Awesome

Safety Score for Pre-Trained Language Models

Paper: An Empirical Study of Metrics to Measure Representational Harms in Pre-Trained Language Models (ACL 2023, TrustNLP workshop)

This repository contains the code used to measure safety scores for pre-trained language models based on ToxiGen human annotated dataset and ImplicitHate dataset.

Evaluation Dataset

• We selected a subset of TxiGen and ImplicitHate datasets. The examples in ImplicitHate subset are either implicit-hate or neutral and we down-sampled the neutral examples to have equal number of harmful and benign exxamples. ImplicitHate does not have any information about the target of the hate for each sentence.
• The examples in ToxiGen dataset include the sentences in which all the annotators agreed on whether the sentence is harmful and more than 2 annotators agreed on the target group of the hate.

Setup

There are few specific dependencies to install before runnung the safety score calculator, you can install them with the command pip install -r requirements.txt.

How to calculate safety score

Now you can run the following script:

python safety_score.py \
   --data data/toxiGen.json \ # Path to evaluation dataset
   --output results \ # local path to a directory for saving results
   --model gpt2 \ # pre-trained model name or local path
   --lmHead clm \ # Type of language model head, i.e. causal or masked
   --force # overwrites the output path if it already exists.

Two files will be saved in the output path:

• 'perplexities.json' which contains the perplexity value for each sentence in the evaluation dataset
• 'safety_scores.json' which contains the statistically significant safety scores for each demographic.

For example, the contetn of 'safety_scores.json' after running the above script is

{"asian": 0.3694922836054574, "black": 0.36662849289967936, "chinese": 0.3731038121619839, "jewish": 0.40661968642101093, "latino": 0.22831884057971014, "lgbtq": 0.2701839434577746, "mental dis": 0.22755361686659398, "mexican": 0.23524720893141945, "middle-eastern": 0.2604830744365628, "muslim": 0.32320982365959877, "native-american": 0.24511818257746595, "physical dis": 0.22460258469801234, "women": 0.23225019516003123}

Safety scores based on ToxiGen

Here are the results based on the ToxiGen dataset:

Safety scores based on ImplicitHate

Here are the results based on the ImplicitHate dataset:

Citation

Please use the following to cite this work:

@misc{hosseini2023empirical,
      title={An Empirical Study of Metrics to Measure Representational Harms in Pre-Trained Language Models}, 
      author={Saghar Hosseini and Hamid Palangi and Ahmed Hassan Awadallah},
      year={2023},
      eprint={2301.09211},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}