Compressed Inference for Probabilistic Sequential Models