Automatic construction of concept maps from unstructured text

Official URL: https://risc01.sabanciuniv.edu/record=b2771025_(Table of contents)

Acquiring knowledge about a topic is a complex process that requires an individual to gradually understand all concepts related to this topic. The learning experience for individuals can be enhanced by visualizing the key characteristics that serve as a guide for learning. Equipping learners with a concept map that orders concepts, represented as nodes, to be studied according to their prerequisite order, indicated by direct edges, helps them stay on track. However, existing algorithms for automatic prerequisite detection are too inaccurate, which reduces learners’ trust in such maps as one assumes the prerequisite relation to be completely reliable. In this work we propose to replace prerequisite relations with less authoritative coverage relations, as they indicate only that one concept is broader and related to another one. Since most of the prerequisites of a given concept are less difficult than the concept itself, we argue that combining the coverage relation with concept’s difficulty scores encodes similar semantics as the prerequisite relation. However, due to the coverage relation being less authoritative, potentially inaccurately detected coverage relations may be ignored by learners.Such relations are considered more like recommendations instead of facts. In turn, this change in perception about the reliability of the edges in the resulting concept map creates less frustration for learners. Further, two additional aspects, the unstructured nature and the abundance of the learning materials should also be considered for devising a salable method for concept map’s construction. With this in mind, this thesis aims to automatically construct from unstructured textual learning materials a concept map that encodes concept difficulty as node color and vi connects concepts through coverage edges. To that end, we divide the problem into two subtasks: extracting concepts from unstructured text and constructing the concept map by estimating the difficulty of each extracted concept before inserting coverage edges among them. Specifically, we first develop an unsupervised method to extract concepts from unstructured textual learning materials and then compute a difficulty score for each of the identified concepts in the second subtask with a novel unsupervised method. We find that our concept extraction method is more accurate than existing state-of-the-art methods. To the best of our knowledge, we have proposed the first unsupervised method for finding the concept’s difficulty. Our experiments demonstrate the feasibility of our proposed difficulty prediction method.It also provides evidence for our core assumption that prerequisites of a given concept tend to be easier than the concept itself, which renders our methodology viable. These findings imply that our proposed methodology yields concept maps for courses that help individuals navigate concepts more successfully in practice.