NN_sentiment

An implementation of CNN/RNN for sentiment classification. This repo has three models: two on sentence level, one on document level.

##Requirment

• python 2.7
• Theano
• keras
• keras-extra

Model

1) sentence-level CNN

The same model as Yoon's Convolutional Neural Networks for Sentence Classification A embedding layer followed by convolution layer.

2) sentence-level LSTM

Similar to model 1, concatenating a embedding layer with a LSTM-RNN module.

3) document-level CNN-LSTM

Implement sentiment classification on document level. The basic idea is to stack CNN and a LSTM. The first layer is a embedding layer initialized by word2vec, transform each word to word embedding representaion. Then a convolution layer to learn a fixed-length representation for each sentence. Then input the sentence-level representaion to a RNN module(GRU/LSTM) for sentiment classification.

Data

• IMDB movie data for sentence-level
• Yelp review data for document-level

Result

• See Yoon's paper for the performance of Model 1
• Model 3 reaches 70.2% accuracy in 5-class classification on yelp-2015 dataset.
• TODO: also worth trying regression as the cost function on yelp data.