Zazaki and English/Structural transfer

From LING073
Jump to: navigation, search

Pre-evaluation

Baseline measurements before implementing any structural transfer rules, using the 7 transfer test sentences.

  • Coverage: 100% (0/26 unknown words)
  • WER: 100.00%
  • PER: 100.00%

The high error rates are expected, the English generator rejects most output because Zazaki-specific tags (<dir>, <obl>, <m>, <f>) are not stripped, postpositions are not reordered, and no articles are inserted. Most content words appear with # markers or are dropped entirely.

Implemented rules

Case and gender stripping on nouns and verbs (Contrastive grammar items 1 and 2)

Zazaki nouns carry case tags (<dir>/<obl>) and gender tags (<m>/<f>) that English nouns don't use. Zazaki past-tense verbs carry person, number, and gender agreement tags that English past-tense verbs don't need. We wrote two transfer rules in the .t1x file that strip these extra tags, outputting only what the English generator expects: lemma<n><number> for common nouns, lemma<np><ant><m><number> for proper nouns, and lemma<vblex><tense> for verbs.

We also fixed the bilingual dictionary entries for proper nouns (Îsa, Şîmûn) to include <ant> and <m> on the English side, and removed the -n flag from cg-proc in modes.xml to fix a pipeline formatting issue.

Example: Îsayî va. → Jesus said.

Tagger:

^Îsa<n><m><sg><obl>$ ^vat<v><past><p3><m><sg>$^.<sent>$

Lex:

^Îsa<n><m><sg><obl>/Jesus<np><ant><m><sg><obl>$ ^vat<v><past><p3><m><sg>/say<vblex><past><p3><m><sg>$^.<sent>/.<sent>$

Transfer:

^n<SN>{^Jesus<np><ant><m><sg>$}$ ^v<SV>{^say<vblex><past>$}$ ^sent<SENT>{^.<sent>$}$

Full pipeline (zza-eng):

Jesus said.

Post-evaluation

After implementing structural transfer rules for noun case/gender stripping and verb tag cleanup:

  • Coverage: 100% (0/26 unknown words)
  • WER: 83.33% (down from 100.00%)
  • PER: 76.67% (down from 100.00%)
  • Test words produced: 26 (up from 7)

The transfer rules eliminate # markers on all nouns and verbs. Remaining errors are due to word order (postpositions not yet reordered), missing definite articles (not yet inserted), and pronoun/determiner tag mismatches.