Can a recurrent autoencoder improve the preformace of a recurrent neural network?
I used the data from "Real or Not? NLP with Disaster Tweets" competition on Kaggle to test this hypothesis.
First I built up a deep recurrent neural network (RNN) and trained it as usual. Then I trained an autoencoder and used the encoder to build a similar RNN.
To improve performance, the GloVe embeddings trained on tweets were used.