In this paper, we study some learner modelling issues underlying the construction of an e-learning system that recommends research papers to graduate students wanting to learn a new research area. In particular, we are interested in learner-centric and paper-centric attributes that can be extracted from learner profiles and learner ratings of papers and then used to inform the recommender system. We have carried out a study of students in a large graduate course in software engineering, looking for patterns in such “pedagogical attributes”. Using mean-variance and correlation analysis of the data collected in the study, four types of attributes have been found that could be usefully annotated to a paper. This is one step towards the ultimate goal of annotating learning content with full instances of learner models that can then be mined for various pedagogical purposes.