RunyaNER

RunyaNER is a Named Entity Recognition (NER) dataset for Runyankore, a Bantu language spoken in western Uganda. It is the first publicly available Runyankore NER benchmark and follows the MasakhaNER 2.0 annotation scheme (PER, LOC, ORG, DATE in BIO format).

The dataset accompanies:

Prosper Arineitwe Asiimwe, Francois Meyer, and Jan Buys. RunyaNER: Auxiliary Language Selection for Runyankore NER.

Statistics

Split	Sentences	Tokens	Named entity spans
Train	15,001	118,622	2,720
Dev	7,498	59,028	1,364
Test	7,508	59,426	1,427
Total	30,007	237,076	5,511

Splits are stratified so that entity distributions stay consistent across train, development, and test.

Sources

RunyaNER is annotated on two publicly available Runyankore–English parallel corpora:

• Sunbird African Language Technology (SALT) — broader domain coverage, including news, public communication, and conversational text.
• Multilingual Parallel Text Corpora (MPTC) — shorter, more regular sentences.

After preprocessing and filtering, about 72% of the sentences come from SALT and 28% from MPTC. Both sources were normalised for encoding artefacts, irregular spacing, and malformed punctuation. Sentences with severe formatting or structural errors were discarded.

Annotation

All sentences were labelled using the MasakhaNER 2.0 guidelines and the four entity types above. All annotation and verification was performed by one of the authors, an L1 speaker of Runyankore. No external annotators were recruited. Before labelling, the annotator studied the MasakhaNER 2.0 guidelines and applied them to a short practice sample. Quality is therefore exhaustive guideline-based correction by a single trained L1 speaker, not inter-annotator agreement.

To reduce annotation time, we used a semi-automated human-in-the-loop workflow:

1. Manually annotate a seed of about 500 sentences, drawn from both sources and including both entity-bearing and entity-free examples.
2. Fine-tune XLM-R on that seed together with Luganda training data from MasakhaNER 2.0.
3. Pre-annotate the remaining corpus with the resulting model.
4. Review and correct every predicted label in Doccano.

Sunflower was consulted only when the English parallel was missing or fragmentary, and only as a reading aid.

Format

The dataset is in CoNLL-style BIO format: one token and one label per line, with sentences separated by a blank line.

Orukungano O
rwa O
African B-ORG
Union I-ORG
rukabaho O
omuri O
Addis B-LOC
Ababa I-LOC
omukwezi B-DATE
kw'okubanza I-DATE
enkumi I-DATE
ibiri I-DATE
ikumi I-DATE
na I-DATE
mushanju I-DATE
. O

Files:

• train.txt
• dev.txt
• test.txt

Usage

RunyaNER can be used for Runyankore NER, cross-lingual transfer, and low-resource multilingual evaluation. For experiments that match the paper, train models on train.txt only and evaluate on test.txt. The development and test splits should not be used to train a tagger or to construct embedding-based similarity metrics.

Citation

If you use RunyaNER, please cite:

@inproceedings{asiimwe-etal-runyaner,
  title     = {{RunyaNER}: Auxiliary Language Selection for {Runyankore} {NER}},
  author    = {Asiimwe, Prosper Arineitwe and Meyer, Francois and Buys, Jan},
  year      = {2026},
  note      = {To appear}
}

License

This dataset is released under the MIT License. SALT and MPTC remain under their original licenses.

Authors

Prosper Arineitwe Asiimwe, Francois Meyer, and Jan Buys
Department of Computer Science, University of Cape Town
arnari002@myuct.ac.za, {francois.meyer, jan.buys}@uct.ac.za