Supplementary materials for our paper "Contextualizing Language Models for Norms Diverging from Social Majority" (Kiehne et al., 2022)

Full paper (EMNLP Findings '22): Here

---

Data available through here or by cloning the repo!*

Update v1.1: We fixed errornous generations in the anti_ms split. For reproducibility, the old data is still there, you can find the new version in anti_ms_llama folder.

* The subfolder original_ms contains the norm-distance split of the Moral Stories dataset by Emelin et al.!

---

Abstract To comprehensibly contextualize decisions, artificial systems in social situations need a high degree of awareness of the rules of conduct of human behavior. Especially transformer-based language models have recently been shown to exhibit some such awareness. But what if norms in some social setting do not adhere to or even blatantly deviate from the mainstream? In this paper, we introduce a novel mechanism based on deontic logic to allow for a flexible adaptation of individual norms by de-biasing training data sets and a task-reduction to textual entailment. Building on the popular ’Moral Stories’ dataset we on the one hand highlight the intrinsic bias of current language models, on the other hand investigate the adaptability of pre-trained models to deviating norms in fine-tuning settings.

---

Code

---

Pre-requisites

Make sure to downloade Social-Chemistry-101 dataset:

1. We require Social Chemistry 101 in "data/social-chem-101/social-chem-101.v1.0.tsv"

   • Downloadable here: https://github.com/mbforbes/social-chemistry-101

2. Our model training and experiment logging is currently outsourced to another repo. It relies on DeepSpeed to dramatically increase training across multiple gpus.

   • FastModelLib needs to be installed from here, TODO!

3. Make sure to install dependencies, e.g. pip install -r requirements.txt.

Running any model training

We support two types of notebook execution:

1. Default jupyter behavior. Just run the notebooks as usual, e.g. cell-by-cell for explorative work.
2. You can use deepspeed and papermill to parametrize the training-notebooks. E.g., there is one central notebook for training the action-classification task, but we use papermill to adapt it to hyperparameters etc. Most of our experiments are run this way, since you can run multiple models in parallel (nice for multi-gpu nodes) and have all of them benefit from the deepspeed speedups!

Data generation

Steps:

1. Train a rot-generator model. Refer to rot-generator.ipynb and the according section in run_model_training.ipynb
2. Run create_flipped_ms.ipynb to create your own contrary norms!

Experiments

Note that ~6TB of disk space are needed if all model checkpoints should be kept.

run_model_training.ipynb: Main class for experiments. Here, all hyper-parameter configurations are deployed and evaluated.

train_action_classification.ipynb, train_textual_entailment, train_polarity_classifier:

• Model training procedure for action classification, textual entailment and polarity components.
• These are deployed via run_model_training.ipynb, but can also be run standalone (without DeepSpeed).

Evaluation

rater-agreement.ipynb: Contains the interrater agreement scores.

---

Citation

@inproceedings{kiehne-etal-2022-contextualizing,
    title = "Contextualizing Language Models for Norms Diverging from Social Majority",
    author = "Kiehne, Niklas  and
      Kroll, Hermann  and
      Balke, Wolf-Tilo",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-emnlp.339",
    pages = "4620--4633",
    }