Summarization as the process of generating a shorter version of a piece of text while preserving important context information is one of the most challenging NLP tasks. Traditionally, there are two main types of summarization: extractive and abstractive summarization. Extractive summarization identifies and ranks the important sentences of the text and sentences with highest ranks form the summary. While abstractive summarization generates novel sentences similar to what a human author does. We focus on abstractive summarization due to the flexibility of abstractive summarization systems and the increasing interest in such systems. Sequence-to-sequence models have recently gained the state of the art performance in summarization. However, there still exists challenges affecting the performance of these systems.
In this talk, we initially present challenges of the existing summarization systems both regarding the data and methods. Next, the WikiHow, a large-scale highly diverse summarization dataset is introduced to address data related issues. Later, the proposed hierarchical reinforcement learning architecture is described which makes decisions in two steps: the high-level policy decides on the sub-goal for generating the next chunk of summary and the low-level policy performs primitive actions to fulfill the specified goal. I will finally talk about the results showing our model outperforms the existing approaches and will conclude with a few possible future directives.