Reinforcement Learning using Sentence-Level Difficulty for Text Simplification

Abstract

Text simplification is a natural language processing task that transforms complex text into simple text while preserving its meaning. People with language disabilities, children, and language learners can benefit from text simplification. Since audiences have different abilities and the difficulty needs to be adjusted according to their ability, some studies have proposed text simplification methods with controllable difficulty. However, previous methods are based on word-level difficulty and do not consider sentence-level difficulty. Accordingly, our approach uses reinforcement learning using a sentence-level difficulty estimator. Experiments results showed that the proposed method improves SARI, an automatic evaluation metric for text simplification.