[SRILM User List] Adaptation in SRILM

[Andreas Stolcke]

Mehdi hoseini wrote:
> hi,
> I have questions in two Model adaptation with SRILM:
>
>    1. referring to "statistical language model adaptation:review and
>       perspectives"(Jerome R. Bellegarda 2003), what does *-mix-lm
>       *option in *ngram* does?* *is it just *interpolate* adaptation
>       language model with background language model and lambda as an
>       interpolate coefficient?  or it use* MAP adaptation* method to
>       merge two language models?
>
-mix-lm (together with the -bayes 0 option)  implements the 
interpolation (mixing)  of multiple LMs at the probability level.
(eqn. 4 in the paper).   Without the -bayes option an approximation to 
the interpolated LM is generated in the form of a single combined 
backoff LM.

>    1. how can i merge two language model in SRILM, using Back-off
>       technique (Besling and Meier,  1995)
>
The fill-in adaptation method is not currently implemented.

Andreas