ALEXSIS: A Dataset for Lexical Simplification in Spanish

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3582–3594 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3582 ALEXSIS: A Dataset for Lexical Simplification in Spanish Daniel Ferr´es, Horacio Saggion Large Scale Text Understanding Systems Lab / TALN Department of Information & Communication Technologies Universitat Pompeu Fabra, Barcelona (Spain) {daniel.ferres,horacio.saggion}@upf.edu Abstract Lexical Simplification is the process of reducing the lexical complexity of a text by replacing difficult words with easier to read (or understand) expressions while preserving the original information and meaning. In this paper we introduce ALEXSIS, a new dataset for this task, and we use ALEXSIS to benchmark Lexical Simplification systems in Spanish. The paper describes the evaluation of three kind of approaches to Lexical Simplification: a thesaurus-based approach, a single transformers-based ap- proach, and a combination of transformers. We also report state of the art results on a previous Lexical Simplification dataset for Spanish. Keywords: Lexical Simplification, Text Simplification, Evaluation Dataset 1. Introduction Text Simplification (Saggion, 2017) is a Natural Language Processing task to transform a text with the aim of reduc- ing its lexical and syntactic complexity while preserving its original meaning. This task can be potentially useful for different audiences, specially children, second language (L2) learners (Petersen and Ostendorf, 2007), low liter- acy readers (Alu´ısio and Gasperin, 2010) and people with cognitive disabilities (Saggion et al., 2015) among others. Automatic Text Simplification has usually been concerned with two different tasks: Lexical Simplification and Syntac- tic Simplification. Lexical Simplification, the focus of the present work, aims at replacing difficult words with easier synonyms while preserving the information and meaning of the original text. Lexical Simplification systems (Shard- low, 2014; Paetzold and Specia, 2017a) usually have com- ponents for: 1) identification of complex terms (Complex Word Identification - CWI), 2) generation of substitution words (Substitute Generation - SG), 3) selection of the sub- stitutes that can fit in the context (Substitute Selection - SS), 4) ranking substitutes by their simplicity (Substitute Ranking - SR), and 5) morphological generation and con- text adaptation. The availability of several lexical simplification datasets for the English language has intensified research in this area making system comparison possible. In this paper we present a dataset for benchmarking lexical simplification in Spanish hoping this will help other researchers who are de- veloping technology in this area. Additionally, we present the first set of experiments on this dataset therefore estab- lishing a number of benchmarks for the task. The dataset, systems, and systems’ outputs will be made available to the research community for reproducible research1. The new dataset contains a set of sentences in which a com- plex work has been identified and a set of substitutes has been proposed by several annotators. Although the first phase of a LS system is CWI, the systems evaluated in this paper do not perform the CWI task and they start in the SG 1https://github.com/LaSTUS-TALN-UPF/ ALEXSIS_lexsim phase, using the provided complex words to propose and rank simpler …

Content truncated to 500 words