The authors report a knowledge-based approach to classification. The proposed methodology uses personal construct theory for interviewing a domain expert to elicit classification knowledge. This interview results in raw data, which, on analysis, yields the relationship between different concepts from a user perspective. After finding the relationships, the user is asked to delineate the boundaries which enclose like concepts. With such a grouping of concepts, the authors develop a methodology to establish a relationship between the concepts and the index terms constituting document representations. This relationship is employed to assign a document to the most appropriate cluster. The knowledge elicited from the expert is mapped to system observable features of documents to develop a classification. The techniques developed are experimentally validated.