Abstract: Systems and methods are provided for retrieving data relevant to a subject of interest. Occurrences of each of a plurality of n-grams within the data record are identified. A multinomial distribution is defined from the respective numbers of occurrence of a subset of the plurality of n-grams. The multinomial distribution is stored in a semantic model as a point on an information manifold. The semantic model is configured to represent an indexed family of probability distributions as points on the information manifold. It is determined if the data record is relevant to the subject of interest according to the position of the point on the information manifold, and the data record is retrieved if the data record is relevant to the subject of interest.