WillHeld/MD3

Original Attribution

MD3-EN: The Multi-Dialect Dataset of Dialogues

• Authors: Jacob Eisenstein (jeisenstein@google.com), Clara Rivera, Vinodkumar Prabhakaran, Jon Clark, Dora Demszky, Sunny Mak, Ravi Rajakumar, David Elworthy, Landis Baker, Devyani Sharma
• Paper: https://arxiv.org/abs/2305.11355, published at InterSpeech 2023
• License: CC BY-SA 4.0

Dataset Description

The MD3 Taboo Game dataset consists of transcripts of two people playing the Taboo game, where one person tries to get the other to guess a secret word without saying certain distractor words.

The dataset includes:

• Transcripts from 3 English dialects: Indian English (en_in), Nigerian English (en_ng), and American English (en_us)
• The correct word to be guessed in each game
• Distractor words that cannot be used during the game

Dataset Structure

The dataset is organized into three splits by dialect:

• en_in: Indian English
• en_ng: Nigerian English
• en_us: American English

Each split contains the following fields:

• id: Unique identifier for the dialogue
• transcript: The dialogue transcript
• correct_word: The target word to be guessed
• distractors: List of words that cannot be used during the game

Usage

This dataset can be used to evaluate language models on their ability to understand contextual clues and infer missing information, similar to how a human would play the Taboo game.

Example usage:

from datasets import load_dataset

dataset = load_dataset("REPO_NAME_HERE")
example = dataset["en_us"][0]
print(f"Transcript: {example['transcript']}")
print(f"Correct word: {example['correct_word']}")
print(f"Distractors: {example['distractors']}")

Citation

If you use this dataset, please cite the original MD3 paper:

@inproceedings{eisenstein2023md3,
  title={MD3: The Multi-Dialect Dataset of Dialogues},
  author={Eisenstein, Jacob and Rivera, Clara and Prabhakaran, Vinodkumar and Clark, Jon and Demszky, Dora and Mak, Sunny and Rajakumar, Ravi and Elworthy, David and Baker, Landis and Sharma, Devyani},
  booktitle={Proceedings of INTERSPEECH},
  year={2023}
}

License

This re-release of the dataset is bound by the original MD3 corpus CC BY-SA 4.0 license.