QA systems distinguish themselves from search engines by the type of manipulated data -- well-formed grammatical units endowed with semantic content -- and its purpose: not just for retrieving facts, but also for more complex learning activities. Current conceptions about learning assume learners as active agents and not passive recipients or simple recorders of information. Therefore, the sequence in which knowledge is assimilated is of high importance. Recently developed recommendation techniques for search engine queries try to leverage the order in which users navigate through them. Although a similar approach might improve the learning experience with QA systems, questions are still considered as any other recommendation item.
In this thesis, two categories of recommendation techniques are defined. The first type, the short-term recommender, considers only the most recently asked question to produce semantically related suggestions using a new semantic similarity measure. The second one, the long-term or learning-oriented question recommender is a novel approach that exploits not only the user's history log, but also two important question attributes: its topic and learning objective. For this purpose, a domain-specific topic-taxonomy and Bloom's learning framework is employed, whereas for modeling the order in which questions are selected, variable length Markov chains are used.