license: cc0-1.0
Synthetic Parallel Corpora
Description
The dataset consists of synthetic parallel corpora for the following language pairs: Lithuanian-English, Lithuanian-French, and Lithuanian-German. The corpora are intended for training and improving neural machine translation systems as well as for other natural language processing tasks. The dataset contains more than 3 million parallel sentence pairs, with over 1 million sentence pairs for each language pair. The data were generated using a context-based template method, which enables the systematic inclusion of named entities (e.g., personal names, locations, organizations, etc.) and ensures their correct usage across various grammatical forms. In addition, the Lithuanian-English corpora include medical terminology-based synthetic data. The dataset also includes the resources used for synthetic data generation: more than 20,000 named entities per language pair (across 11 named entity categories), as well as more than 50 context templates for each category. The corpora feature a controlled structure and linguistic diversity, as the sentences are generated from templates derived from real language usage examples. The data are provided in a parallel format, ensuring direct sentence alignment across languages, and are suitable for immediate use in model training. The dataset is available in TXT and TMX formats, making it compatible with both machine learning environments and translation memory-based software. This resource is particularly useful for tasks involving the translation of named entities and domain-specific terminology, as well as for model evaluation and error analysis.
Publisher
State Digital Solutions Agency
Date issued
2026-04-13
Type
Corpus, text
Size
3328707 sentences
Licence
CC0 1.0
Acknowledgement
NextGenerationEU/Naujos kartos Lietuva Project code:02-107-P-0001 Project name:Development of synthetic parallel corpora / Sintetinių lygiagrečių tekstynų sukūrimas