Mindie/wikiqa-eval

pretty_name: WikiQA Evaluation Dataset tags:

• wikiqa
• information-retrieval
• retrieval
• evaluation
• question-answering
• benchmark

---

WikiQA Evaluation Dataset

This dataset is a processed evaluation dataset derived from the WikiQA development split.

The dataset was constructed for evaluating query-document relevance in information retrieval and retrieval evaluation experiments.

Dataset Description

The original WikiQA dataset provides questions and candidate answer sentences with human-annotated relevance labels.

For this dataset, the WikiQA development split (WikiQA-dev.tsv) was used as the source data. Positive question-sentence pairs were extracted from the original annotations, and additional negative examples were randomly sampled to construct an evaluation dataset with a specified positive-to-negative ratio.

No model-based hard-negative mining was performed when constructing this evaluation dataset.

Source Dataset

This dataset is derived from the WikiQA dataset introduced by:

Yang, Yi, Wen-tau Yih, and Christopher Meek. "WikiQA: A Challenge Dataset for Open-Domain Question Answering." Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2015.

The original WikiQA corpus was released by Microsoft Research.

The source files used for preprocessing were obtained from the following Kaggle repository:

https://www.kaggle.com/datasets/saurabhshahane/wikiqa-corpus

The Kaggle dataset is used here as the source from which the WikiQA development split was obtained; the original dataset should be attributed to Microsoft Research and the original WikiQA authors.

Dataset Construction

The dataset was constructed from WikiQA-dev.tsv using the following procedure:

1. The WikiQA development split was loaded as the source dataset.
2. Question-sentence pairs labeled as relevant in the original WikiQA annotations were extracted as positive examples.
3. Non-relevant question-sentence pairs were randomly sampled from the available negative examples.
4. The number of negative examples was controlled to obtain the desired positive-to-negative ratio.
5. The resulting examples were used as the evaluation dataset.

The purpose of this processing was to create a compact evaluation set while preserving the relevance annotations provided by WikiQA.

Negative Sampling

Negative examples in this dataset are randomly sampled negatives.

Dataset Structure

Each example contains the following fields:

Field	Description
Question	The question from WikiQA
Sentence	A candidate answer sentence
label	Binary relevance label

Where:

• 1 indicates that the sentence is annotated as a relevant answer to the question.
• 0 indicates that the sentence is not annotated as a relevant answer.

Update the field names above if the actual Hugging Face dataset uses different column names.

Intended Use

This dataset can be used for:

• evaluating information retrieval models,
• evaluating query-document relevance models,
• evaluating retrieval evaluators,
• comparing retrieval or reranking methods,
• experiments involving binary relevance classification.

Although this dataset was originally created for retrieval-evaluation experiments, it is not restricted to a particular model or retrieval framework.

Relationship to the Original WikiQA Dataset

This repository contains a processed subset/reconstruction of the WikiQA development split, rather than the original WikiQA corpus.

The original question, sentence, and relevance annotations originate from WikiQA. The positive-example extraction and negative-example sampling used to construct this evaluation dataset were performed separately for this repository.

This dataset should therefore not be considered an official release of the WikiQA benchmark.

License

This dataset is derived from the WikiQA dataset distributed by Microsoft Research under the Open Use of Data Agreement v1.0 (O-UDA-1.0).

The original license permits use, modification, and distribution of the data subject to its terms.

Users of this dataset should review the original license and retain the required attribution and license information.

Original license:

Open Use of Data Agreement v1.0 (O-UDA-1.0)

License identifier: O-UDA-1.0

Attribution

Please attribute the original WikiQA dataset to Microsoft Research and the original authors:

Yang, Yi, Wen-tau Yih, and Christopher Meek. "WikiQA: A Challenge Dataset for Open-Domain Question Answering." Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2015.

Citation

If you use this dataset, please cite the original WikiQA paper:

@inproceedings{yang2015wikiqa,
  title     = {WikiQA: A Challenge Dataset for Open-Domain Question Answering},
  author    = {Yang, Yi and Yih, Wen-tau and Meek, Christopher},
  booktitle = {Proceedings of the 2015 Conference on Empirical Methods
               in Natural Language Processing},
  year      = {2015}
}

Acknowledgements

This dataset is based on the WikiQA corpus released by Microsoft Research.

The author of this repository is responsible only for the preprocessing and sampling procedure described above and does not claim ownership of the original WikiQA data.