Scruples

A corpus and code for understanding norms and subjectivity.

This repository is for the paper: Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-life Anecdotes. Scruples provides datasets for studying norm understanding in anecdotes and language. This repo contains code for building and analyzing the Scruples datasets, running the baselines, demoing the models and the BEST estimator, and using the BEST estimator directly to estimate the best possible performance on classification datasets.

Jump to a section of this readme to accomplish different goals:

Data: Download the Scruples data.
Demos: View or run demos for the BEST estimator or the model trained to predict people's ethical judgments on anecdotes and moral dilemmas.
Setup: Install the code in this repository.
Quickstart: Get started running the code in this repo.
Citation: Cite the Scruples project.
Contact: Reach out with questions or comments.
Disclaimer: Understand the intended purpose of this work as well as it's limitations.

In addition, the following documents dive deep into particular topics:

Annotation Guidelines: Learn how we annotated data for Scruples to validate the extraction performance.
Demos: Set up and run the demos on your own machine.

Note: This repository is intended for research purposes only. It is NOT intended for use in production environments, and there is no intention for ongoing maintenance. See the Disclaimer for more information.

Data

Scruples has two primary datasets: the Anecdotes and the Dilemmas.

Anecdotes

The Anecdotes provide 32,000 anecdotes of real-life situations with ethical judgments collected from community members about who was in the wrong. See Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-life Anecdotes for more information.

You can download the Anecdotes here.

Dilemmas

The Scruples Dilemmas provide 10,000 ethical dilemmas in the form of paired actions, where the model must identify which one was considered less ethical by crowd workers on Mechanical Turk. See Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-life Anecdotes for more information.

You can download the Dilemmas here.

Demos

Scruples has two demos associated with it.

Scoracle

Visit scoracle to compute the BEST (Bayesian Estimated Score Terminus) performance for a classification dataset. BEST uses the annotations to estimate the upper bound for how well models can possibly do on a dataset under various metrics (accuracy, cross entropy, etc.). See the paper for details.

Norms

The norms demo shows how current neural models can learn to predict basic ethical judgments using the Scruples data. It let's you run anecdotes and dilemmas through a model to view its predictions. In addition, it visualizes how Dirichlet-multinomial layers allow models to separate intrinsic from model uncertainty. The paper elaborates on these techniques.

Running the Demos

Running the demos yourself is quite easy! If you want to run these demos on your own hardware, check out the demo documentation.

Setup

This project requires Python 3.7, and was tested with Python 3.7.6 specifically. To setup the project:

Make sure you have the MySQL client (on ubuntu):
```
sudo apt-get install libmysqlclient-dev
```
Upgrade pip:
```
pip install --upgrade pip
```
Install PyTorch using the directions on their site.
Install the package requirements:
```
pip install -r requirements.txt
```
Install this repository:
```
pip install --editable .
```
Download the english model for spacy:
```
python -m spacy download en
```
(optional) Run the tests to make sure that everything is working. They'll take about 5 minutes to complete, or you can pass the --skip-slow (-s) option to run a smaller, faster test suite:
```
pip install pytest
pytest
```

Quickstart

Once you've installed the package, you'll have the Scruples CLI available. It's a hierarchical, self-documenting CLI that contains all the commands necessary to build and analyze Scruples:

$ scruples --help
Usage: scruples [OPTIONS] COMMAND [ARGS]...

  The command line interface for scruples.

Options:
  --verbose  Set the log level to DEBUG.
  --help     Show this message and exit.

Commands:
  analyze   Run an analysis.
  demo      Run a demo's server.
  evaluate  Evaluate models on scruples.
  make      Make different components of scruples.

To build the dataset, you'll need to download the reddit posts and comments. The initial version of Scruples used November 2018 through April 2019.

Also, Scruples comes with demos that you can run and view locally in the browser:

$ scruples demo --help
Usage: scruples demo [OPTIONS] COMMAND [ARGS]...

  Run a demo's server.

Options:
  --help  Show this message and exit.

Commands:
  norms     Serve the norms demo.
  scoracle  Serve the scoracle demo.

Citation

If you build off of this code, data, or work, please cite the paper as follows:

@article{Lourie2020Scruples,
    author = {Nicholas Lourie and Ronan Le Bras and Yejin Choi},
    title = {Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-Life Anecdotes},
    journal = {arXiv e-prints},
    year = {2020},
    archivePrefix = {arXiv},
    eprint = {2008.09094},
}

Contact

For public, non-sensitive questions and concerns, please file an issue on this repository.

For private or sensitive inquiries email mosaic on the allenai.org website.

Disclaimer

This code and corpus is intended for research purposes only.

As AI agents become more autonomous, they must understand and apply human ethics, values, and norms. One step towards better norm understanding is to reproduce normative judgments drawn from various communities. This skill would enable computers to anticipate people's reactions and understand deeper situational context.

Scruples encourages progress on this research problem by providing a corpus of real-world ethical situations with community sourced normative judgments. The norms expressed by this corpus represent those of the community from which they're drawn—and thus they are neither representative of other communities nor necessarily the right norms to use in any particular application scenario.

Any organization looking to incorporate normative understanding into a product or service should carefully consider, investigate, and evaluate which norms are correct for their particular application.

allenai/scruples