# Dataset: mocha

Languages: en
Multilinguality: monolingual
Size Categories: 10K<n<100K
Language Creators: found
Annotations Creators: crowdsourced
Source Datasets: original

# Dataset Card for Mocha

### Dataset Summary

Posing reading comprehension as a generation problem provides a great deal of flexibility, allowing for open-ended questions with few restrictions on possible answers. However, progress is impeded by existing generation metrics, which rely on token overlap and are agnostic to the nuances of reading comprehension. To address this, we introduce a benchmark for training and evaluating generative reading comprehension metrics: MOdeling Correctness with Human Annotations. MOCHA contains 40K human judgement scores on model outputs from 6 diverse question answering datasets and an additional set of minimal pairs for evaluation. Using MOCHA, we train a Learned Evaluation metric for Reading Comprehension, LERC, to mimic human judgement scores. LERC outperforms baseline metrics by 10 to 36 absolute Pearson points on held-out annotations. When we evaluate robustness on minimal pairs, LERC achieves 80% accuracy, outperforming baselines by 14 to 26 absolute percentage points while leaving significant room for improvement. MOCHA presents a challenging problem for developing accurate and robust generative reading comprehension metrics.

English

## Dataset Structure

### Data Instances

MOCHA contains 40K human judgement scores on model outputs from 6 diverse question answering datasets and an additional set of minimal pairs for evaluation. MOCHA pairs reading comprehension instances, which consists of a passage, question, and reference, with candidates and human judgement scores.

### Data Fields

• constituent_dataset: the original QA dataset which the data instance came from.
• id
• context: the passage content.
• question: the question related to the passage content.
• reference: the correct answer for the question.
• candidate: the answer generated from the reference by source
• score: the human judgement score for the candidate. Not included in test split, defaults to -1
• metadata: Not included in minimal pairs split.
• scores: list of scores from difference judges, averaged out to get final score. defaults to []
• source: the generative model to generate the candidate

In minimal pairs, we'll have an additional candidate for robust evaluation.

• candidate2
• score2

### Data Splits

Dataset Split Number of Instances in Split
Train 31,069
Validation 4,009
Test 6,321
Minimal Pairs 200

## Considerations for Using the Data

CC BY-SA 4.0

### Citation Information

@inproceedings{Chen2020MOCHAAD,
author={Anthony Chen and Gabriel Stanovsky and Sameer Singh and Matt Gardner},
title={MOCHA: A Dataset for Training and Evaluating Generative Reading Comprehension Metrics},
booktitle={EMNLP},
year={2020}
}

### Contributions

Thanks to @mattbui for adding this dataset.

None yet