Luganda STS Benchmark

Overview

This dataset is a Luganda adaptation of the Semantic Textual Similarity (STS) Benchmark. It is designed to evaluate how well language models understand semantic similarity between pairs of Luganda sentences.

The dataset is derived from the original STS Benchmark and is being manually adapted and expanded for Luganda. The goal is to evaluate semantic understanding rather than simple lexical or syntactic matching.

Current Status

The initial version contains approximately 200 sentence pairs from the test split of the original STS Benchmark.

The dataset is under active expansion. Additional examples will be added to the test split, and the development (dev) split will subsequently be expanded.

Split	Current status
Test	Initial Luganda adaptation, ~200 examples
Dev	To be expanded
Train	Not currently included

Task

Each example contains two Luganda sentences representing a semantic relationship. A model should determine the degree to which the two sentences convey the same meaning.

The benchmark is intended to test semantic understanding rather than simple word overlap. Two sentences may use different vocabulary or grammatical structures while expressing essentially the same meaning, while sentences sharing similar words may still express different meanings.

Intended Use

The benchmark can be used to evaluate:

• Luganda sentence embedding models
• Semantic similarity models
• Cross-lingual representation models
• Information retrieval systems
• Multilingual language models
• Natural language understanding systems
• Luganda-specific language models

It can also be used to evaluate whether improvements in Luganda language modeling result in better semantic representations.

Dataset Development

The current test set is an initial release of the benchmark. Additional examples will be progressively adapted from the original STS Benchmark and expanded where necessary.

The adaptation focuses on preserving the semantic relationship between sentence pairs rather than performing literal word-for-word translation. A faithful semantic adaptation may therefore require changes in wording, grammatical structure, or lexical choice in Luganda.

Limitations

The current release is relatively small and should be considered an early development version rather than a complete Luganda STS benchmark.

The test split currently contains approximately 130 examples, while the dev split is still being expanded. Evaluation results on this version should therefore be interpreted with the size of the benchmark in mind.

Future versions are expected to provide broader coverage of semantic relationships and linguistic variation in Luganda.

Attribution

This dataset is a Luganda adaptation of the STS Benchmark, originally introduced as part of SemEval-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual Focused Evaluation.

The original work should be cited as:

Cer, D., Diab, M., Agirre, E., Lopez-Gazpio, I., & Specia, L. (2017). SemEval-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual Focused Evaluation. Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), 1–14.

Citation

If you use this dataset, please cite both the original STS Benchmark and this Luganda adaptation.

Original STS Benchmark

@inproceedings{cer2017semeval,
  title     = {SemEval-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual Focused Evaluation},
  author    = {Cer, Daniel and Diab, Mona and Agirre, Eneko and Lopez-Gazpio, Inigo and Specia, Lucia},
  booktitle = {Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)},
  year      = {2017},
  pages     = {1--14},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/S17-2001}
}

Luganda Adaptation

@misc{matovu2026lugandasts,
  title     = {Luganda STS Benchmark},
  author    = {MATOVU CALEB},
  year      = {2026},
  publisher = {SAGE POND},
  note      = {Luganda adaptation and expansion of the STS Benchmark}
}

Acknowledgements

This benchmark builds upon the work of the STS Benchmark and the researchers who contributed to the original semantic textual similarity evaluation resources.

The Luganda adaptation and ongoing expansion are developed by Caleb Matovu at SAGE POND.