A neural network model for sentiment analysis on a twitter dataset.
This model receives a tweet of variable length and classifies it as either positive or negative. The structure is just a simple Bi-LSTM plus Multi-Layer Perceptron Classifier.
I ran an experiment of positive-negative classification on the Twitter Japanese Reputation Analysis Dataset distributed in the link below.
http://bigdata.naist.jp/~ysuzuki/data/twitter/
For validation set, I randomly selected 20% of the data with positive or negative label and used the rest for training. The ratio of positive and negative labels is almost 1:2.
As a result, this model achieved 82% accuracy which is much higher than that of my Bag of Words based model whose accuracy is 63%.