From Texts to Prerequisites. Identifying and Annotating Propaedeutic Relations in Educational Textual Resources

Prerequisite Relations (PRs) are dependency relations established between two distinct concepts expressing which piece(s) of information a student has to learn first in order to understand a certain target concept. Such relations are one of the most fundamental in Education, playing a crucial role not only for what concerns new knowledge acquisition, but also in the novel applications of Artificial Intelligence to distant and e-learning. Indeed, resources annotated with such information could be used to develop automatic systems able to acquire and organize the knowledge embodied in educational resources, possibly fostering educational applications personalized, e.g., on students' needs and prior knowledge. The present thesis discusses the issues and challenges of identifying PRs in educational textual materials with the purpose of building a shared understanding of the relation among the research community. To this aim, we present a methodology for dealing with prerequisite relations as established in educational textual resources which aims at providing a systematic approach for uncovering PRs in textual materials, both when manually annotating and automatically extracting the PRs. The fundamental principles of our methodology guided the development of a novel framework for PR identification which comprises three components, each tackling a different task: (i) an annotation protocol (PREAP), reporting the set of guidelines and recommendations for building PR-annotated resources; (ii) an annotation tool (PRET), supporting the creation of manually annotated datasets reflecting the principles of PREAP; (iii) an automatic PR learning method based on machine learning (PREL). The main novelty of our methodology and framework lies in the fact that we propose to uncover PRs from textual resources relying solely on the content of the instructional material: differently from other works, rather than creating de-contextualised PRs, we acknowledge the presence of a PR between two concepts only if emerging from the way they are presented in the text. By doing so, we anchor relations to the text while modelling the knowledge structure entailed in the resource. As an original contribution of this work, we explore whether linguistic complexity of the text influences the task of manual identification of PRs. To this aim, we investigate the interplay between text and content in educational texts through a crowd-sourcing experiment on concept sequencing. Our methodology values the content of educational materials as it incorporates the evidence acquired from such investigation which suggests that PR recognition is highly influenced by the way in which concepts are introduced in the resource and by the complexity of the texts. The thesis reports a case study dealing with every component of the PR framework which produced a novel manually-labelled PR-annotated dataset.