Introducing Simplicity in Document Summarization by Leveraging Coreference and Pseudo-Summary

Abstract

Document summarization has rapidly gained importance due to the exponentially increasing data. Generally, studies in document summarization focused on generating summaries having high coverage and fluency. Such summaries can be challenging for readers with limited language proficiency. This paper introduces the simplification concept in document summarization tasks. Our method is divided into two phases to handle the challenges of the task. The first phase is to tackle the problem of long documents with unnecessary details that can affect key information or coverage of the summaries. Thus, we propose a method to condense key information by utilizing coreference resolution. The second phase uses the condensed documents as inputs. This phase handles the challenge of having no dataset with a simplification concept in summarization tasks. Therefore, this research proposes an unsupervised training framework without relying on golden summaries. The training first outputs summaries with high coverage called pseudo-summaries. Then, it is used as a reference to generate final summaries with words that are more familiar and commonly used, resulting in easier-to-understand summaries for readers.