TiQuAD: Tigrinya Question Answering Dataset

TiQuAD is a human-annotated question-answering dataset for the Tigrinya language. The dataset contains 10.6K question-answer pairs (6.5K unique questions) spanning 572 paragraphs extracted from 290 news articles covering diverse topics including politics, sports, culture, and events.

Published with the paper: Question-Answering in a Low-resourced Language: Benchmark Dataset and Models for Tigrinya (ACL 2023)

Related repositories:

The paper's GitHub repository
Tigrinya-SQuAD, a silver dataset for training purposes

Dataset Overview

TiQuAD follows the SQuAD format and is designed for extractive question answering in Tigrinya, a low-resource Semitic language primarily spoken in Eritrea and Ethiopia. The dataset features:

Human-annotated: All question-answer pairs are created by native Tigrinya speakers
Multiple annotations: Validation and test sets include up to 3 answer annotations per question
Diverse content: Questions span factual, inferential, and analytical reasoning types
Article-based partitioning: Train/validation/test splits are divided by source articles to prevent data leakage. An article could be split into multiple paragraphs when applicable.

Split	Articles	Paragraphs	Questions	Answers
Train	205	408	4,452	4,454
Validation	43	76	934	2,805
Test*	42	96	1,122	3,378
Total	290	572	6,508	10,637

*Test set is not publicly available (see section below)

Dataset Construction

The dataset was constructed through a careful multi-stage process:

Article Collection: 290 news articles were gathered from Tigrinya news sources
Paragraph Extraction: 572 informative paragraphs were selected and cleaned
Question Generation: Native speakers created questions based on paragraph content
Answer Annotation: Extractive answers were marked with character-level spans
Multi-annotation: Validation and test samples received additional annotations for reliability

How to Load TiQuAD

Install the datasets library installed by running pip install -U datasets in the terminal.

Make sure the latest datasets library is installed as older versions may not properly load the data.

Then pull and load the dataset using Python, as follows:

from datasets import load_dataset

# Load the dataset
tiquad = load_dataset("fgaim/tiquad")
print(tiquad)

That will print the dataset features:

DatasetDict({
    train: Dataset({
        features: ['id', 'question', 'context', 'answers', 'article_title', 'context_id'],
        num_rows: 4452
    })
    validation: Dataset({
        features: ['id', 'question', 'context', 'answers', 'article_title', 'context_id'],
        num_rows: 934
    })
})

Data Fields

id: Unique identifier for each question
question: The question to be answered (in Tigrinya)
context: The paragraph containing the answer (in Tigrinya)
answers: A list of dictionaries of candidate answers, each entry containing:
- text: An answer string (training data has one answer per question)
- answer_start: A starting position of answer string in the context
article_title: Title of the source article
context_id: Unique identifier of the context in the data split

Sample Entry

{
    "id": "5dda7d3e-f76f-4500-a3af-07648a1afa51",
    "title": "ሃብቶም ክብረኣብ (ሞጀ)",
    "context": "ሃብቶም ክብረኣብ (ሞጀ)\nሞጀ ኣብ 80’ታትን ኣብ ፈለማ 90’ታትን ካብቶም ናይ ክለብ ኣልታሕሪር ንፉዓት ተኸላኸልቲ ነይሩ፣ ብድሕሪ’ዚ’ውን ኣብ ውድድራት ሓይልታት ምክልኻል ንንውሕ ዝበለ ዓመታት ከም ኣሰልጣኒ ክለብ በኒፈር ኮይኑ ዝነጥፍ ዘሎ ገዲም ተጻዋታይን ኣሰልጣንን’ዩ። ምሉእ ስሙ ሃብቶም ክብርኣብ (ሞጀ) እዩ። ሞጀ ብ1968 ኣብ ኣስመራ ተወሊዱ ዓብዩ። ንሱ ካብ ንኡስ ዕድሚኡ ብኩዕሶ ጨርቂ ጸወታ ጀሚሩ። ብድሕሪኡ ብደረጃ ምምሕዳር ኣብ ዝካየድ ዝነበረ ናይ ‘ቀበሌ’ ጸወታታት ምስ ጸሓይ በርቂ ምስ እትበሃል ጋንታ ተጻዊቱ። ኣብ 1987 ምስ ዳህላክ እትበሃል ዝነበረት ጋንታ ንሓደ ዓመት ድሕሪ ምጽዋቱ ከኣ ኣብ መወዳእታ ወርሒ 1987 ናብ ጋንታ ፖሊስ (ናይ ሎሚ ኣልታሕሪር) ብምጽንባር ክሳብ 1988 ተጻዊቱ። ምስታ ናይ ቅድሚ ናጽነት ጋንታ ፖሊስ ኣብ ዝተጻወተሉ ሰለስተ ዓመታት ከኣ ዝተፈላለየ ዓወታት ተጐናጺፉ ዋናጩ ከልዕል በቒዑ’ዩ። ድሕሪ ናጽነት ስም ክለቡ ኣልታሕሪር ምስ ተቐየረ፣ ሞጀ ናይታ ክለብ ተጻዋታይ ኮይኑ ውድድሩ ቀጺሉ። ኣብ መጀመርታ ናጽነት (1991) ኣብ ዝተኻየደ ናይ ፋልማይ ዋንጫ ስውኣት መን ዓተረ ውድድር ሞጀ ምስ ክለቡ ኣልታሕሪር ዋንጫ ከልዕል በቒዑ። ብዘይካ’ዚ ኣብ 1992 ብብሉጻት ተጻወትቲ ተተኽቲኻ ዝነበረት ኣልታሕሪር ናይ ፋልማይ ዋንጫ ናጽነት ከምኡ’ውን ሻምፕዮን ክትከውን ከላ ሞጀ ኣባል’ታ ጋንታ ነይሩ። ምስ ክለብ ኣልታሕሪር ፍቕርን ሕውነትን ዝመልኦ ምቁር ናይ ጸወታ ዘመን ከም ዘሕለፈ ዝጠቅስ ሞጀ፣ ምስ ኣልታሕሪር ናብ ከም ሱዳንን ኢትዮጵያን ዝኣመሰላ ሃገራት ብምጋሽ ኣህጉራዊ ጸወታታት’ውን ኣካይዱ’ዩ።",
    "question": "ኣልታሕሪር ናይ ቅድም ስማ እንታይ ኔሩ?",
    "answers": {
        "text": ["ፖሊስ", "ፖሊስ", "ጋንታ ፖሊስ"], 
        "answer_start": [414, 414, 410]
    }
}

Note: Samples in the validation set include up to three answer annotations by different annotators to enable human performance estimation and robust evaluation.

TiQuAD Test Set Access

To maintain evaluation integrity and avoid data contamination in models trained on web-crawled data, the test set of TiQuAD is not publicly available. This practice helps ensure fair and reliable evaluation of question-answering systems.

For Researchers

If you are a researcher interested in evaluating your models on the TiQuAD test set, please email [email protected] with the following information:

Subject: TiQuAD Test Set Request
Your full name
Affiliation (university or company)
Purpose and usage plan of request for TiQuAD test set
Acknowledgment that the dataset will be used for evaluation only, not for model training

We will review your request and share the test set directly with appropriate usage agreements. This controlled access helps maintain the dataset's value for fair benchmarking while supporting legitimate research.

Evaluation Protocol

For reproducible evaluation, we recommend:

Using exact match (EM) and token-level F1 score as primary metrics
Handling multiple reference answers appropriately (max score across references)
Reporting results on both validation and test sets when available
Or use the official TiQuAD evaluation script for consistency, which also handles normalization of common articles.

Baseline Results

The original paper reports the following baseline results on TiQuAD:

Model	EM (Val)	F1 (Val)	EM (Test)	F1 (Test)
mBERT	42.1	58.6	39.7	56.2
XLM-R	45.8	62.4	43.1	59.8

For detailed experimental setup and additional baselines, please refer to the original paper.

Citation

If you use this dataset in your work, please cite:

@inproceedings{gaim-etal-2023-tiquad,
    title = "Question-Answering in a Low-resourced Language: Benchmark Dataset and Models for {T}igrinya",
    author = "Fitsum Gaim and Wonsuk Yang and Hancheol Park and Jong C. Park",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.661",
    pages = "11857--11870",
}

Data Concerns and Corrections

We acknowledge that we do not own the copyright to the original news articles from which this dataset was derived. The articles were used under fair use principles for academic research purposes only.

If you identify any of the following issues with the dataset, please contact us:

Sensitive or inappropriate content that should be removed from the dataset
Annotation errors or inconsistencies in the question-answer pairs
Copyright concerns regarding the use of specific content
Factual inaccuracies in the source material or annotations

How to report issues:

Email us with subject "TiQuAD Data Concern"
GitHub: Open an issue at https://github.com/fgaim/TiQuAD

We are committed to addressing legitimate concerns promptly and maintaining the quality and appropriateness of this resource for the research community.

Acknowledgments

We thank the native Tigrinya speakers who contributed to the annotation process and to the Eritrean Ministry of Information (https://shabait.com) and Hadas Ertra newspaper whose articles formed the basis of this dataset. This work contributes to advancing NLP research for low-resource African languages.

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.