Look to the material on Rough Sets. Categorical data is not always perfect. No mechanism short of DNA analysis is perfect at classifying gender. Hence we must often account for membership that is uncertain. Machine learning methods and methods from rough sets can do this. Otherwise the standard statistical packages have some methods for clustering.

If you can do clustering algorithms on your data (i.e. using similarity/proximity metric or rule), then it is possible to define tree structures (dendrograms) that specify clusters of clusters, and so on. Then you can create ultrametric for similarity measures. I recommend these two papers:

The Evolving Tree of Industrial Life: An Approach to the Transformation of European Industry

http://www.business.aau.dk/~esa/evolution/esapapers/esa03/torino.pdf

Ultrametric and Generalized Ultrametric in Logic and in Data Analysis

http://arxiv.org/pdf/1008.3585

You can also look at Jaccard Similarity and Locality Sensitive Hashing. See: Chapter 3 in http://infolab.stanford.edu/~ullman/mmds.html

Jaccard is in fact very good and has the big advantage to work with sparse matrixes that are very common in science.

for feature reduction to text processing usually use tf/idf to get weight for each term. then if the term weight smaller than threshold is must be reduced. you also can use stopword list to reduce unimportant term. for clustering you can use bayesian classtering, or SOM.

There is a very important one nowadays wich is about the semantic similarity: http://www.academia.edu/2884350/Evolutionary_conceptual_clustering_of_semantically_annotated_resources

http://www.academia.edu/2974373/Semantic_clustering_for_a_functional_text_classification_task