A Data Augmentation approach to Automated Readability Assessment

• Autor(a) principal: Cunha de Menezes, Luiza
• Data de Publicação: 2023
• Outros Autores: Paes, Aline, Finatto, Maria José Bocorny
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Domínios de Lingu@gem
• Texto Completo: https://seer.ufu.br/index.php/dominiosdelinguagem/article/view/68318
• topic: Processamento de Linguagem Natural; Substituição por Sinônimo; Retrotradução; Aumento de Dados; Avaliação Automática de Leiturabilidade; Natural Language Processing; Synonym Replacement; Back-translation; Data Augmentation; Automatic Readability Assessment

Studies about how to measure text readability reassemble the last century. Nonetheless, there is no consensus on which could be the best metrics. Tools regarding the field of Natural Language Processing (NLP) may support this task but are dependent on a high number of samples for training, and that is a bottleneck to its advancement. The main goal of this paper is to analyze the impact of a couple of data augmentation (DA) methods to support the readability classification task in Brazilian Portuguese (BP) to mitigate the bottleneck problem. In this sense, we worked on a paired and classified corpus created by linguists. The corpus is about science, and each text contemplates its original and simplified versions. About the methodology, we considered two agnostic tasks: synonym replacement and back-translation and evaluated 75 models with different techniques and combinations of input features. For the trained model with the corpus without DA, the best score reached 94.0% of the hit rate. When combining the NILC-Metrix metrics and contextualized word embeddings, the results overtook 95.2%. Compared to other papers applied to the BP, the proposed methodology improved the hit rate considering a distinct training domain. Our results demonstrate that the capacity of DA methods can be equal to or greater than those trained without augmentation and, at the same time, present greater generalization when applied to other domains.