metadata

license: apache-2.0
language:
  - ha
  - so
  - sw
  - yo
task_categories:
  - text-retrieval
mutilinguality:
  - multilingual
viewer: true

Dataset Summary

CIRAL is a collection for cross-lingual information retrieval research across four (4) African languages. The collection comprises English queries and query-passage relevance judgements for passages in the African languages. This dataset repo contains only the queries and relevance judgements. The corpus collection can be found here here

Dataset Structure

To download the files: The queries can be found under ciral-{lang}/topics and are in .tsv formats with each line in the form:
```
 qid\tquery
```
while the judgements are in the folder ciral-{lang}/qrels, with each file in the standard TREC format:
```
 qid Q0 docid relevance
```

To access the dataset via datasets:

ciral_dataset = load_dataset("ciral/ciral", "hausa") #or swahili, somali, yoruba

for data in ciral_dataset['dev']:  # or 'testA' or 'testB'
   query_id = data['query_id']
   query = data['query']
   pos_qrels = data['positive_passages']
   neg_qrels = data['negative_passages']

   # To load test set A's pool judgments
   pools_pos_qrels = data['pools_positive_passages']
   pools_neg_qrels = data['pools_negative_passages']

   for qrel in pos_qrels:
      docid = qrel['docid']
      text = qrel['text']

Citation

@misc{CiralHfCite,
  title = {{CIRAL: A Test Suite for {CLIR} in {A}frican Languages}},
  author = {Mofetoluwa Adeyemi and
    Akintunde Oladipo and
    Xinyu Zhang and
    David Alfonso-Hermelo and
    Mehdi Rezagholizadeh and
    Boxing Chen and
    Jimmy Lin},
  year = 2023,
  url = {https://huggingface.co/datasets/CIRAL/ciral},
  urldate = {2023-12-19}
}