Comparison and evaluation of approaches for automatic extraction of knowledge structure from learning materials

Abstract

Given a semi-structured learning material, can we infer which keywords are important and how they are connected? We formalize this question as the automatic extraction of knowledge structure and develop approaches based on measures for analyzing the relationship of keywords in the learning material. The amount of learning materials increases exponentially owing to the advance of the IT technology, and the size of the online education market gets larger because of the prevalence of the Internet. However, many people experience difficulties in exactly finding the proper learning material for them. The extraction of knowledge structure can facilitate this task since it finds or recommends proper learning materials in terms of knowledge structure. Along this direction, location based approaches have been widely used. However, most learning materials are semi-structured so that it is hard to find exact semantic separation. There are two major contributions in this thesis: (a) We use SVM(Singular Vector Decomposition) and a flexible edit distance matching algorithm to find semantic separation of leaning materials and compare proposed methods and location based methods (b) We conduct user study with domain experts for about a hundred learning materials to measure actual expert’s knowledge structure. Results demonstrate 20% superior performance of the proposed methods over the location based methods.