Phrase-Level Simplification for Non-native Speakers

Abstract

AbstractTypical Lexical Simplification systems replace single words with simpler alternatives. We introduce the task of Phrase-Level Simplification, a variant of Lexical Simplification where sequences of words are replaced as a whole, allowing for the substitution of compositional expressions. We tackle this task with a novel pipeline approach by generating candidate replacements with lexicon-retrofitted POS-aware phrase embedding models, selecting them through an unsupervised comparison-based method, then ranking them with rankers trained with features that capture phrase simplicity more effectively than other popularly used feature sets. We train and evaluate this approach using BenchPS, a new dataset we created for the task that focuses on annotations on the needs of non-native English speakers. Our methods and resources result in a state-of-the-art phrase simplifier that correctly simplifies complex phrases 61% of the time.KeywordsPhrase simplificationLexical simplificationText simplification