Extended Multidimensional Conceptual Spaces in Document Classification
Mulugeta Hadish · OhioLink ETD Center (Ohio Library and Information Network) · 2008
Content based retrieval, entailing knowledge representation, can be essentially described as assessing the similarity between objects which constitute primitives as their building blocks.This thesis presents a novel approach in document classification using an Extended Multidimensional Conceptual Spaces (EMCS) framework to address problem areas that require similarity assessment.As a typical problem domain, articles on Breast, Brain, and Colon Cancers were obtained from PubMed 1 as search results of binary based queries.Since not all terms carry equal discriminatory information, only those that do were identified and treated as primitives by carrying out document pre-processing.Salient weights associated to each term thereafter were assessed statistically-computing their normalized term frequencies and inverse document frequencies.The product of these frequencies were determined as ideal features for their concept category.Example documents were also preprocessed and expressed in terms of the normalized frequency values for each of their stemmed terms.Then, the ideal normalized frequencies of the selected features are compared with every term of each example documents using the degree of difference measure.This computation essentially transformed the frequencies of terms into degrees of membership and cardinality fuzzy sets.Each of the fuzzy sets from the example documents were aggregated using cardinality of fuzzy sets and transformed into co-occurrence matrix in respect to the three concepts.Similarly, fuzzy sets representing test documents were also transformed into co-occurrence matrices in respect to the concepts.Similarity was finally assessed based on a Frobenius distance measure between a core concept matrix and a test matrix.The experiment demonstrated feasibility of the framework in document classification.