How to implement text autoencoder for outlier detection?

Question

I have a imbalanced text dataset, with distribution of the classes like this:

class 1 - 11k samples class 2 - 7k samples class 3 - 4k samples class 4 - 600 class 5 - 400