This repository contains source code for the experiments from the paper presented at CIKM 2019.
Find our interactive demo that visualizes the results on three Question Answering datasets here: https://visbert.demo.datexis.com
For probing the language abilities in BERT's layers, we used the Jiant Probing Suite by Wang et al. We added two additional tasks to their suite: Question Type Classification and Supporting Fact Extraction. The code for creating these tasks can be found in the probing directory.
To train and evaluate BERT QA models we used the 🤗 Transformers framework by Huggingface. A simple way to visualize how tokens are transformed by a QA transformer model can be found in the visualization directory. We use a single question as input and output the token representations for each layer of the model within a 2D vector space.
When building up on our work, please cite our paper as follows:
@article{van_Aken_2019,
title={How Does BERT Answer Questions?},
journal={Proceedings of the 28th ACM International Conference on Information and Knowledge Management - CIKM ’19},
publisher={ACM Press},
author={van Aken, Betty and Winter, Benjamin and Löser, Alexander and Gers, Felix A.},
year={2019}
}