gector-bert-de / language-rules.json
Imperatore's picture
German 1.0.0: run 2026-11-08-0749 (gold 168/84/10/0, Tatoeba 0.53%)
d37ed2d verified
Raw History Blame Contribute Delete
1.16 kB
{
"version": 1,
"language": "de",
"_comment": "Per-bundle language facts (see C# LanguageRules / ADR-004). stripTrailingTerminator true: the DE bundle follows the same BERT/WordPiece recipe as Italian — validate at eval. clauseConjunctions: connectors after a comma = independent-clause boundary for GectorAnalyzer.Correct's clause split (see the EN/IT files). German caveat: weil/wenn/dass/obwohl/während introduce verb-FINAL subordinate clauses, so the split hands the model a fragment with non-main-clause word order; the Italian precedent (perché/mentre, HARMFUL 0) suggests this is safe because GECToR only makes word-level agreement/form edits, but PRUNE these five at the DE eval if comma_context shows harm. The coordinators (aber/und/oder/denn/sondern/doch) and V2 adverbial connectors (dann/also/deshalb/...) are unproblematic. Validate against lang/de/gold comma_context at eval.",
"stripTrailingTerminator": true,
"clauseConjunctions": ["aber", "und", "oder", "denn", "sondern", "doch", "sonst", "dann", "also", "deshalb", "deswegen", "trotzdem", "dennoch", "jedoch", "außerdem", "weil", "wenn", "dass", "obwohl", "während"]
}