Research on modeling the identification of materials related to a given topic, person or country is reported. The models use Bayesian analysis, and a sparsity-inducing prior distribution of the chance that a term will be useful. The result is concise computer-generated models, which can be understood, and improved, by human users of the tools. Examples are given based on technical literature, and on materials of interest to intelligence and policy analysts. The methods are particularly effective in learning to recognize materials pertinent to a specific topic, from very small sets of learning materials, provided that general background information (such as might be found in a text-book or encyclopedia) can be used to set the prior probability that a term will be used in the machine learning model.