AfriXIR Passage Retrieval with Hard Negatives
Dataset Summary
AfriXIR-Passage-Retrieval-With-Negatives is a large-scale, multilingual passage retrieval dataset designed for training and evaluating information retrieval systems across eight African languages. The dataset contains 313,592 training examples and 34,849 test examples, with each example consisting of a query-passage pair accompanied by 16 hard negative passages mined through semantic similarity.
This dataset enables the development of robust cross-lingual and multilingual retrieval models for underrepresented African languages, addressing the critical need for information retrieval resources in low-resource language settings.
Languages Covered
The dataset includes eight African languages from diverse language families:
| Language | ISO Code | Train Examples | Test Examples | % of Dataset |
|---|---|---|---|---|
| Swahili | sw | 76,288 | 8,477 | 24.3% |
| Hausa | ha | 68,526 | 7,614 | 21.9% |
| Amharic | am | 61,469 | 6,832 | 19.6% |
| Yoruba | yo | 32,728 | 3,637 | 10.4% |
| Somali | so | 29,755 | 3,307 | 9.5% |
| Shona | sn | 17,010 | 1,890 | 5.4% |
| Zulu | zu | 14,620 | 1,625 | 4.7% |
| Oromo | om | 13,196 | 1,467 | 4.2% |
Total: 313,592 train + 34,849 test = 348,441 query-passage pairs
Data Collection Methodology
Source Datasets
The dataset was constructed through a weak supervision approach by aggregating and repurposing multiple publicly available datasets. Query-passage pairs were extracted from:
- News Articles (MasakhaNEWS, AfriCast, DSFSI News)
- Headlines → Full article text pairs
- Available for: All 8 languages
- Provides real-world journalistic content
- XL-SUM Summarization Dataset
- Summaries → Full text pairs
- Available for: Amharic, Hausa, Oromo, Somali, Swahili, Yoruba, Zulu
- High-quality editorial content
- Wikipedia Articles
- Title → Article content pairs
- Available for: All 8 languages
- Filtered by length (100-1000 characters)
- Encyclopedic knowledge coverage
- Question-Answer Datasets (AfriQA, KenSwQuAD)
- Questions → Context passages
- Available for: Hausa, Swahili, Yoruba, Zulu
- Natural information-seeking queries
- CIRAL Corpus
- Repurposed pre-existing passages into query-passage pairs
- Available for: Hausa, Somali, Swahili, Yoruba
- Amharic Passage Retrieval Dataset V2 (Ras Yosef, 2024)
- Pre-built query-passage pairs with hard negatives
- Available for: Amharic
- News, XL-SUM, Wikipedia, and QA sources
Hard Negative Mining
Hard negatives are crucial for training discriminative retrieval models. Following the methodology established in Amharic-Passage-Retrieval-Dataset-V2-With-Negatives, for each query-passage pair:
- Embedding Generation: All passages in the dataset were embedded using
sentence-transformers/paraphrase-multilingual-mpnet-base-v2 - Similarity Computation: Cosine similarity was computed between the query embedding and all passage embeddings
- Hard Negative Selection: The top-16 most similar passages (excluding the positive passage) were selected as hard negatives
- Score Retention: Each negative passage includes its cosine similarity score
Data Splitting
- Train/Test Split: 90% / 10% stratified by language and source
- Random Seed: 42 (for reproducibility)
- Minimum Examples per Source: Ensured adequate representation from each source dataset
Acknowledgments
We gratefully acknowledge the creators of the source datasets that made this work possible:
- MasakhaNEWS: Adelani et al. (2023)
- XL-SUM: Hasan et al. (2021)
- AfriQA: Ogundepo et al. (2023)
- CIRAL: Nair et al. (2022)
- Wikipedia: Wikimedia Foundation
- Amharic Passage Retrieval Dataset: Ras Yosef (2024)
- AfriCast/KenSwQuAD: Various contributors
License
This dataset is released under the MIT License. Users are free to use, modify, and distribute this dataset with attribution.
- Downloads last month
- 33