KanuriMT: A Low-Resource English–Kanuri Parallel Corpus and Benchmark for Neural Machine Translation

Dataset Description

KanuriMT is the first publicly available cleaned, annotated English–Kanuri parallel corpus designed for Neural Machine Translation (NMT) research. Kanuri is a low-resource Nilo-Saharan language spoken primarily in northeastern Nigeria, Niger, Chad, and Cameroon, with an estimated 4–8 million speakers.

• Language pair: English (en) → Kanuri (knc, Latin script)
• Total pairs: 7,513
• Created by: Isah Mallam Bukar, Yobe State University, Nigeria
• Supervisor: Bashir Maina Saleh, Yobe State University, Nigeria
• License: CC BY 4.0 (see license notes below)

⚠️ License Notice: KanuriMT is a derivative work compiled from three source corpora, each carrying its own license. Users must comply with all applicable source licenses as described in the Data Sources section below.

---

Dataset Structure

Columns

Column	Type	Description
id	int	Unique row identifier
source	string	English sentence
target	string	Kanuri translation (Latin script)
source_lang	string	Source language code (en)
target_lang	string	Target language code (kr)
domain	string	Text domain (medical, science, news, government, education, legal, social, general)
polarity	string	Sentiment polarity (positive, negative, neutral)
is_question	string	Whether the sentence is a question (true/false)

Data Splits

Split	Size	%
Train	6,010	80%
Validation	751	10%
Test	752	10%
Total	7,513	100%

---

Corpus Statistics

Basic Statistics

Statistic	English	Kanuri
Total tokens	104,742	102,452
Vocabulary size	10,791	20,518
Type-token ratio	0.103	0.200
Avg sentence length (tokens)	13.94	13.64
Median sentence length (tokens)	11	10
Min sentence length	4	1
Max sentence length	79	89

Note: The Kanuri vocabulary size (20,518) is nearly twice that of English (10,791), reflecting the agglutinative morphology of the Kanuri language where multiple affixes combine with root words to form complex word forms.

Domain Distribution

Domain	Count	%
Social	1,993	26.5%
General	1,781	23.7%
Medical	1,366	18.2%
Science	869	11.6%
News	630	8.4%
Government	393	5.2%
Education	262	3.5%
Legal	219	2.9%

Sentiment Polarity Distribution

Polarity	Count	%
Neutral	4,940	65.8%
Negative	1,478	19.7%
Positive	1,095	14.6%

Sentence Type Distribution

Type	Count	%
Statements	7,000	93.2%
Questions	513	6.8%

---

Baseline Results

Machine Translation (BLEU scores on test set, n=752)

System	BLEU	Notes
Google Translate	4.34	Partial Kanuri lexical coverage
NLLB-200 (600M)	1.19	Listed as supported but severe repetition artifacts
Helsinki-NLP opus-mt	N/A	No English–Kanuri model available

Sentiment Classification (weighted F1 on test set, n=752)

System	Accuracy	F1 (weighted)	Type
TextBlob	51.93%	0.5356	Rule-based
VADER	55.79%	0.5808	Rule-based
mBERT fine-tuned	74.20%	0.7256	Neural

---

Dataset Creation

Data Sources and Attribution

KanuriMT was compiled by merging, cleaning, and annotating data from three publicly available parallel corpora. Full attribution and license terms for each source are provided below.

---

1. TWB Gamayun Portal

• Provider: Translators Without Borders (TWB)
• Portal: https://gamayun.translatorswithoutborders.org
• License: TWB Gamayun Portal Access and License Agreement
• Copyright: Copyright © 2020, Translators Without Borders – US, Inc.
• Usage: Used under the TWB open data license. This dataset is a Derivative Work as defined in the TWB license agreement. Any redistribution of this dataset must comply with the TWB license terms, including that Derivative Works be made available at no cost and in an open manner.
• Attribution: "Copyright 2020, Translators Without Borders – US, Inc."

---

2. TICO-19 (via OPUS)

• Provider: Translation Initiative for COVID-19 (TICO-19)
• OPUS URL: http://opus.nlpl.eu/tico-19-v2020-10-28.php
• License: Creative Commons CC0 1.0 Universal (Public Domain)
• Description: A collection of COVID-19 related translation memories compiled by the Translation Initiative for COVID-19.
• Citation: Anastasopoulos et al. (2020). TICO-19: the Translation Initiative for COVID-19. Proceedings of the EMNLP 2020 Workshop on NLP for COVID-19.

---

3. Tatoeba (via OPUS)

• Provider: Tatoeba Project
• OPUS URL: http://opus.nlpl.eu/Tatoeba-v2023-04-12.php
• License: Creative Commons CC BY 2.0 FR
• Copyright: See https://tatoeba.org/eng/terms_of_use
• Description: A collection of translated sentences from the Tatoeba community translation project.
• Citation: Tiedemann, J. (2012). Parallel Data, Tools and Interfaces in OPUS. In Proceedings of LREC 2012.

---

Cleaning Pipeline

The raw merged corpus was processed through the following steps:

1. Deduplication — Removal of exact duplicate English sentences
2. Artifact removal — Filtering of image references, table headers, email headers, and document formatting artifacts
3. Length filtering — Removal of paragraph-length sequences (>80 words) and fragments shorter than 4 words
4. LLM-based domain classification — Each sentence classified into one of 8 domains using Groq LLaMA-3.1-8B-Instant
5. LLM-based sentiment annotation — Two-pass sentiment labeling (positive/negative/neutral) with cross-validation using Groq LLaMA-3.1-8B-Instant
6. LLM-based question classification — Sentences classified as questions or statements using Groq LLaMA-3.1-8B-Instant

---

License

The KanuriMT dataset — including all cleaning, annotations, domain labels, sentiment labels, question flags, and derived splits — is released by the authors under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

However, the underlying source data retains its original licenses:

Source	License
TWB Gamayun	TWB Gamayun License Agreement
TICO-19	CC0 1.0 (Public Domain)
Tatoeba	CC BY 2.0 FR

By using KanuriMT, you agree to comply with the terms of all applicable source licenses in addition to the CC BY 4.0 license covering the KanuriMT annotations and derived work.

---

Usage

from datasets import load_dataset

dataset = load_dataset("IsahMBukar/kanurimt")

# Access splits
train = dataset["train"]
val   = dataset["validation"]
test  = dataset["test"]

# Example row
print(train[0])
# {
#   'id': '1',
#   'source': 'how long have you had this fever?',
#   'target': 'Kanuri translation here...',
#   'source_lang': 'en',
#   'target_lang': 'kr',
#   'domain': 'medical',
#   'polarity': 'neutral',
#   'is_question': 'true'
# }

---

Citation

If you use KanuriMT in your research, please cite:

@dataset{bukar2026kanurimt,
  title     = {{KanuriMT}: A Low-Resource {English}--{Kanuri} Parallel 
               Corpus and Benchmark for Neural Machine Translation},
  author    = {Bukar, Isah Mallam and Saleh, Bashir Maina},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/IsahMBukar/kanurimt},
  license   = {CC BY 4.0}
}

Please also cite the original data sources:

@misc{twb2020gamayun,
  title        = {Gamayun Open Data Portal},
  author       = {{Translators Without Borders}},
  year         = {2020},
  howpublished = {\url{https://gamayun.translatorswithoutborders.org}},
  note         = {Copyright 2020, Translators Without Borders -- US, Inc.}
}

@inproceedings{anastasopoulos2020tico,
  title     = {{TICO-19}: the Translation Initiative for {CO}vid-19},
  author    = {Anastasopoulos, Antonios and others},
  booktitle = {Proceedings of the EMNLP 2020 Workshop on NLP for COVID-19},
  year      = {2020}
}

@inproceedings{tiedemann2012parallel,
  title     = {Parallel Data, Tools and Interfaces in {OPUS}},
  author    = {Tiedemann, J{\"{o}}rg},
  booktitle = {Proceedings of the 8th International Conference on 
               Language Resources and Evaluation (LREC 2012)},
  year      = {2012}
}

@misc{tatoeba2023,
  title        = {Tatoeba: Collection of Translated Sentences},
  author       = {{Tatoeba Project}},
  year         = {2023},
  howpublished = {\url{https://tatoeba.org}},
  note         = {License: CC BY 2.0 FR}
}

---

Contact

• Author: Isah Mallam Bukar
• Institution: Yobe State University, Damaturu, Yobe State, Nigeria
• Language: Kanuri (Central Kanuri, ISO 639-3: knc)
• HuggingFace: IsahMBukar