Abstract: A computer-based method and system for establishing topic words to represent a document, the topic words being suitable for use in document retrieval. The method includes determining document keywords from the document; classifying each of the document keywords into one of a plurality of preestablished keyword classes; and selecting words as the topic words, each selected word from a different one of the preestablished keyword classes, to minimize a cost function on proposed topic words. The cost function may be a metric of dissimilarity, such as cross-entropy, between a first distribution of likelihood of appearance by the plurality of document keywords in a typical document and a second distribution of likelihood of appearance by the plurality of document keywords in a typical document, the second distribution being approximated using proposed topic words. The cost function can be a basis for sorting the priority of the documents.

Abstract: A facility is provided for normalizing the format of stored data records using a dictionary that is generated from a training set of data records having predefined formats.