HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering

HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering is a 2018 paper by Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov and Christopher D. Manning, published at EMNLP 2018 (arXiv).

It introduces HotpotQA, a question-answering dataset of around 113k question-answer pairs based on Wikipedia. Answering each question requires finding and reasoning over more than one supporting document (multi-hop reasoning).

The dataset also labels the sentence-level supporting facts needed to answer each question, so models can be trained and evaluated on explaining their answers, not just getting them right.