This repository contains the datasets, code, and scripts to conduct the analysis in paper in Lin and Ng (2023).
Ruixi Lin and Hwee Tou Ng (2023). Mind the Biases: Quantifying Cognitive Biases in Language Model Prompting. Proceedings of the Findings-ACL 2023.
Please cite:
@inproceedings{lin-ng-2023-mind,
title = "Mind the Biases: Quantifying Cognitive Biases in Language Model Prompting",
author = "Lin, Ruixi and Ng, Hwee Tou",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
year = "2023",
url = "https://aclanthology.org/2023.findings-acl.324",
pages = "5269--5281"
}
We release our saved models (prompt-based fine-tuned BERT models), which are used for measuring the availability bias scores and the framing effect paraphrase conditional probabilities.
Table of contents
Run availability bias experiments
Run framing effect experiments
Go to /your/home/CBPrompt, run
source setup_env.sh
Activate your environment,
source env/bin/activate
For software, we build upon the open source code of ADAPET (Tam et al., 2021) and added our scripts and code for the experiments in this work. For the DDI dataset used in this work, we preprocess the DDIExtraction dataset (Segura-Bedmar et al., 2013) for our experiments, and the processed datasets can be found at ADAPET-CB/data/ddi.
Go to /your/home/CBPrompt/run_avail, copy ADAPET-CB to the current directory
- You can train your own models, for full training and few-shot training settings
./run_single_ddi.sh
- Get test predictions for the dummy prompts on the pretrained models, and compute availability bias scores. Specify paths to pretrained models in run_avail_test.sh, and in the pretrained model directory, add config files for 3 repeated runs (for 3 repeats of sample drawing) according to config_avail_100_0_example.json, then
cd /your/home/CBPrompt/ADAPET-CB
./run_avail_test.sh
If you want to create your own dummy test prompts on biomedical texts, functions in avail_bias_util.py and umls_mapping.py may be useful.
Go to /your/home/CBPrompt/run_framing, copy ADAPET-CB to the current directory
- Get the original and GPT-3 generated toned-down paraphrase datasets (including train, val, test sets) at
/your/home/CBPrompt/ADAPET-CB/data/ddi/framing/ori/ train.jsonl val.jsonl test.jsonl
/your/home/CBPrompt/ADAPET-CB/data/ddi/framing/gpt_completion/ train.jsonl val.jsonl test.jsonl
If you want to generate your own GPT-3 paraphrases with custom data, kindly refer to get_data.sh
- You can train your own models on the original data and paraphrase data
./run_framing_exp_train.sh
- Get pretrained models, evaluate the models on original and paraphrase test data, where $framing_ori and $framing_gpt are your paths to trained model directories, e.g., /your/home/CBPrompt/ADAPET-CB/exp_out/generic/bert-base-uncased/20XX-XX-XX-XX-XX-XX
./run_framing_exp_eval.sh $framing_ori $framing_gpt
- Get Framing Effect Paraphrase (FEP) scores, and compute conditional probabilies, given FEP ranges
python measure_framing_effect.py \
-ori_train /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/ori/train.jsonl \
-ori_val /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/ori/val.jsonl \
-ori_test /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/ori/test.jsonl \
-gpt_train /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/gpt_completion/train.jsonl \
-gpt_val /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/gpt_completion/val.jsonl \
-gpt_test /your/home/CBPrompt/ADAPET-CB/data/ddi/framing/gpt_completion/test.jsonl \
The source code and models in this repository are licensed under the GNU General Public License v3.0 (see LICENSE). For commercial use of this code and models, separate commercial licensing is also available. Please contact Ruixi Lin ([email protected]) and Prof. Hwee Tou Ng ([email protected]).