Combination of Multiple predictors to Improve Confidence Measure Based on Local Posterior Probabilities
Fu Yue-wen, Limin Du · 2006
Recently local word posterior probabilities computed from word expansion during stack decoding search was proposed to be a confidence measure under real-time condition. However, much approximation in its computation limits its quality. In this paper, we intend to improve its performance by using a decision tree to combine it with other real-time predictors. A series of other predictors are constructed and the experiments on different combination of predictors using the decision tree are carried out. The experimental results show that confidence measure based on local word posterior probability can be improved significantly (18.9% confidence error rate improvement relatively in our experiments) by combining with other real-time predictors. The experiments also show that local posterior probabilities of adjacent words are relatively effective predictors.