Towards Improving the Robustness of Neural Abstractive Summarization

Recent deep learning and sequence-to-sequence learning technology have produced impressive results on automatic summarization. However, the models have limited insights on the underlying language and it remains challenging for system-generated summaries to be truthful to the original input or cover the most important information. This is especially the case for generating abstractive summaries using neural models. My work aims for a flexible and controllable summarization system that can be adapted to cater to different scenarios. It is designed to incorporate linguistic structure information into deep neural networks, have the capability to produce abstracts by re-using a varying amount of source text, and take language characteristics into consideration for summary generation and selection. My dissertation provides a comprehensive overview to the problem of text summarization. I will present a number of approaches to incorporate linguistic structure into state-of-the-art deep neural models to help system summaries remain grammatical and retain the most salient meaning of the source text. I will also describe a summarization approach that is controllable during training and produce diverse summaries during the decoding and re-ranking processes. Finally, I will conclude with a novel approach for selecting optimal summaries from a collection of candidates and discuss the opportunities and challenges in this promising area of research