A framework for structuring prerequisite relations between concepts in educational textbooks

In our age we are experiencing an increasing availability of digital educational resources and self-regulated learning. In this scenario, the development of automatic strategies for organizing the knowledge embodied in educational resources has a tremendous potential for building personalized learning paths and applications such as intelligent textbooks and recommender systems of learning materials. To this aim, a straightforward approach consists in enriching the educational materials with a concept graph, i.a. a knowledge structure where key concepts of the subject matter are represented as nodes and prerequisite dependencies among such concepts are also explicitly represented. This thesis focuses therefore on prerequisite relations in textbooks and it has two main research goals. The first goal is to define a methodology for systematically annotating prerequisite relations in textbooks, which is functional for analysing the prerequisite phenomenon and for evaluating and training automatic methods of extraction. The second goal concerns the automatic extraction of prerequisite relations from textbooks. These two research goals will guide towards the design of PRET, i.e. a comprehensive framework for supporting researchers involved in this research issue. The framework described in the present thesis allows indeed researchers to conduct the following tasks: 1) manual annotation of educational texts, in order to create datasets to be used for machine learning algorithms or for evaluation as gold standards; 2) annotation analysis, for investigating inter-annotator agreement, graph metrics and in-context linguistic features; 3) data visualization, for visually exploring datasets and gaining insights of the problem that may lead to improve algorithms; 4) automatic extraction of prerequisite relations. As for the automatic extraction, we developed a method that is based on burst analysis of concepts in the textbook and we used the gold dataset with PR annotation for its evaluation, comparing the method with other metrics for PR extraction.