meldynamics/lt-stressed-corpus

LT Stressed Corpus

Lithuanian sentences with word stress marks.

The dataset combines MATAS v1.0 and ALKSNIS v3.0. It is useful for speech technology, pronunciation work, language learning, and research on Lithuanian stress.

This release contains sentences where every Lithuanian word that needs stress has a selected stressed form. Sentences with Arabic or Roman numerals are left out because reading a number correctly depends on context and grammatical form.

Some Lithuanian words allow more than one stress pattern. The dataset keeps the selected form and also records ambiguity counts. Here, fully stressed means that every applicable word has a selected form; it does not mean that every choice was unambiguous.

Quick Start

from datasets import load_dataset

dataset = load_dataset("meldynamics/lt-stressed-corpus", "default")

Choose a Config

Both configs contain the same sentences:

• default: recommended for most users; includes sentence text, tokens, and stress information.
• full: also includes the original serialized CoNLL-U sentence and all sentence metadata.

What Is Included

• 111,358 sentences
• 1,456,222 tokens
• 1,147,025 stressable words
• 0 unresolved stressable words in this export
• 147,670 source sentences considered before filtering
• 36,312 sentences omitted because they contain unresolved words, numbers, or both

The published data is a clean subset of the source corpora, not a complete export of every source sentence.

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.

Data Origin and Stress Annotation

The sentence text and linguistic annotation come from the MATAS v1.0 and ALKSNIS v3.0 CoNLL-U corpora. Their token forms, lemmas, and Jablonskis/XPOS morphology were retained; ALKSNIS dependency annotation was also preserved. The source corpora did not contain the stress marks added by this project.

Most stress forms were added in the 2020 build by querying the former VDU Kirtis SOAP service for one word at a time (through the vdu-nlp-services Python client). Kirtis returned stressed alternatives with morphological labels. The builder converted those labels to the corpus tag vocabulary and selected the alternative with the same part of speech, the greatest tag overlap, and then the fewest extra tags. It used the existing MATAS/ALKSNIS morphology; it did not call a separate VDU morphology service or use sentence dependencies for disambiguation.

VDU describes Kirtis as a morphology- and rule-based accentuator using Lemuoklis, syllabification, and lexical resources derived from Lithuanian dictionaries and accentuation references (2005 report; 2010 description). If Kirtis did not supply a stressed form, the build tried an exact lookup in stressed entries collected from LKIIS dictionaries.

The current release preserved 1,138,365 forms from those 2020 stressed archives as legacy, added 1,681 exact LKIIS-derived lookups, and added 6,979 sentence-level fallbacks from phonology_engine, derived from LIEPA-1. Because the old archives did not record a provider for each token, legacy usually means VDU-selected but can also include an older cache or dictionary fallback. The stress layer is automatic rather than exhaustively human-verified.

The public export keeps only fully resolved sentences without Arabic or Roman numerals.

Main Fields

• original_text: original sentence text
• stressed_text_unicode: stressed sentence with Unicode combining stress marks
• tokens: words and punctuation with linguistic analysis, stressed forms, and stress sources

See COLUMNS.md for the complete field reference.

License

The dataset files are published under Creative Commons Attribution 4.0. The current CLARIN-LT records list both source corpora under the same licence.

• MATAS source page: MATAS v1.0
• ALKSNIS source page: Lithuanian Treebank - ALKSNIS v3.0

This derivative keeps the original sentence content and adds stress annotation. Please also credit the original corpora. Software used to build the data has separate terms; see the phonology_engine licence and upstream attribution.

Project

Build scripts and source archives are maintained at matas-alksnis-stressed.

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.