As the growth of online data continues, automatic summarization is integral in generating a condensed version of a text while preserving the meaning of the original input. Although most of the earlier works on automatic summarization use extractive approaches to identify the most important information of a document, recent research works focus on the more challenging task of making the summaries abstractive. Sequence-to-sequence models with attention have quantitatively shown to be effective for abstractive summarization, but the quality of the generated summaries is often poor with incorrect and redundant information. In this thesis, we present an end-to-end neural network framework which combines a hierarchical content selector and pointer generator networks abstractor through a multi-level attention mechanism that uses the sentence importance scores from the former model to help the word-level attention of the latter model make better decisions when generating the output words. Hence, words from key sentences will be attended more than words in less salient sentences of input. Our approach is motivated by human writers who tend to focus only on the relevant portions of an article when summarizing while ignoring anything irrelevant that might degrade the output quality. We conduct experiments on the challenging CNN/Daily Mail dataset, which consists of long newswire articles paired with multiple-sentence summaries. Experimental results show that our end-to-end architecture outperforms the extractive systems and strong lead-3 baseline and achieves competitive ROUGE and METEOR scores with previous abstractive systems on the same dataset. Qualitative analysis of test data shows that the generated summaries are fluent as well as informative.
Copyright is held by the author.
This thesis may be printed or downloaded for non-commercial research and scholarly purposes.
Supervisor or Senior Supervisor
Thesis advisor: Popowich, Fred
Member of collection