Towards automatic identification of core concepts in educational resources

Automatically identifying and extracting key ideas and concepts from educational resources is an important but challenging computational task. We present a supervised machine learning approach to assessing the “coreness” of concepts expressed by resource sentences. The algorithm has been developed and evaluated in the domain of science education where coreness refers to the degree to which a sentence embodies key concepts important to developing a robust understanding of the domain. Our method operates by automatically computing and leveraging the degree of semantic similarity between resource sentences and standard domain concepts designed by human experts for various STEM domains. In our experiments, the algorithm demonstrates high accuracy in identifying sentence coreness when there is agreement between human experts on the coreness rating. We also present performance comparisons with a number of baseline systems.