Encoding Document Semantic into Binary Codes Space

Abstract

We develop a deep neural network model to encode document semantic into compact binary codes with the elegant property that semantically similar documents have similar embedding codes. The deep learning model is constructed with three stacked auto-encoders. The input of the lowest auto-encoder is the representation of word-count vector of a document, while the learned hidden features of the deepest auto-encoder are thresholded to be binary codes to represent the document semantic. Retrieving similar document is very efficient by simply returning the documents whose codes have small Hamming distances to that of the query document. We illustrate the effectiveness of our model on two public real datasets – 20NewsGroup and Wikipedia, and the experiments demonstrate that the compact binary codes sufficiently embed the semantic of documents and bring improvement in retrieval accuracy.