Controllable Sentence Simplification

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4689–4698 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4689 Controllable Sentence Simplification Louis Martin1,2,3 ´Eric Villemonte de la Clergerie2 Benoˆıt Sagot2 Antoine Bordes1 Facebook AI Research1, Inria2, Sorbonne Universit´e3. 6 Rue M´enars, 75002 Paris1; 2 rue Simone Iff, 75012 Paris2. louismartin@fb.com, {benoit.sagot, eric.de la clergerie}@inria.fr, antoine.bordes@fb.com Abstract Text simplification aims at making a text easier to read and understand by simplifying grammar and structure while keeping the underlying information identical. It is often considered an all-purpose generic task where the same simplification is suitable for all; however multiple audiences can benefit from simplified text in different ways. We adapt a discrete parametrization mechanism that provides explicit control on simplification systems based on Sequence-to-Sequence models. As a result, users can condition the simplifications returned by a model on attributes such as length, amount of paraphrasing, lexical complexity and syntactic complexity. We also show that carefully chosen values of these attributes allow out-of-the-box Sequence-to-Sequence models to outperform their standard counterparts on simplification benchmarks. Our model, which we call ACCESS (as shorthand for AudienCe-CEntric Sentence Simplification), establishes the state of the art at 41.87 SARI on the WikiLarge test set, a +1.42 improvement over the best previously reported score. Keywords: Text Simplification, Sequence-to-Sequence models, ACCESS 1. Introduction In Natural Language Processing, the Text Simplification task aims at making a text easier to read and understand. Text simplification can be beneficial for people with cog- nitive disabilities such as aphasia (Carroll et al., 1998), dyslexia (Rello et al., 2013) and autism (Evans et al., 2014) but also for second language learners (Xia et al., 2016) and people with low literacy (Watanabe et al., 2009). The type of simplification needed for each of these audiences is dif- ferent. Some aphasic patients struggle to read sentences with a high cognitive load such as long sentences with intri- cate syntactic structures, whereas second language learners might not understand texts with rare or specific vocabulary. Yet, research in text simplification has been mostly focused on developing models that generate a single generic simpli- fication for a given source text with no possibility to adapt outputs for the needs of various target populations. In this paper, we propose a controllable simplification model that provides explicit ways for users to manipulate and update simplified outputs as they see fit. This work only considers the task of Sentence Simplification (SS) where the input of the model is a single source sentence and the output can be composed of one sentence or split into multiple. Our work builds upon previous work on controllable text gen- eration (Kikuchi et al., 2016; Fan et al., 2017; Scarton and Specia, 2018; Nishihara et al., 2019) where a Sequence-to- Sequence (Seq2Seq) model is modified to control attributes of the output text. We tailor this mechanism to the task of SS by considering relevant attributes of the output sen- tence such as the output length, the amount of …

Content truncated to 500 words