Prediction of Essay Cohesion in Portuguese Based on Item Response Theory in Machine Learning

The essay is considered a useful mechanism for evaluating learning outcomes in writing. Essay correction is a manual task that presents difficulties related to time, cost, reliability, and the subjectivity of the examiner. Cohesion is a fundamental aspect of the text, as it helps to establish a meaningful relationship between its different parts. The automated scoring of cohesion in essays presents a challenge in the field of artificial intelligence in education. This is primarily due to the fact that machine learning algorithms, commonly employed for text evaluation, often overlook the unique characteristics of individual instances within the analyzed corpus. To address this issue, item response theory can be adapted to the machine learning context. This adaptation involves characterizing aspects such as ability, difficulty, discrimination, and guessing in the utilized models. This research aims to analyze the performance of cohesion score prediction in Brazilian basic education essays, using item response theory to estimate the scores generated by machine learning models. The research extracted 325 linguistic features and treated it as a regression problem. Initial results indicate that the proposed approach has the potential to outperform conventional models. The research presents a promising avenue for a more precise evaluation of cohesion in educational essays.