Nonparametric method of topic identification using granularity concept and graph-based modeling

This paper aims to classify the large unstructured documents into different topics without involving huge computational resources and a priori knowledge. The concept of granularity is employed here to extract contextual information from the documents by generating granules of words (GoWs), hierarchically. The proposed granularity-based word grouping (GBWG) algorithm in a computationally efficient way group the words at different layers by using co-occurrence measure between the words of different granules. The GBWG algorithm terminates when no new GoW is generated at any layer of the hierarchical structure. Thus multiple GoWs are obtained, each of which contains contextually related words, representing different topics. However, the GoWs may contain common words and creating ambiguity in topic identification. Louvain graph clustering algorithm has been employed to automatically identify the topics, containing unique words by using mutual information as an association measure between the words (nodes) of each GoW. A test document is classified into a particular topic based on the probability of its unique words belong to different topics. The performance of the proposed method has been compared with other unsupervised, semi-supervised, and supervised topic modeling algorithms. Experimentally, it has been shown that the proposed method is comparable or better than the state-of-the-art topic modeling algorithms which further statistically verified with the Wilcoxon Rank-sum Test.