Learning Meanings for Sentences with Recursive Autoencoders

The objective of this project is to implement the recursive auto encoder (RAE) method to learn a model to predict sentiments for sentences and reproduce the result in [1]. To learn the weights for recursive functions, we implement forward and backward propagation algorithms. We validate the gradient computed from forward and backward algorithm by comparing it to the gradient computed from numerical approximation. Our implementation could achieve accuracy 48.5% with the basic RAE implementation. We show that the system achieves accuracy 76.0% by using normalized transfer function and adjustable meaning vectors.