Today, an increasing number of universities use distance learning systems. However, such systems may be dull and unhelpful for students studying on their own, because the learning materials are the same for every student. To overcome this problem, it is necessary to provide learning materials that are not only well designed, but also personalized. A metadata and Semantic Web approach is effective for such personalization. Since 2001, we have been developing and utilizing a multimedia e-Learning system, called e-Math, for economical mathematics. The goal of our e-Math system is to attain ever more natural interactions between a system and a student. To automate the interactions and learning materials in the e-Math system, we have adopted a metadata framework approach. We have also introduced an inference engine and a knowledge base in our implementation of the interaction agent in the e-Math system. This paper describes the e-Math Interaction Agent that dynamically automates personalized Web-based learning materials and explains its metadata framework.