Avaliação do conhecimento do estudante em sistemas tutores inteligentes baseados em passos: uma abordagem baseada em Deep Knowledge Tracing

Abstract

Intelligent Tutoring Systems (ITS) are intelligent learning environments that, as they have a model of the content to be taught, as well as student information, are able to offer individualized teaching and assistance. ITS are almost as efficient as one-to-one tutoring, in which the student receives individualized assistance from a teacher. This greater efficiency is shown to be even more effective when students use step-based ITS, in which they receive feedback from the system at each step taken to solve a problem and not only after the end of the exercise. To provide individualized instruction and assistance, ITS use artificial intelligence (AI) techniques to identify and model the difficulties and knowledge of each student, returning quality content, adequate for the needs of the student. The component of the system responsible for assessing the student’s knowledge and skills is the student model; this component is responsible both for identifying the degree of proficiency of students in each knowledge unit, and for assessing the chance of success in the next task or step, in order to provide assistance. Bayesian Networks, Bayesian Knowledge Tracing (BKT) and Deep Knowledge Tracing (DKT) are the most applied current tools for the development of the ITS student model, and the DKT technique is the one that currently presents the best performance for student models that seek to predict success in the next task. However, this technique was only used in answer-based ITS, that is, ITS in which only the student’s final response is analyzed. The aim of the proposed work is to use DKT to predict student success in the next step in a step-based ITS. Step-based ITS are those in which each intermediate step of the solving process presented by the student is analyzed, providing feedback and help with each new interaction. When using data from a step-based ITS, the system receives a greater amount of information, since to complete an exercise, in general, the student provides more than one step, making the step-based student model more complex and efficient. For this study, data extracted from the PAT2Math tutor system was used, which is a step-based ITS that assists students in solving first degree equations. The results demonstrate that the Neural Networks using the multi-layer perceptron architecture achieved an overall result very close to those obtained by the related works. However, when using step-based ITS data, this result is even better. While the related work reached AUC of 79.24 % with LSTM networks for answer-based ITS, the proposed work reached AUC of 50% and of F1-Score of 0.8506 using the same LSTM network with step-based ITS , and AUC of 76.78 % and F1-Score of 0.8657, when applied to MLP with the same data, allowing to identify that such student model has potential for use in step-based ITS.