Abstract:

Adaptation in education is an important technique that can be used to enhance and improve learning. Different contents and links can be personalized to students depending on their preferences, needs or behaviors. Intelligent Tutoring Systems (ITSs) aim to provide personalized contents to students contributing to create an adaptive and innovative choice to learn. These contents are usually exercises. ITSs make use of different types of data, student models and mathematical and probabilistic algorithms to achieve the adaptation.

One of the main sources of information used for adaptation are the skills of the student. In order to provide these personalized contents, there is a need of student modeling. A student modelling defines a set of features for each student. The contents are personalized depending on these features. We can divide the student models into two main general groups: skill modelling and non-skill modelling. The former provides information about the knowledge level of the student in different skills, while the latter provides any other information about the student, e.g. their engagement, emotions, habits or preferences.

This paper presents a review of four main techniques used for skill modelling and adaptation of questions: knowledge spaces (KS), item response theory (IRT), bayesian networks (BN), and semantic solutions. Moreover, we provide examples of real systems that use these methods.

These skill models make different assumptions to model the reality so they have some limitations. In this paper, we make an analysis of the advantages and disadvantages of these methods for adaptation of learning contents.

The methodology used in this review has been to search and read many different papers of previous researches about adaptive learning using skill modelling. Next, the main methods of skill modelling will be reported with our own words, making the proper citations where appropriate. The methodology used for the comparison of the different methods of skill modelling is based on the following dimensions:
- The power of the domain knwoledge models.
- The time required by teachers or designers.
- The maintainability of the solution
- The number of skills considered
- The complexity modeling for calculating the skill.

About the main results:
1) The IRT models does not provide any semantic information about the relationship among items.The KS can give relationships about the items but only the precedence relationship. The BN can give relationships among the different skills and exercises but they should be limited to the rules of the BN. The semantic solutions can provide more fained-grained relationships among skills and items based on semantic relationships. One of the main disadvantage of the KS, BN, or semantic solutions is the difficulty of creating a solid structure that models the domain knowledge and maintain it.
2) While the IRT does not require any modeling, all the other methods require different modelings more or less complex. This requires the effort of course designers and teachers. In addition, the complexity of maintainability increases, because if there is a new skill or new exercise, this should be included in the modeling.
3)The IRT just consider one skill in the traditional model. BN and KS can consider different skills as well as the semantic solutions. This is a limitation of IRT since sometimes it is better to model the domain with different skills.
4) The IRT way of calculation of skills provides a more exact solution.

Keywords:

Item response theory, e-learning, bayesian networks, knowledge spaces, intelligent tutoring systems, educational data mining, learning analytics.