PoliLean Reposiotry

This is the official repo for From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair NLP Models @ ACL 2023.

Evaluate the Political Leaning of Language Models

Any environment with the HuggingFace Transformers that support pipelines should work. You might need to additionally install selenium for step 3.

Step 0: Sanity Check

We mainly implement things with the text generation pipeline of Huggingface Transformers. Check out your HuggingFace model compatibility by running:

python step0_hftest.py --model <your_model> --device <your_device>

If you see success! printed out, you are good to go. If not, or if your model is not compatible with Huggingface Transformers (e.g. OpenAI models), you can skip this step. The default device is -1 (CPU), but you can specify a GPU device by setting --device <your_device>.

Step 1: Generate Responses

If your step 0 is successful, run:

python step1_response.py --model <your_model> --device <your_device>

There should be a jsonl file in response/ with your model name. If you want to generate responses with your own prompts, you can modify line 22: make sure to keep the <statement> placeholder in your prompt template.

Note that 1) we only prompt once for clarity and efficiency, while the paper used an average of 5 runs; 2) we used the default prompt in the script, while different models might work better with different prompts to better elicit political biases. These two factors, among others (e.g. LM checkpoint update, etc.), that might lead to result varaition.

If your model is not compatible with Huggingface Transformers, feel free to get them to respond to the political statements in response/example.jsonl in your own fashion, change the response fields, and save the file as response/<your_model>.jsonl.

Step 2: Get Agree/Disagree Scores

We use an NLI-based model to evaluate whether the response agrees or disagrees with the political statement. Run:

python step2_scoring.py --model <your_model> --device <your_device>

There should be a txt file in score/ with your model name. Each line presents the agree/disagree probabilities for each political statement.

Step 3: Get Political Leaning with the Political Compass Test

Important: Run this step on your local computer. We need to use selenium to simulate the Chrome browser and auto-click based on the scores in step 2.

1. Download the Chrome browser execuatble at link. Make sure to check the current version of your Chrome browser and download the same version.

2. Download the adblocker crx file at link.

3. change the paths to the browser executable and adblocker in step3_testing.py (lines 64 and 69).

4. Run python step3_testing.py --model <your_model>. The script will automatically open the Chrome browser and take the test. The final political leaning will be displayed on the website. Please note that the browser will first be on the adblocker tab, make sure not to close it and switch to the political compass test tab after the ad blocker is successfully loaded.

Partisan Corpora and Language Models

For partisan news corpora, visit POLITICS. For partisan social media corpora, please include your name, affiliation, institutional email address and apply for access at link. Due to ethical concerns, we are not directly releasing the further pre-trained partisan language models.

Downstream Tasks

Hate speech detection: link

Fake news detection: link

Citation

If you find this repo useful, please cite our paper:

@inproceedings{feng-etal-2023-pretraining,
    title = "From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair {NLP} Models",
    author = "Feng, Shangbin  and
      Park, Chan Young  and
      Liu, Yuhan  and
      Tsvetkov, Yulia",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.656",
    doi = "10.18653/v1/2023.acl-long.656",
    pages = "11737--11762",
    abstract = "Language models (LMs) are pretrained on diverse data sources{---}news, discussion forums, books, online encyclopedias. A significant portion of this data includes facts and opinions which, on one hand, celebrate democracy and diversity of ideas, and on the other hand are inherently socially biased. Our work develops new methods to (1) measure media biases in LMs trained on such corpora, along social and economic axes, and (2) measure the fairness of downstream NLP models trained on top of politically biased LMs. We focus on hate speech and misinformation detection, aiming to empirically quantify the effects of political (social, economic) biases in pretraining data on the fairness of high-stakes social-oriented tasks. Our findings reveal that pretrained LMs do have political leanings which reinforce the polarization present in pretraining corpora, propagating social biases into hate speech predictions and media biases into misinformation detectors. We discuss the implications of our findings for NLP research and propose future directions to mitigate unfairness.",
}