Query-oriented single-document summarization using unsupervised deep learning

Over the past half a century, machine-based text summarization has been addressed from many different perspectives in a variety of application domains. Deep neural networks recently show promising results for text summarization and this thesis explores this application domain. In this research study, a deep auto-encoder is used to rank sentences based on the most salient information. More precisely, a deep neural network has been used for extractive query-oriented single-document summarization. Also, the use of an Ensemble Noisy Auto-Encoder (ENAE) for this task has been evaluated. ENAE is a stochastic version of an auto-encoder that adds noise to the input text and selects the top sentences from an ensemble of runs. Our experiments show that although a deep auto-encoder can be an effective summarizer, deep auto-encoders trained with stochastic noise in the input and run multiple times with different noise in the input can make improvements. The architecture of ENAE changes the application of the auto-encoder from a deterministic feed-forward network to a stochastic model. To cover a wide range of topics and structures, we perform experiments on two different publicly available email corpora that are specifically designed for text summarization.