Abstract: After combining of organically the recognition information on single Manchu characters from relevant system with the information on phrases to set up a statistical information database of Manchu phrases and underdetermined word sets, Bayes rules are used to synthesize the prior probability of underdetermined Manchu word sets and posterior probability of phrases. A data construction is thus developed to improve efficiently the recognition rate, which is rational and easy to implement especially available to detect and correct those rejected and incorrectly recognized words output from the SCR single character recognition system. Experiment shows that the post-processing performance depends on not only the language model but the accurate estimate of posterior probability. In addition, the higher the recognition rate of SCR, the stronger the rectifiability of postprocessing.