Abstract

Generative topic models such as LDA are limited by their inability to utilize nontrivial input features to enhance their performance, and many topic models assume that topic assignments of different words are conditionally independent. Some work exists to address the second limitation but no work exists to address both. This paper presents a conditional topic random field (CTRF) model, which can use arbitrary nonlocal features about words and documents and incorporate the Markov dependency between topic assignments of neighboring words. We develop an efficient variational inference algorithm that scales linearly in terms of topic numbers, and a maximum likelihood estimation (MLE) procedure for parameter estimation. For the supervised version of CTRF, we also develop an arguably more discriminative max-margin learning method. We evaluate CTRF on real review rating data and demonstrate the advantages of CTRF over generative competitors, and we show the advantages of max-margin learning over MLE.