meldynamics/lt-stressed-corpus

LT Stressed Corpus

A sentence-level Lithuanian corpus with word stress marks added to the text.

This release combines MATAS v1.0 and ALKSNIS v3.0 and keeps only fully stressed sentences. Sentences containing numeric expressions are excluded because Lithuanian number normalization is handled separately.

Each row keeps the original sentence, the stressed sentence, token-level analysis, and stress-source provenance.

Two configs are published from the same sentence set:

• default: Keeps the sentence text, token records, and stress/provenance counts.
• full: Adds serialized CoNLL-U and full sentence metadata on top of the slim schema.

The main fields are listed below. For a full field-by-field schema reference, see COLUMNS.md.

Methodology

The corpus is built from the stressed CoNLL-U releases of MATAS v1.0 and ALKSNIS v3.0. Each CoNLL-U sentence becomes one dataset row. The original text is reconstructed from the source tokens, while the stressed text is reconstructed from the stressed token forms. Source file names, sentence identifiers, document and paragraph identifiers, CoNLL-U metadata, token analyses, and per-word stress provenance are retained so that every exported sentence can be traced back to its source.

Stress resolution follows this order: existing stressed output or curated corrections; the reusable stress cache; morphology-aware vdu-nlp-services results; lki.words; and finally the sentence-level fallback recorded as phonology_engine. The fallback is described in terms of LIEPA synthesizer functionality: it receives the ordered sequence of words forming a sentence, assigns stress in context, and the resulting forms are mapped back to the corresponding token positions in that same sequence. It is not a general-purpose standalone phonology claim.

Rows are selected only after all sentence fields and stress metadata have been assembled. The published release keeps rows with no unresolved stressable words (fully_stressed_only) and removes rows containing Arabic digits or Roman numerals (exclude_number_sentences). Number normalization is not performed in this project, so excluded numeric sentences are not silently converted or guessed. Ambiguity counts and stress-source counts remain available in the row metadata and token records.

The default and full configurations contain the same selected sentence set. They differ only in schema: default is lighter, while full additionally retains serialized CoNLL-U and complete sentence metadata.

What Is Included

• 111,358 sentences
• 1,456,222 tokens
• 1,147,025 stressable words
• 0 unresolved stressable words in this export

Sources

• Lithuanian Treebank - ALKSNIS v3.0: Lithuanian dependency treebank distributed here in CoNLL-U. Used here for 2,603 exported sentences.
• MATAS v1.0: Manually checked, morphologically annotated Lithuanian corpus distributed here in CoNLL-U. Used here for 108,755 exported sentences.

How Stress Was Added

Stress was added with a conservative pipeline that prefers reusable and auditable sources over guessing.

1. Existing stressed MATAS and ALKSNIS releases were reused first.
2. Cached and morphology-aware sources are part of the pipeline when available.
3. lki.words was used as a dictionary fallback.
4. A final sentence-level fallback was used only for still-unresolved words in number-free sentences.

In the provenance columns this final fallback is named phonology_engine, but in practical terms it means Lithuanian speech-synthesis preprocessing of the kind used by the LIEPA synthesizer: sentence normalization followed by accent assignment in context.

Most words in this release come from reused stressed archives. Smaller portions were filled in from lki.words and the final sentence-level fallback.

• Reused existing stressed archives: 1,138,365 words
• lki.words fallback: 1,681 words
• Final sentence-level fallback: 6,979 words

Main Fields

• original_text: original sentence text
• stressed_text_unicode: stressed sentence with Unicode combining stress marks
• tokens: token-level records with tags, features, dependency fields, stressed form, and stress-source provenance

The default config omits conllu_sentence and sentence_metadata to keep the download lighter. The full config adds those two provenance-heavy fields back.

Both configs keep the full tokens records and omit duplicate flat token_* arrays and other convenience columns that can be reconstructed from the remaining data.

License

• Both source corpora are marked on their CLARIN-LT item pages as publicly available under PUB_CLARIN-LT_End-User-Licence-Agreement_EN-LT.
• MATAS source page: MATAS v1.0
• ALKSNIS source page: Lithuanian Treebank - ALKSNIS v3.0
• Licence text: PUB_CLARIN-LT_End-User-Licence-Agreement_EN-LT
• This stressed derivative preserves the original sentence content and adds stress annotation. Use it in compliance with the upstream CLARIN-LT licence terms; no broader rights are asserted here.

Citation

@misc{ltstressedcorpus,
  title        = {LT Stressed Corpus},
  author       = {Aleksas Pielikis},
  year         = {2026},
  note         = {Sentence-level Lithuanian corpus with Lithuanian word stress annotations},
  howpublished = {Hugging Face dataset, MEL DYNAMICS}
}

Also cite both original corpora below.

• MATAS: Rimkutė E., Daudaravičius V., Utka A. 2007: Morphological Annotation of the Lithuanian Corpus. Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics; Workshop Balto-Slavonic Natural Language Processing 2007, Prague, 94–99.
• ALKSNIS: Bielinskienė A., Boizou L., Kovalevskaitė J., Rimkutė E. 2016: Lithuanian Dependency Treebank ALKSNIS. Proceedings of the Seventh International Conference Baltic HLT 2016. Amsterdam: IOS Press, 107–114.