N-Gram Model Smoothing with Independently Controllable Parameters

摘要: Described is a technology by which probability estimated for token in sequence of tokens based upon number zero or more times (actual counts) that the was observed training data. The may be word sequence, and used statistical language model. A discount parameter set independently interpolation parameters. If at least once data, an are computed summed to provide probability. not observed, computing backoff Also described various ways obtain