Compensating for Word Posterior Estimation Bias in Confusion Networks

This paper looks at the problem of confidence estimation at the word network level, where multiple hypotheses from a recognizer are represented in a confusion network. Given features of the network, an SVM is used to estimate the probability that the correct word is missing from a candidate slot and then other word probabilities are normalized accordingly. The result is a reduction in overall bias of the estimated word posteriors and an improvement in the confidence estimate for the top word hypothesis in particular