Code of Directional Self-Attention Network (DiSAN)
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
Fast-DiSA Update May 8, 2018
ReSAN add new model May 3, 2018
SNLI_disan add new model May 6, 2018
SST_disan add new model May 6, 2018
.gitignore update May 6, 2018
LICENSE Initial commit Nov 10, 2017 Update Apr 28, 2018 upload all codes Nov 10, 2017

Directional Self-Attention Network

Cite this paper using BibTex:

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}

Overall Requirements

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

Python Packages:

  • numpy

This repo includes three part as follows:

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

The Usage of 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.

Usage of


  • 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.