Document-Level Planning for Text Simplification

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 993–1006 May 2-6, 2023 ©2023 Association for Computational Linguistics Document-Level Planning for Text Simplification Liam Cripwell Université de Lorraine CNRS/LORIA liam.cripwell@loria.fr Joël Legrand Université de Lorraine Centrale Supélec CNRS/LORIA joel.legrand@inria.fr Claire Gardent CNRS/LORIA Université de Lorraine claire.gardent@loria.fr Abstract Most existing work on text simplification is limited to sentence-level inputs, with at- tempts to iteratively apply these approaches to document-level simplification failing to coher- ently preserve the discourse structure of the document. We hypothesise that by providing a high-level view of the target document, a simplification plan might help to guide gener- ation. Building upon previous work on con- trolled, sentence-level simplification, we view a plan as a sequence of labels, each describing one of four sentence-level simplification oper- ations (copy, rephrase, split, or delete). We propose a planning model that labels each sen- tence in the input document while considering both its context (a window of surrounding sen- tences) and its internal structure (a token-level representation). Experiments on two simpli- fication benchmarks (Newsela-auto and Wiki- auto) show that our model outperforms strong baselines both on the planning task and when used to guide document-level simplification models. 1 Introduction Text simplification aims to transform a given text into a simpler version of itself that preserves the core meaning such that it can be better understood by a wider audience (Gooding, 2022). Simplifica- tion has also been shown to be a useful preprocess- ing step for downstream NLP tasks such as relation extraction (Miwa et al., 2010; Niklaus et al., 2016) and machine translation (Chandrasekar et al., 1996; Mishra et al., 2014; Li and Nenkova, 2015; Štajner and Popovic, 2016). Previous research has mostly considered the sim- plification of isolated sentences. Much work has focused on training a statistical or a neural model on pairs of complex and simplified sentences as- suming that such models will learn to perform sim- plification operations (e.g. sentence splitting, lexi- cal simplification or syntactic rephrasing) implic- itly from the inductive bias present in the train- ing data (Zhang and Lapata, 2017; Nisioi et al., 2017; Jiang et al., 2020). However, because the training data is obtained using distant supervision techniques and is often imbalanced in terms of sim- plification operations (many of which occur infre- quently (Jiang et al., 2020)), system outputs have been found to be overly conservative, often making no changes or being limited to the paraphrasing of short word sequences (Alva-Manchego et al., 2017; Maddela et al., 2021). In addition, these systems provide limited capacity for controllability and are unable to express alternative variants of the sim- plified text (Alva-Manchego et al., 2017; Cripwell et al., 2021). In response, controllable simplification systems have been proposed which either constrain at- tributes of the output (length, amount of paraphras- ing, lexical and syntactic complexity) (Martin et al., 2020) or explicitly specify which simplification op- eration to perform (Alva-Manchego et al., 2017; Dong et al., 2019; Malmi et al., 2019; …

Content truncated to 500 words