Dirichlet Compound Multinomials Statistical Models.docVIP

Applied Mathematics, 2012, 3, 2089-2097 /10.4236/am.2012.312A288 Published Online December 2012 (http://www.SciRP.org/journal/am) Dirichlet Compound Multinomials Statistical Models Paola Cerchiello, Paolo Giudici Department of Economics and Management, University of Pavia, Pavia, Italy Email: paola.cerchiello@unipv.it Received August 31, 2012; revised September 31, 2012; accepted October 6, 2012 ABSTRACT This contribution deals with a generative approach for the analysis of textual data. Instead of creating heuristic rules for the representation of documents and word counts, we employ a distribution able to model words along texts considering different topics. In this regard, following Minka proposal (2003), we implement a Dirichlet Compound Multinomial (DCM) distribution, then we propose an extension called sbDCM that takes explicitly into account the different latent topics that compound the document. We follow two alternative approaches: on one hand the topics can be unknown, thus to be estimated on the basis of the data, on the other hand topics are determined in advance on the basis of a prede- fined ontological schema. The two possible approaches are assessed on the basis of real data. Keywords: Textual Data Analysis; Mixture Models; Ontology Schema; Reputational Risk 1. Introduction well. By the way the goal was achieved only partially, in fact the multinomial mixtures, which components can be interpreted as topics, offer a probabilistic justification at words but not at documents level. In fact the latter are represented merely as list of mixing proportions derived from mixture components. Moreover, the multinomial distribution presents as many values as there are in the training documents and therefore it learns topic mixture on those trained documents. The extension to previously unseen documents is not appropriate since there can be new topics. In order to overcome the asymmetry between words and do