Source code and data for ADEPT: A DEbiasing PrompT Framework (AAAI-23).

An illustration of how debiasing works using ADEPT and for downstream tasks:

(a) While debiasing, ADEPT only trains the prompt parameters and keeps the base model frozen.

(b) When performing downstream tasks, ADEPT conditions the base model or both the prompt and the base model.

We recently received feedback from a researcher regarding the code. Part of the original message was:

When evaluating the debiased model using StereoSet, I found that my replication results differ from those in the paper. I suspect there may be an issue with my StereoSet evaluation dataset, and the paper mentions using a "filtered StereoSet-Intrasentence." Could you please provide details on the specific criteria used for filtering?

We appreciate the feedback. Our response is:

Thank you for your question. This is an important issue, and if we have the opportunity to update the next paper version, we will include more relevant details. When testing with StereoSet, we found that the performance of the debiased model is related to whether or not it has encountered the target words during the debiasing process. For example, with the word "schoolgirl" as a target, the debiased model showed less sensitivity, which is related to the fact that the debiasing training objective only includes a limited set of words, making it difficult to generalize the debiased gender subspace to less common words. Therefore, the StereoSet collection we tested was filtered to include only those targets that appeared in the debiasing word list. This was a summer research project from a few years ago, and in the interest of code simplicity, the benchmark testing part was not included (we used a debiasing benchmark repo for that). Unfortunately, due to the loss of machine access at that time, that part of the code was lost. However, using the method described above should allow you to recreate the test dataset.

We conduct experiments on the bert-large-uncased pre-trained model from HuggingFace. By using ADEPT, we need only train 1.97M parameters when prompt-tuning with 40 prompt tokens, orders of magnitude smaller than the 335M parameters required for finetuning.

We provide bash scripts and codes to replicate our findings. Our environment is:

• Ubuntu servers with NVIDIA GeForce RTX 3090 (24G) GPUs
• cuda 11.1
• packages with certain versions

Create environment:

conda create -n debias python=3.8.5
conda activate debias

Install pytorch and python packages:

conda install -n pt2 pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
cd ADEPT
pip install -r requirements.txt

We've already included word lists for attributes in the ./data folder, so there is no need to acquire them from other resources. As for larger corpora, you can download News-Commentary v15 here and Hugging Face's BookCorpus replica here. New-Commentary alone can support gender debiasing. You may need to create a new text file by combining the two corpora mentioned above so that there are sufficient sentences for religion debiasing.

Collect sentences:

cd script
bash ./collect_sentences.sh bert [corpus_path] gender final

Debias:

bash ./debias.sh bert 0 ADEPT gender # for ADEPT
bash ./debias.sh bert 0 DPCE gender # for DPCE

Preprocess corpus for plotting words' correlation (for better visualizing pairwise words' correlation, we highly suggest that you choose a large corpus, like a subset of BookCorpus sampled with function sample_sentences_from_bookcorpus in utils.py, because we have set the minimum threshold for $len(S_m^{a(i)})$ if word $w_m^{a(i)}$ is to be plotted):

bash ./preprocess_plot_word_correlation.sh bert 0 gender [corpus_path]

Plot words' correlation:

bash ./plot_word_correlation.sh bert 0 ADEPT gender [model_name_or_path] # for ADEPT
bash ./plot_word_correlation.sh bert 0 DPCE gender [model_name_or_path] # for DPCE

We color neutral words beige, male words blue, and female words red.

(a) original:

(b) DPCE:

(c) ADEPT-finetuning:

(d) ADEPT:

We color neutral words grey, Judaism words yellow, Christianity words blue, and Islam words green.

(a) original:

(b) ADEPT: