• This repo is the codes of DiSAN: Directional Self-Attention Network for RNN/CNN-free Language Understanding.
• This is python based codes implementation under tensorflow 1.2 DL framework.
• The leaderboard of Stanford Natural Language Inference is available here.
• Please contact Tao Shen or open an issue for questions/suggestions.

Cite this paper using BibTex:

@inproceedings{shen2018disan,
Author = {Shen, Tao and Zhou, Tianyi and Long, Guodong and Jiang, Jing and Pan, Shirui and Zhang, Chengqi},
Booktitle = {AAAI Conference on Artificial Intelligence},
Title = {DISAN: Directional self-attention network for rnn/cnn-free language understanding},
Year = {2018}
}

• Python3 (verified on 3.5.2, or Anaconda3 4.2.0)
• tensorflow>=1.2

• numpy

1. Directionnal Self-Attention Network independent file -> file disan.py
2. DiSAN implementation for Stanford Natural Language Inference -> dir SNLI_disan
3. DiSAN implementation for Stanford Sentiment Classification -> dir SST_disan

The Usage of disan.py will be introduced below, and as for the implementation of SNLI and SST, please enter corresponding folder for further introduction.

And, Code for the other experiments (e.g. SICK, MPQA, CR etc.) appeared in the paper is under preparation.

• param rep_tensor: 3D tensorflow dense float tensor [batch_size, max_len, dim]
• param rep_mask: 2D tensorflow bool tensor as mask for rep_tensor, [batch_size, max_len]
• param scope: tensorflow variable scope
• param keep_prob: float, dropout keep probability
• param is_train: tensorflow bool scalar
• param wd: if wd>0, add related tensor to tf collectoion "reg_vars" for further l2 decay
• param activation: disan activation function [elu|relu|selu]
• param tensor_dict: a dict to record disan internal attention result (insignificance)
• param name: record name suffix (insignificance)

2D tensorflow dense float tensor, which shape is [batch_size, dim] as the encoding result for each sentence.

• Some basic neural networks are copied from Minjoon's Repo, including RNN cell, dropout-able dynamic RNN etc.