| Sumario: | This paper presents a study on text complexity of Open Educational Resources (OER) in Brazilian Portuguese. In a data analysis of the Brazilian Ministry of Education Integrated Platform (MEC-RED) carried out in September 2020, 86% of the resources on the platform did not have any grade level classification, making it difficult to find, use, and expand them. The text complexity task in the Natural Language Processing research area can be used to identify texts that have adequate linguistic complexity for specific grade levels, allowing to complete the stage of education metadata in MEC-RED. However, some types of MEC-RED's resources do not present any information about their stage of education, making it unfeasible to compile a balanced dataset of OER for training a text complexity predictor. This study is driven and enabled by a recently created corpus of transcribed spoken narratives produced by fourth graders to first graders of high school which were collected to evaluate the development of language abilities. A multi-task learning (MTL) approach via hard parameter sharing of hidden layers was adopted to train three models that share all parameters in their hidden layers. The main objective of this study was to explore the relationship between three text complexity tasks by jointly learning to predict text readability, using coarse and fine-grained datasets of written, spoken and domain texts (a small dataset of OER resources) to overcome the lack of grade classified resources in MEC-RED. Our MTL model with two auxiliary tasks presents a F-measure of 0.955, an improvement of 0.15 points over our previous results.
|
|---|