interpolate-ngram

NAME

interpolate-ngram - Constructs a statically interpolated n-gram language model from component n-gram models.

SYNOPSIS

interpolate-ngram [Options]

DESCRIPTION

Interpolates multiple n-gram models by computing appropriate interpolation weights from optional features and constructing a statically interpolated n-gram model. Parameters can be optionally tuned to optimize development set performance.

Filename argument can be an ASCII file, a compressed file (ending in .Z or .gz), or '-' to indicate stdin/stdout.

OPTIONS

`-h`, `-help`	Print this message.
`-verbose` (1)	Set verbosity level.
`-o`, `-order` (3)	Set the n-gram order of the estimated LM.
`-v`, `-vocab`	Fix the vocab to only words from the specified file.
`-u`, `-unk`	Replace all out of vocab words with `<unk>`.
`-l`, `-lm`	Interpolate specified LM files.
`-t`, `-text`	Interpolate models trained from text files.
`-c`, `-counts`	Interpolate models trained from counts files.
`-s`, `-smoothing` (`ModKN`)	Specify smoothing algorithms. (`ML`, `FixKN`, `FixModKN`, `FixKN`d, `KN`, `ModKN`, `KN`d)
`-wf`, `-weight-features`	Specify n-gram weighting features.
`-i`, `-interpolation` (`LI`)	Specify interpolation mode. (`LI`, `CM`, `GLI`)
`-if`, `-interpolation-features`	Specify interpolation features.
`-tpo`, `-tie-param-order` (true)	Tie parameters across n-gram order.
`-tpl`, `-tie-param-lm` (false)	Tie parameters across LM components.
`-p`, `-params`	Set initial model params.
`-oa`, `-opt-alg` (`Powell`)	Specify optimization algorithm. (`Powell`, `LBFGS`, `LBFGSB`)
`-op`, `-opt-perp`	Tune params to minimize dev set perplexity.
`-ow`, `-opt-wer`	Tune params to minimize lattice word error rate.
`-om`, `-opt-margin`	Tune params to minimize lattice margin.
`-wb`, `-write-binary`	Write LM/counts files in binary format.
`-wp`, `-write-params`	Write tuned model params to file.
`-wv`, `-write-vocab`	Write LM vocab to file.
`-wl`, `-write-lm`	Write ARPA backoff LM to file.
`-ep`, `-eval-perp`	Compute test set perplexity.
`-ew`, `-eval-wer`	Compute test set lattice word error rate.
`-em`, `-eval-margin`	Compute test set lattice margin.

INTERPOLATION

MITLM currently supports the following interpolation algorithms:

LI - Interpolates the component models using linear interpolation. Constructs a model consisting of the union of all observed n-grams, with probabilities equal to the weighted sum of the component n-gram probabilities, using backoff, if necessary. Backoff weights are computed to normalize the model. In linear interpolation, the weight of each component is a fixed constant.
CM - Interpolates the component models using count merging (Bacchiani et al., 2006; Hsu, 2007). Unlike linear interpolation, the weight of each component depends on the n-gram history count for that component.
GLI - Interpolates the component models using generalized linear interpolation (Hsu, 2007). Unlike count merging, the weight of each component depends on arbitrary features of the n-gram history, specified through the -interpolation-features option.

REFERENCES

M. Bacchiani, M. Riley, B. Roark, and R. Sproat, MAP adaptation of stochastic grammars, Computer Speech & Language, vol. 20, no. 1, pp. 41-68, 2006.
B. Hsu, Generalized linear interpolation of language models, in Proc. ASRU, 2007.
F. Jelinek and R. Mercer, Interpolated estimation of Markov source parameters from sparse data, in Proc. Workshop on Pattern Recognition in Practice, 1980.