PyTorch code for paper: On Attribution of Recurrent Neural Network Predictions via Additive Decomposition. It has been accepted in WWW2019.

We propose an attribution method, called REAT, to provide interpretations to RNN predictions in a faithful manner. REAT decomposes the final prediction of a RNN into additive contribution of each word in the input text. This additive decomposition enables REAT to flexibly generate phrase-level attribution scores. In addition, REAT is generally applicable to various RNN architectures, including GRU, LSTM and their bidirectional versions.

• Clone the code from Github:

git clone https://github.com/mndu/REAT.git
cd REAT

• Download the pretrained word2vec pre-trained using Google News corpus from GoogleNews-vectors-negative300.bin.gz. Unzip the .gz file:

gunzip GoogleNews-vectors-negative300.bin.gz

• Train RNN models, we consider the following three models, GRU, LSTM and Bidirectional GRU. As the focus of this paper is to provide post-hoc attribution for the predictions of RNNs, rather than increase their predictive accuracy, thus we use standard practices to train our models.

python train_batch.py --m gru
python train_batch.py --m lstm
python train_batch.py --m bigru

• Provide attribution for an input text for three kinds of RNN models, GRU, LSTM and Bidirectional GRU:

python reat_gru
python reat_lstm
python reat_bigru

Python 2.7, torch 0.3, torchtext 0.2.3, nltk, and tqdm

@inproceedings{du2019www,
    author    = {Mengnan Du, Ninghao Liu, Fan Yang, Shuiwang Ji,  Xia Hu},
    title     = {On Attribution of Recurrent Neural Network Predictions via Additive Decomposition},
    booktitle = {The Web Conference (WWW)},
    year      = {2019}
}