This is the repository for the paper: Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps (COLING 2020).

• We release the para_with_hyperlink.zip file that contains all articles/paragraphs with hyperlink information (except for some error paragraphs). Each paragraph has the following information:
  • id: Wikipedia id of an article
  • title
  • sentences: a list of sentence
  • mentions: a list of hyperlink information, each element has the following information: id, start, end, ref_url, ref_ids, sent_idx.
• Here is the link of the dataset that has fixed the inconsistency of sentence segmentation in paragraphs.

• Due to the multiple names of an entity in Wikidata, we add evidences_id and answer_id to our dataset. Here are the details:
  • For Inference and Compositional questions: we add to all questions.
  • For Comparison and Bridge_comparison questions: we add to questions that have relations: country, country of origin, and country of citizenship.
• We update the new evaluation script 2wikimultihop_evaluation_v1.1.py. We can use this evaluation script to evaluate the dataset with evidences_id and answer_id.
• We also update the results of the baseline model by using the new evaluation script and the dataset with evidences_id and answer_id. The updated results of tables 5, 6, and 7 in the paper are in the folder update_results.
• Here is the link of the dataset with evidences_id and answer_id. File id_aliases.json is used for evaluation.

To evaluate your model on the test data, please contact us. Please prepare the following information:

1. Your prediction file (follow format in file: prediction_format.json)
2. The name of your model
3. Public repository of your model (optional)
4. Reference to your publication (optional)

The full dataset is in here.

Our dataset follows the format of HotpotQA. Each sample has the following keys:

• _id: a unique id for each sample
• question: a string
• answer: an answer to the question. The test data does not have this information.
• supporting_facts: a list, each element is a list that contains: [title, sent_id], title is the title of the paragraph, sent_id is the sentence index (start from 0) of the sentence that the model uses. The test data does not have this information.
• context: a list, each element is a list that contains [title, setences], sentences is a list of sentences.
• evidences: a list, each element is a triple that contains [subject entity, relation, object entity]. The test data does not have this information.
• type: a string, there are four types of questions in our dataset: comparison, inference, compositional, and bridge-comparison.
• entity_ids: a string that contains the two Wikidata ids (four for bridge_comparison question) of the gold paragraphs, e.g., 'Q7320430_Q51759'.

Our baseline model is based on the baseline model in HotpotQA. The process to train and test is quite similar to HotpotQA.

• Process train data

python3 main.py --mode prepro --data_file wikimultihop/train.json --para_limit 2250 --data_split train

• Process dev data

python3 main.py --mode prepro --data_file wikimultihop/dev.json --para_limit 2250 --data_split dev

• Train a model

python3 -u main.py --mode train --para_limit 2250 --batch_size 24 --init_lr 0.1 --keep_prob 1.0

• Evaluation on dev (Local Evaluation)

python3 main.py --mode test --data_split dev --para_limit 2250 --batch_size 24 --init_lr 0.1 --keep_prob 1.0 --save WikiMultiHop-20201024-023745 --prediction_file predictions/wikimultihop_dev_pred.json

python3 2wikimultihop_evaluate.py predictions/wikimultihop_dev_pred.json data/dev.json

• Use new evaluation script

python3 2wikimultihop_evaluate_v1.1.py predictions/wikimultihop_dev_pred.json data_ids/dev.json id_aliases.json

If you plan to use the dataset, please cite our paper:

@inproceedings{xanh2020_2wikimultihop,
    title = "Constructing A Multi-hop {QA} Dataset for Comprehensive Evaluation of Reasoning Steps",
    author = "Ho, Xanh  and
      Duong Nguyen, Anh-Khoa  and
      Sugawara, Saku  and
      Aizawa, Akiko",
    booktitle = "Proceedings of the 28th International Conference on Computational Linguistics",
    month = dec,
    year = "2020",
    address = "Barcelona, Spain (Online)",
    publisher = "International Committee on Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.coling-main.580",
    pages = "6609--6625",
}

The baseline model and the evaluation script are adapted from https://github.com/hotpotqa/hotpot