Using Unsupervised Deep Learning for Automatic Summarization of Arabic Documents

Traditional Arabic text summarization (ATS) systems are based on bag-of-words representation, which involve a sparse and high-dimensional input data. Thus, dimensionality reduction is greatly needed to increase the power of features discrimination. In this paper, we present a new method for ATS using variational auto-encoder (VAE) model to learn a feature space from a high-dimensional input data. We explore several input representations such as term frequency (tf), tf-idf and both local and global vocabularies. All sentences are ranked according to the latent representation produced by the VAE. We investigate the impact of using VAE with two summarization approaches, graph-based and query-based approaches. Experiments on two benchmark datasets specifically designed for ATS show that the VAE using tf-idf representation of global vocabularies clearly provides a more discriminative feature space and improves the recall of other models. Experiment results confirm that the proposed method leads to better performance than most of the state-of-the-art extractive summarization approaches for both graph-based and query-based summarization approaches.