Linear-chain conditional random fields (LC-CRFs) have been successfully applied in manystructured prediction tasks. Many previous ex-tensions, e.g. replacing local factors by neuralnetworks, are computationally demanding. Inthis paper, we extend conventional LC-CRFs byreplacing the local factors with sum-product net-works, i.e. a promising new deep architecture al-lowing for exact and efficient inference. The pro-posed local factors can be interpreted as an exten-sion of Gaussian mixture models (GMMs). Thus,we provide a powerful alternative to LC-CRFsextended by GMMs. In extensive experiments,we achieved performance competitive to state-of-the-art methods in phone classification and opti-cal character recognition tasks.

Top