julianfuchs/twittersentimentcharcnn

Character based CNN for Twitter Sentiment Classification

TwitterSentimentCharCNN

Character based CNN for Twitter Sentiment Classification

Takes the Twitter data from the Computational Intelligence Lab at ETH from the spring semester 2016 (Project in Exercise 6) and performs a Text Sentiment Classification using a Character based Convolutional Network (CNN).

The CNN used is 9 layers deep with 6 convolutional layers and 3 fully connected layers, heavily inspired by the CNN introduced in the paper 'Character-level Convolutional Networks for Text Classification' by Xiang Zhang, Junbo Zhao, Yann LeCun. https://arxiv.org/abs/1509.01626

As a base code, the code written by https://github.com/scharmchi was used found under https://github.com/scharmchi/char-level-cnn-tf. The code was modified however:

• Padding was introduced (a padding of 'SAME' instead of 'VALID' in tensorflow)
• Since the maximal twitter length is 140 and the original paper has a sequence length of 1014 in mind, we only use 2 pooling steps instead of 3 as in the original paper, since otherwise the vectors would get too small
• You can manually set the number of filters for each convolutional layer