In this paper we deal with the problem of seeking relevant information sources selected from scientific or other electronic publications. In the era of information surfeit, it is getting more and more difficult to extract relevant and reliable sources of information from the huge number of e-sources. The starting point is user's query for a given concept or topic. Our algorithm applies machine learning methods in order to propose hypothetic explications of the sought terms based on pieces of information extracted from the potentially relevant e-sources. Hypotheses, formalized in the TIL-Script language, are incrementally built by applying heuristic functions. The user thus obtains as closed approximations of the meaning of the sought terms as possible that at the same time provide fine-grained keyword definitions. As a result, it should be much easier to decide which of the e-sources are relevant for user's interest.