README.md

DREAM

Overview

This repository maintains DREAM, a multiple-choice Dialogue-based REAding comprehension exaMination dataset.

• Paper: https://arxiv.org/abs/1902.00164

@article{sundream2018,
  title={{DREAM}: A Challenge Dataset and Models for Dialogue-Based Reading Comprehension},
  author={Sun, Kai and Yu, Dian and Chen, Jianshu and Yu, Dong and Choi, Yejin and Cardie, Claire},
  journal={Transactions of the Association for Computational Linguistics},
  year={2019},
  url={https://arxiv.org/abs/1902.00164v1}
}

• Leaderboard: https://dataset.org/dream/

Files in this repository:

• data folder: the dataset.
• annotation folder: question type annotations.
• dsw++ folder: code of DSW++.
• ftlm++ folder: code of FTLM++.
• license.txt: the license of DREAM.
• websites.txt: list of websites used for the data collection of DREAM.

Dataset

data/train.json, data/dev.json and data/test.json are the training, development and test sets, respectively. The format of them is as follows:

[
  [
    [
      dialogue 1 / turn 1,
      dialogue 1 / turn 2,
      ...
    ],
    [
      {
        "question": dialogue 1 / question 1,
        "choice": [
          dialogue 1 / question 1 / answer option 1,
          dialogue 1 / question 1 / answer option 2,
          dialogue 1 / question 1 / answer option 3
        ],
        "answer": dialogue 1 / question 1 / correct answer option
      },
      {
        "question": dialogue 1 / question 2,
        "choice": [
          dialogue 1 / question 2 / answer option 1,
          dialogue 1 / question 2 / answer option 2,
          dialogue 1 / question 2 / answer option 3
        ],
        "answer": dialogue 1 / question 2 / correct answer option
      },
      ...
    ],
    dialogue 1 / id
  ],
  [
    [
      dialogue 2 / turn 1,
      dialogue 2 / turn 2,
      ...
    ],
    [
      {
        "question": dialogue 2 / question 1,
        "choice": [
          dialogue 2 / question 1 / answer option 1,
          dialogue 2 / question 1 / answer option 2,
          dialogue 2 / question 1 / answer option 3
        ],
        "answer": dialogue 2 / question 1 / correct answer option
      },
      {
        "question": dialogue 2 / question 2,
        "choice": [
          dialogue 2 / question 2 / answer option 1,
          dialogue 2 / question 2 / answer option 2,
          dialogue 2 / question 2 / answer option 3
        ],
        "answer": dialogue 2 / question 2 / correct answer option
      },
      ...
    ],
    dialogue 2 / id
  ],
  ...
]

Question Type Annotations

annotation/{annotator1,annotator2}_{dev,test}.json are the question type annotations for 25% questions in the development and test sets from two annotators.

In accordance with the format explanation above, the question index starts from 1.

We adopt the following abbreviations:

Abbreviation	Question Type
m	matching
s	summary
l	logic
a	arithmetic
c	commonsense

Code

• DSW++

  1. Copy the data folder data to dsw++/.
  2. Download numberbatch-en-17.06.txt.gz from https://github.com/commonsense/conceptnet-numberbatch, and put it into dsw++/data/.
  3. In dsw++, execute python run.py.
  4. Execute python evaluate.py to get the accuracy on the test set.

• FTLM++

  1. Download the pre-trained language model from https://github.com/openai/finetune-transformer-lm, and copy the model folder model to ftlm++/.
  2. Copy the data folder data to ftlm++/.
  3. In ftlm++, execute python train.py --submit. You may want to also specify --n_gpu (e.g., 4) and --n_batch (e.g., 2) based on your environment.
  4. Execute python evaluate.py to get the accuracy on the test set.

Note: The results you get may be slightly different from those reported in the paper. For example, the dev and test accuracy for DSW++ in this repository is 51.2 and 50.2 respectively, while the reported accuracy in the paper is 51.4 and 50.1. That is due to (1) we refactor the code with different dependencies to make it portable, and (2) some of the code is non-deterministic due to GPU non-determinism.

Environment: The code has been tested with Python 3.6/3.7 and Tensorflow 1.4

Other Useful Code

You can refer to this repository for a finetuned transformer baseline based on BERT.