Abstract

Reading comprehension is a key process in academic and personal development. Different instruments are systematically applied to analyze the behavior of this indicator, such as the PISA tests, which reflect that the situation in OECD countries such as Chile is far from good. This paper presents a proposal that, through Language Models based on Deep Learning, can be used to improve reading comprehension. Using pre-trained models, together with an additional dataset, fine-tuning has been made to obtain pairs of questions and answers from documents supplied by users. These models are used for the automated preparation of questionnaires with which reading comprehension can be evaluated. A set of experiments was carried out to determine the best adjustment parameters of these models, and these were validated in books of the Chilean National Reading Curriculum, specifically in Readings present in basic education. In all cases, good results were obtained.