = index.lookup(token.reading, limit=pool_size) |
| if len(candidates) < 2: |
| continue |
| ambiguous += 1 |
| if token.surface not in {candidate.surface for candidate in candidates}: |
| oracle_miss += 1 |
| continue |
| oracle_in_pool += 1 |
| examples.append( |
| BenchmarkExample( |
| request=RerankRequest( |
| reading=token.reading, |
| candidates=candidates, |
| left_context=surfaces[:target_index], |
| right_context=( |
| surfaces[target_index + 1 :] |
| if context_mode == "bidirectional" |
| else () |
| ), |
| ), |
| expected=token.surface, |
| ) |
| ) |
| return PreparedSplit( |
| examples=tuple(examples), |
| coverage=SplitCoverage( |
| eligible_tokens=eligible_tokens, |
| ambiguous_known_reading=ambiguous, |
| oracle_in_pool=oracle_in_pool, |
| oracle_miss=oracle_miss, |
| ), |
| ) |
|
|
