TransformersConceptNet

Code and Data Associated with Can LLMs facilitate interpretation of pre-trained language models?

Data

The data can be downloaded from the project page

Data Description

We use agglomerative hierarchical clustering over contextualized representations and then annotate these concepts using GPT annotations. Our study involved several 12-layered transformer models, including BERT-cased, RoBERTa, XLNet, and ALBERT. For more details, please refer to the paper

Load data into display

Clone this repository:

git clone https://github.com/qcri/TransformersConceptNet.git
cd TransformersConceptNet

Create and activate virtual environment:

python -m venv .envs/tcn
source .envs/tcn/bin/activate

Install requirements:

pip install -r requirements.txt

Start the webapp using the following command:

python -u app.py -d <path-to-downloaded-data>

and visit http://localhost:8080 in your browser. The port and hostname can be passed as additional arguments to app.py.

Concept probing and Neuron Analysis

We used the NeuroX package to train linear probes and perform neuron analysis. You can view the documentation of the package here

Todo

• Add cross architectural comparison between the models in the display

Citation

If you used the dataset please cite

@inproceedings{mousi2023llms,
      title = "Can LLMs Facilitate Interpretation of Pre-trained Language 
      Models?",
      author = "Mousi, Basel  and
      Durrani, Nadir  and 
      Dalvi, Fahim", 
      booktitle = "Proceedings of the 2023 Conference on Empirical Methods 
      in Natural Language Processing",
      month = dec, 
      year = "2023", 
      publisher = "Association for Computational Linguistics", 
      url = "https://browse.arxiv.org/pdf/2305.13386.pdf"
}