Inserting rhetorical predicates for quasi-abstractive summarization

Abstract

We investigate the problem of inserting rhetorical predicates (e.g. "to present", "to discuss", "to indicate", "to show") during non extractive summary generation and compare various algorithms for the task which we trained over a set of human written summaries. The algorithms which use a set of features previously introduced in the summarization literature achieve between 57% to 62% accuracy depending on the machine learning algorithm used. We draw conclusions with respect to the use of context during predicate prediction.