/rcnn

Recurrent & convolutional neural network modules

Primary LanguagePythonApache License 2.0Apache-2.0

Latest updates

Check code/rationale for the project of learning rationales for neural prediction:

       

Adam Yala has implemented a Pytorch version of the rationale project!

Check it out at: https://github.com/yala/text_nn


About this repo

This repo contains Theano implementations of popular neural network components and optimization methods. Source code of the following papers are also available:

[1] Molding CNNs for text: non-linear, non-consecutive convolutions. EMNLP 2015

[2] Semi-supervised Question Retrieval with Gated Convolutions. NAACL 2016

[3] Rationalizing Neural Predictions. EMNLP 2016

Features

  • Basic modules including feedforward layer, dropout, word embedding, RNN, LSTM, GRU and CNN
  • Optimization methods including SGD, AdaGrad, AdaDelta and Adam
  • Advanced modules from recent papers such as attention and gated convolution.
  • Transparent to use GPU

Projects

1. Neural question retrieval for community-based QA

The directories code/qa and code/pt contain the implementation of the model described in paper [2]. Datasets and and pre-trained word vectors are available at here.

2. Sentiment analysis / document classification

The directory code/sentiment contains the implementation of the model described in paper [1]. Datasets and and pre-trained word vectors are available at here.

3. Rationalizing neural predictions

The directory code/rationale contains the implementation of the model described in paper [3].


Dependencies

Theano >= 0.7, Python >= 2.7, Numpy