HECTOR: A Hybrid TExt SimplifiCation TOol for Raw Texts in French

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4620–4630 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4620 HECTOR: A Hybrid TExt SimplifiCation TOol for Raw Texts in French Amalia Todirascu1, Rodrigo Wilkens2, Eva Rolin2, Thomas Franc¸ois2, Delphine Bernhard1, N´uria Gala3 1 Universit´e de Strasbourg, LiLPa UR 1339, F-67000 Strasbourg, France 2 CENTAL, Universit´e catholique de Louvain, Belgium 3 Aix-Marseille Universit´e, Laboratoire Parole et Langage, LPL CNRS (UMR 7309), France {todiras, dbernhard}@unistra.fr, {rodrigo.wilkens,eva.rolin,thomas.francois}@uclouvain.be, nuria.gala@univ-amu.fr Abstract Reducing the complexity of texts by applying an Automatic Text Simplification (ATS) system has been sparking interest in the area of Natural Language Processing (NLP) for several years and a number of methods and evaluation campaigns have emerged targeting lexical and syntactic transformations. In recent years, several studies exploit deep learning techniques based on very large comparable corpora. Yet the lack of large amounts of corpora (original-simplified) for French has been hindering the development of an ATS tool for this language. In this paper, we present our system, which is based on a combination of methods relying on word embeddings for lexical simplification and rule-based strategies for syntax and discourse adaptations. We present an evaluation of the lexical, syntactic and discourse-level simplifications according to automatic and human evaluations. We discuss the performances of our system at the lexical, syntactic, and discourse levels. Keywords: Automatic Text Simplification, hybrid architecture, French corpora, evaluation metrics. 1. Introduction Automatic Text Simplification (ATS) is a challeng- ing Natural Language Processing (NLP) task aiming at making content more readable and understandable for specific target readers (e.g. people with learning or cognitive disorders, language learners and people with low literacy). Several systems have been proposed for English (Carroll et al., 1998) or Ibero-Romance lan- guages (Ferr´es et al., 2017) focusing on lexical and syntactic transformations. However, fewer systems yet propose discourse-level simplifications (Siddharthan, 2014; Siddharthan, 2006) and no specific end-to-end multilevel simplification tool is available for French, apart from MUSS (Martin et al., 2020b), a multilin- gual language model trained using sentence-level para- phrases for several languages including French. There has been much work on writing rules for domain- specific ATS over the last two decades but the field has significantly changed during the last few years due to deep learning techniques (Al-Thanyyan and Azmi, 2021). Whereas rule-based systems require time- consuming, hand-made rules based on the study of cor- pora, ML approaches – and, in particular, deep learning approaches – consider the simplification as a monolin- gual variant of a machine translation (MT) task: sim- plification operations are learned from complex-simple sentence pairs, automatically extracted from available corpora (Zhu et al., 2010). However, the lack of large available parallel corpora has been hindering the de- velopment of MT-ATS systems in various languages, including French. Faced with this challenge, NLP de- velopers have investigated two main paths for French: • to develop a MT system by automatically trans- lating existing corpora such as Newsela from En- glish to French and then applying …

Content truncated to 500 words