Need help with dan?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

229 Stars 56 Forks MIT License 16 Commits 3 Opened issues


Deep Averaging Networks

Services available


Need anything else?

Contributors list

Deep Averaging Networks (DAN)

code for model described in along with negation dataset (negation_dataset.txt). feel free to email me at [email protected] with any comments/problems/questions/suggestions.


  • python 2.7.9, numpy 1.9.2 (might work w/ other versions but not tested), nltk

commands to run DAN on Stanford Sentiment Treebank:

  • bash (downloads word embeddings and dataset, preprocesses PTB trees into DAN format)
  • python (can tweak hyperparameters via command-line arguments, currently this runs the fine-grained experiment on only root-level labels and should take a few minutes to finish training)

QA DAN code available in the repository for our full quiz bowl system


DAN input format (for your own data!):

  • each training/test instance must be a tuple with the following format: ([list of word embedding lookup indices associated with text], label)
  • if you want to use pretrained word embeddings, you should also pass a pickled matrix using the --We argument, where the matrix is of size d x V (each column stores the embedding for the corresponding word lookup index)

important hyperparameters:

  • batch size (the smaller the better, but also slower)
  • adagrad initial learning rate (should be decreased as the batch size is decreased)
  • word dropout probability (30% is the default but might be too high for some tasks)
  • number of epochs (increase when using random initialization)

if you use this code, please cite:

    Title = {Deep Unordered Composition Rivals Syntactic Methods for Text Classification},
    Booktitle = {Association for Computational Linguistics},
    Author = {Mohit Iyyer and Varun Manjunatha and Jordan Boyd-Graber and Hal {Daum\'{e} III}},
    Year = {2015},
    Location = {Beijing, China}

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.