Iterative Edit-based Unsupervised Sentence Simplification

Loading...
Thumbnail Image

Date

2020-07-28

Authors

Kumar, Dhruv

Advisor

Golab, Lukasz

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

We present a new iterative approach towards unsupervised edit-based sentence simplification. Our approach is guided by a scoring function to select simplified sentences generated after iteratively performing word and phrase-level edits on the complex sentence. The scoring function measures different aspects of simplification: fluency, simplicity, and preservation of meaning. As a result, unlike past approaches, our method is controllable and interpretable and does not require a parallel training set since it is unsupervised. At the same time, using the Newsela and WikiLarge datasets, we experimentally show that our solution is nearly as effective as state-of-the-art supervised approaches.

Description

Keywords

Natural Language Processing, Machine Learning, Text Simplification

LC Subject Headings

Citation