Self-adaptive GA, quantitative semantic similarity measures and ontology-based text clustering

As the common clustering algorithms use vector space model (VSM) to represent document, the conceptual relationships between related terms which do not co-occur literally are ignored. A genetic algorithm-based clustering technique, named GA clustering, in conjunction with ontology is proposed in this article to overcome this problem. In general, the ontology measures can be partitioned into two categories: thesaurus-based methods and corpus-based methods. We take advantage of the hierarchical structure and the broad coverage taxonomy of Wordnet as the thesaurus-based ontology. However, the corpus-based method is rather complicated to handle in practical application. We propose a transformed latent semantic analysis (LSA) model as the corpus-based method in this paper. Moreover, two hybrid strategies, the combinations of the various similarity measures, are implemented in the clustering experiments. The results show that our GA clustering algorithm, in conjunction with the thesaurus-based and the LSA-based method, apparently outperforms that with other similarity measures. Moreover, the superiority of the GA clustering algorithm proposed over the commonly used k-means algorithm and the standard GA is demonstrated by the improvements of the clustering performance.