offelia39/pms-wiki-mined-corpora

Piedmontese–Italian Parallel Corpora and MT Evaluation

Data for Italian→Piedmontese machine translation, released with:

• Yulia Panchenko. "When metrics miss the language: Machine translation evaluation for Piedmontese". In: Proceedings of the Twelfth Italian Conference on Computational Linguistics. CLiC-it '26. Palermo, Italy: CEUR Workshop Proceedings, 2026. Forthcoming.
• Yulia Panchenko. Low-Resource Machine Translation for Piedmontese: Model Training, Human Evaluation, and Metric Reliability. Master's thesis, University of Turin, 2026.

Code (corpus building, model training, evaluation): https://github.com/panchenko-ling/piedmontese_wiki_ml

Contents

File	Description	Size
pmswiki_noleak.jsonl	Cleaned Piedmontese Wikipedia dump (no overlap with the mined parallel corpus)	~59k documents, ~4M tokens
wiki_mined_clean.csv	Italian–Piedmontese sentence pairs mined from Wikipedia via LaBSE alignment, cleaned	4,714 pairs
it-pms_test_translations.csv	FLORES+ Italian sentences with original and standardised Piedmontese references and NLLB outputs	140 sentences

Dataset Details

Piedmontese Wikipedia Corpus (pmswiki_noleak.jsonl)

• Source: Piedmontese Wikipedia dump (01/02/2026)
• Cleaning: Articles cleaned with mwparserfromhell: removed categories, file/image links, wiki markup, URLs, section headers, image caption directives, pixel dimensions, ISO language code references etc.
• Format: JSONL, one article per line with fields id, title, text
• Note: Deduplicated against the mined parallel corpus

Mined Parallel Corpus (wiki_mined_clean.csv)

• Source: Piedmontese–Italian Wikipedia article pairs aligned via Wikidata interlanguage links
• Alignment: LaBSE cosine similarity ≥ 0.85; sentences between 5–100 words
• Columns: source (Italian), target (Piedmontese), similarity (LaBSE cosine similarity), article_title (title of the Wikipedia article the pair was mined from)
• Note: Cleaned of identical pairs and near-copies, non-Italian sources and non-Piedmontese targets (LLM-assisted)

Test Translations (it-pms_test_translations.csv)

• Source sentences: Italian, from the FLORES+ dev set
• Reference: Human translations from Vico & Libovický (2026), in the speakers' original (non-standard) orthography
• Standardised reference: The same references respelled in standard Piedmontese orthography following Tosco's grammar (LLM-assisted conversion, uncertain cases reviewed by a native speaker). Only spelling was changed; lexicon, morphology and word order are the speakers' own.
• MT outputs: Four NLLB-200 conditions:
  • NLLB-avg: zero-shot; Piedmontese token initialised from the average of Italian, Occitan, Ligurian and Lombard embeddings
  • NLLB-avg+MT: fine-tuned on the mined parallel corpus
  • NLLB-avg+DAE-50K+MT: denoising pretraining on 50K Piedmontese Wikipedia sentences, then MT fine-tuning (NLLB-avg+DAE+MT in the CLiC-it paper)
  • NLLB-avg+DAE-full+MT: same, with the full corpus (191,884 sentences)
• Columns: flores_id, source, reference, reference_std_revised, NLLB-avg, NLLB-avg+MT, NLLB-avg+DAE-50K+MT, NLLB-avg+DAE-full+MT

Language

• Piedmontese (ISO 639-3: pms): endangered Romance language of the Piedmont region in Italy
• Italian (ISO 639-3: ita)

Citation

@inproceedings{panchenko2026metrics,
  author    = {Panchenko, Yulia},
  title     = {When metrics miss the language: {M}achine translation evaluation for {P}iedmontese},
  booktitle = {Proceedings of the Twelfth Italian Conference on Computational Linguistics},
  series    = {CLiC-it '26},
  address   = {Palermo, Italy},
  publisher = {CEUR Workshop Proceedings},
  year      = {2026},
  pubstate  = {forthcoming},
}

@mastersthesis{panchenko2026thesis,
  author = {Panchenko, Yulia},
  title  = {Low-Resource Machine Translation for {P}iedmontese: Model Training, Human Evaluation, and Metric Reliability},
  school = {University of Turin},
  year   = {2026}
}

License

CC BY-SA 4.0

Contact

Yulia Panchenko, yulia.panchenko@edu.unito.it