ALEXSIS-PT: A New Resource for Portuguese Lexical Simplification

Proceedings of the 29th International Conference on Computational Linguistics, pages 6057–6062 October 12–17, 2022. 6057 ALEXSIS-PT: A New Resource for Portuguese Lexical Simplification Kai North1, Marcos Zampieri1, Tharindu Ranasinghe2 1George Mason University, USA 2University of Wolverhampton, UK knorth8@gmu.edu Abstract Lexical simplification (LS) is the task of automatically replacing complex words for easier ones making texts more accessible to various target populations (e.g. individuals with low literacy, individuals with learning disabilities, second language learners). To train and test models, LS systems usually require corpora that feature complex words in context along with their candidate substitutions. To continue improving the performance of LS systems we introduce ALEXSIS- PT, a novel multi-candidate dataset for Brazilian Portuguese LS containing 9,605 candidate substitutions for 387 complex words. ALEXSIS-PT has been compiled following the ALEXSIS protocol for Spanish opening exciting new avenues for cross- lingual models. ALEXSIS-PT is the first LS multi-candidate dataset that contains Brazilian newspaper articles. We evaluated four models for substitute generation on this dataset, namely mDistilBERT, mBERT, XLM-R, and BERTimbau. BERTimbau achieved the highest performance across all evaluation metrics. 1 Introduction The development of lexical simplification (LS) systems provides a cost-effective means of making texts accessible to individuals with reading disabilities or low-literacy who are at an economic and social disadvantage (Nogueira et al., 2022). LS systems aim to replace difficult to understand (complex) words or phrases with simpler alternatives (North et al., 2022). Consider the examples in Table 2. An LS system would firstly identify the word shown in bold as being complex (Paetzold and Specia, 2016). It would then generate top-k number of candidate substitutions that preserve the meaning of the original complex word in the provided context, yet are easier to understand or are more familiar to the user. These top-k candidate substitutions are then filtered based on their appropriateness, referred to as substitute generation, and finally ranked in accordance to their suitability, known as substitute ranking. Various studies have been published on Brazilian Portuguese (pt-BR) LS (Aluísio and Gasperin, 2010; Leal et al., 2018; de Lima et al., 2021; Leal et al., 2022). Brazil is a country with important literacy challenges where only 1% of its population working in the agricultural sector are proficient readers (Leal et al., 2018). These educational challenges motivate the development of technology to assist readers and the creation of resources for pt-BR. One of the most popular pt-BR LS datasets is SIMPLEX-PB 2.0 (Hartmann et al., 2020b,a) featuring substitutions for over 1,500 complex words. However, in SIMPLEX-PB 2.0 only 5 ranked candidate substitutions are available for 730 of its complex words making the dataset less useful as a benchmark for state-of-the-art large pre-trained language models. Moreover, it contains texts from children’s books which makes it domain-specific and therefore not a great fit for general LS. To address these gaps we introduce ALEXSIS- PT, a pt-BR LS dataset featuring excerpts of newspaper articles containing a larger number of candidate substitutions per target word (up to 25). To the best of our knowledge, ALEXSIS-PT has …

Content truncated to 500 words