acl24-trigger-warnings

This repository contains the data and code for If there's a Trigger Warning, then where's the Trigger? Investigating Trigger Warnings at the Passage Level.

For all commands, you can read the docstrings or add --help for usage explanations.

Get the Data

Note: This is obfuscated for review. Please consider the dataset or the excerpt in /samples/ Data will be published via Zenodo for the final release.

Setup

1. Run pip install .
2. Run python3 main.py --help for usage instructions of the individual methods.

Dataset

Unzip the dataset dataset.jsonl in resources.

To reconstruct this file from the annotation result files from label studio, run:

poetry run python3 main.py data parse_ls

Contact the authors to get the source files.

Experiments

sequence_level_trigger_warning_assignment/experiment-runners contains scripts to run every part of the experiments. In particular:

• run-compile-datasets.sh creates the cross-validation folds.
• run-evaluate-dataset.sh created the dataset statistics.
• run-experiments-* runs the respective experiments. Those create score- and prediction- files for each run under resources/models-*. Copy these prediction files to a shared directory for evaluation (default: resources/classification-results/predictions).
• run-evaluate-classification.sh scores the predictions in the evaluation directory in a binary fashion.

Note: For the GPT experiments, you need to add your OpenAI API key in a file resources/prompt/gpt-secret-sentence-triggers.txt

Evaluation

sequence_level_trigger_warning_assignment/notebooks contains jupyter notebooks used to evaluate and plot the classification results. Run the cells in order.