Learning document representations using subspace multinomial model

Subspace multinomial model (SMM) is a log-linear model andcan be used for learning low dimensional continuous representationfor discrete data. SMMand its variants have been used forspeaker verification based on prosodic features and phonotacticlanguage recognition. In this paper, we propose a new variantof SMM that introduces sparsity and call the resulting modelas `1 SMM. We show that `1 SMM can be used for learningdocument representations that are helpful in topic identificationor classification and clustering tasks. Our experiments in documentclassification show that SMM achieves comparable resultsto models such as latent Dirichlet allocation and sparse topicalcoding, while having a useful property that the resulting documentvectors are Gaussian distributed.