Representing sentence with unfolding recursive autoencoders and dynamic average pooling

This paper proposes a new composition method to represent semantic compositionality of sentences. Using the unfolding recursive autoencoders, we build sentence representing trees from the original sentences of words. We utilize trained word embeddings and sentence parser to train the model, and we can build sentence representing trees from the trained model. We further propose to use dynamic average pooling to pool the trees and get fix-size vector representation for sentences. The fix-size vector representation after dynamic average pooling can then be used to represent sentences. We verify the validity of sentence representations by using them to classify sentence paraphrase. Experiment shows that compared to the baseline representation, using proposed representations together with sub-vector Euclidean distance feature, the classification performance can be improved by 1.39% for testing accuracy, which proves the proposed method can better represent semantic compositionality of sentences.