Interactively Generating Explanations for Transformer Language Models

An XAI method to get insights into the decision-making process of transformer language models through prototypical explanations. Additionally a XIL method to interact with the trained (explainable) model.

Setup and Run Model

To setup the model, we need to first:

• download https://github.com/peterbhase/InterpretableNLP-ACL2020/tree/master/text/data/rt-polarity to data/rt-polarity/
• download https://www.yelp.com/dataset to data/restaurant/
• download https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data to data/jigsaw/
• set up a (python) virtual environment meeting requirements.txt
• to run SGPT models, please update the sentence-transformer package using the following repository: https://github.com/Muennighoff/sgpt

The core file is run_proto_nlp.py

To run this model on sentence-/ word-level, e.g. do:

python run_proto_nlp.py --data_name rt-polarity --level sentence --language_model SentBert --compute_emb True

python run_proto_nlp.py --data_name rt-polarity --level word --language_model GPT2 --proto_size 4  --compute_emb True

You can find the results in experiments/train_results/. Further parameter choices are explained in the core file.

Interaction

To interact with the model, you first need to train a model. After the training you can interact with the model in the following fashion:

python run_proto_nlp.py --data_name rt-polarity --level sentence --language_model SentBert --mode replace

The mode sets the interaction type you want to apply. In the core file you can refine the interaction method, i.e. set e.g. the prototype you want to remove or replace. The last trained model is automatically selected for the interaction. If you trained several models and want to interact with a certain one, the path has to be set manually.

Citation

If you like or use our work please cite us:

@inproceedings{friedrich2022hhai,
    title={Interactively Providing Explanations for Transformer Language Models},
    author={Felix Friedrich and Patrick Schramowski and Christopher Tauchmann and Kristian Kersting},
    year={2022},
    Keywords = {Transformer, Large Language Models, Prototype Layers, Explainable AI, Explanatory Interactive Learning},
    booktitle= {Proceedings of the 1st Conference of Hybrid Human Artificial Intelligence (HHAI) and in Frontiers in Artificial Intelligence and Applications}
}