CapsNet-for-Sentiment-Analysis

Sentiment analysis using CapsNet architecture

Convolutional Neural Networks have shown state of the art performance in Text Classification . But there is still a loss of information which happens due to the presence of MaxPool layer in CNNs. With Hinton's CapsNet , the loss of information is avoided . Not just that , the 'pose' or the'orientation' of an object or a representation is preserved .

This is another attempt of using CapsNet for Sentiment Analysis in a mixed language corpus (Hinglish).

For text preprocessing , 100 dimensional Glove embeddings were used.

Credits https://github.com/stefan-it , this is a modification of his CapsNet class implementation.

Improvements

Used Glove embeddings for only english words and Fastext for hindi words. Improvement in accuracy was noted.