Unsupervised Joint Learning for Headline Generation and Discourse Structure of Reviews

Abstract

Recently, using a large amount of reference summary, supervised neural summarization models have achieved success. However, such datasets are rare, and trained models cannot be shared across domains. Although an unsupervised approach is a possible solution, models applicable for single-document summary or headline generation have not been established. Our work focuses on generating headlines for reviews, without supervision. We assume that reviews contain a discourse tree in which the headline is the root and the child sentences elaborate on the parent. By estimating the parent from their child recursively, our model learns such a structure and generates a headline that describes the entire review. Through the evaluation of the generated headline on actual reviews, our model achieved competitive performance with supervised models, especially on relatively long reviews. In induced structures, we confirmed that the child-sentences explain the parent in detail and generated headline abstracts for the entire review.