Layerwise Relevance Visualization in Convolutional Text Graph Classifiers

Code accompanying paper

@inproceedings{Schwarzenberg_lrv_2019,
  title = {Layerwise Relevance Visualization in Convolutional Text Graph Classifiers},
  booktitle = {Proceedings of the EMNLP 2019 Workshop on Graph-Based Natural Language Processing},
  author = {Schwarzenberg, Robert and Hübner, Marc and Harbecke, David and Alt, Christoph and Hennig, Leonhard},
  location = {Hong Kong, China},
  year = {2019}
}

Approach

Apply LRP to GCNs. Visualize relevance in each layer.

Installation

0. Create environment w/ python 3.6, e.g.

conda create --name research-xgcn python=3.6
source activate research-xgcn

1. Install requirements.txt, e.g.

pip install -r requirements.txt 

2. Install English scispacy language model

pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.2.0/en_core_sci_sm-0.2.0.tar.gz

3. Download (last accessed 2019-05-16)

https://dl.fbaipublicfiles.com/fasttext/vectors-english/wiki-news-300d-1M.vec.zip

and extract

wiki-news-300d-1M.vec

into

./data/embeddings/

Data

The input data is contained in

./data/PubMed_20k_RCT/

It was downloaded on 2019-05-16 from

https://github.com/Franck-Dernoncourt/pubmed-rct

Experiments

1. Check config.json, to run the full pipeline, all values in the pipeline namespace should be true.

2. Run main.py which runs

   • preprocess.py (prepares dictionary, input data)
   • train.py (train XGCN)
   • explain.py (LRV on the trained XGCN, writes explanations to json lines)
   • postprocess.py (summarizes occlusion experiments in plots, writes latex document)

3. run pdflatex on data/explanations/explanations.tex (make sure tiks-dependency is installed)

4. View the explanations in ./data/explanations/explanations.pdf (if config was not changed)