Semantics-based topic inter-relationship extraction

Maintaining large collection of documents is an important problem in many areas of science and industry. Different analysis can be performed on large document collection with ease only if a short or reduced description can be obtained. Topic modeling offers a promising solution for this. Topic modeling is a method that learns about hidden themes from a large set of unorganized documents. Different approaches and alternatives are available for finding topics, such as Latent Dirichlet Allocation (LDA), neural networks, Latent Semantic Analysis (LSA), probabilistic LSA (pLSA), probabilistic LDA (pLDA). In topic models the topics inferred are based only on observing the term occurrence. However, the terms may not be semantically related in a manner that is relevant to the topic. Understanding the semantics can yield improved topics for representing the documents. The objective of this paper is to develop a semantically oriented probabilistic model based approach for generating topic representation from the document collection. From the modified topic model, we generate 2 matrices-a document-topic and a term-topic matrix. The reduced document-term matrix derived from these two matrices has 85% similarity with the original document-term matrix i.e. we get 85% similarity between the original document collection and the documents reconstructed from the above two matrices. Also, a classifier when applied to the document-topic matrix appended with the class label, shows an 80% improvement in F-measure score. The paper also uses the perplexity metric to find out the number of topics for a test set.