Code and dataset for The World of an Octopus: How Reporting Bias Influences a Language Model's Perception of Color.

This repository is roughly split into 2 parts:

• probing: The probing implementations, including code for generating CoDa.
• mturk-survey: Instruction pages and used for crowdsourcing annotations.

If you'd like to use CoDa, we highly recommend using the version hosted on the Huggingface Hub as it requires no additional dependencies.

from datasets import load_dataset

ds = load_dataset('corypaik/coda')

You can find more details about how to use Huggingface Datasets here.

This repository is developed and tested on linux systems and uses Bazel. If you are on other platforms, you might consider running Bazel in a docker container. If you'd like more guidance on this, please open an Issue on GitHub.

First, clone the project

# clone project
git clone https://github.com/nala-cub/coda

# goto project
cd coda

You can run the specific tasks as:

# run zeroshot
bazel run //projects/coda/probing/zeroshot
# representation probing
bazel run //projects/coda/probing/representations
# ngrams
bazel run //projects/coda/probing/ngram_stats
# generate dataset from annotations (relative to workspace root)
bazel run //projects/coda/probing/dataset:create_dataset -- \
  --coda_ds_export_dir=<export_dir>

To see help for any of the commands, use:

bazel run <target> -- --help
# for example:
# bazel run //projects/coda/probing/zeroshot -- --help

Annotations were collected using an Angular app on Firebase. The included files contain all instructions, but not the app itself. If you're interested in the latter please open an issue on GitHub.

If this code was useful, please cite the paper:

@misc{paik2021world,
      title={The World of an Octopus: How Reporting Bias Influences a Language Model's Perception of Color},
      author={Cory Paik and Stéphane Aroca-Ouellette and Alessandro Roncone and Katharina Kann},
      year={2021},
      eprint={2110.08182},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

CoDa is licensed under the Apache 2.0 license. The text of the license can be found here.