Back-off method for n-gram smoothing based on binomial posteriori distribution

The n-gram language model is powerful for treating natural spoken language, however it requires large amounts of spoken language corpus to estimate reliable model parameters. To estimate n-gram probabilities from sparse data, Katz´s (1987) back-off smoothing method is promising. However, this approach is sometimes unstable because it uses singleton heuristics based on Turing´s formula. This paper proposes a new back-off method based on binomial posteriori distribution of n-gram probabilities, which achieves stable and more effective n-gram smoothing using a sophisticated calculation formula with no heuristics