Train convolutional network for sentiment analysis. Based on "Convolutional Neural Networks for Sentence Classification" by Yoon Kim, link. Inspired by Denny Britz article "Implementing a CNN for Text Classification in TensorFlow", link.

We experimented this implementation for comparative study of sentiment analysis on Keras for Teano, Tesorflow, CNTK.

CNN-non-static loss and accuracy:

• In theano val_loss and val_acc in 1st epoch is 0.6931 and 0.5126 respectively. While in 10 epoch it is 0.3801 and 0.8708.
• In Tensorflow val_loss and val_acc in 1st epoch is 0.6932 and 0.5000 respectively. While in 10 epoch it is 0.3434 and 0.8690.
• In CNTK val_loss and val_acc in 1st epoch is 0.3691 and 0.5000 respectively. While in 10 epoch it is 0.2695 and 0.8690.

• larger IMDB corpus, longer sentences; sentence length is very important, just like data size
• smaller embedding dimension, 20 instead of 300
• 2 filter sizes instead of original 3
• much fewer filters; experiments show that 3-10 is enough; original work uses 100
• random initialization is no worse than word2vec init on IMDB corpus
• sliding Max Pooling instead of original Global Pooling

• The Keras Deep Learning library

• The Theano backend should be installed.

• The Tensorflow backend should be installed.

• The CNTK backend should be installed.