Sequence-to-Sequence Models for Automated Text Simplification

Abstract

A key writing skill is the capability to clearly convey desired meaning using available linguistic knowledge. Consequently, writers must select from a large array of idioms, vocabulary terms that are semantically equivalent, and discourse features that simultaneously reflect content and allow readers to grasp meaning. In many cases, a simplified version of a text is needed to ensure comprehension on the part of a targeted audience (e.g., second language learners). To address this need, we propose an automated method to simplify texts based on paraphrasing. Specifically, we explore the potential for a deep learning model, previously used for machine translation, to learn a simplified version of the English language within the context of short phrases. The best model, based on an Universal Transformer architecture, achieved a BLEU score of 66.01. We also evaluated this model’s capability to perform similar transformation to texts that were simplified by human experts at different levels.