outputs of nbest-optimize...

[Andreas Stolcke]

In message <3C5B113F.A2420B83 at inzigo.com>you wrote:
> Hi,
> 
> The followings are the output of "nbest-optimize":
> 
> 14 reference words
>    weights = 0.125 1 0
>    normed = 1 8 0
>    scale = 8
> NEW BEST ERROR: 0 (0/word)
>    weights = 0.125 1 0
>    normed = 1 8 0
>    scale = 8
> stopping due to convergence
> original errors = 0 (0/word)
> best errors = 0 (0/word)
>    weights = 0.125 1 0
>    normed = 1 8 0
>    scale = 8
> 
> What are the meanings of those? For example, weights=0.125 1 0 0.125
> what weights do they refer to?
> normed=1 8 0, what are these numbers?

"weights=" shows the raw weights to be applied to the scores, in the order
in which they were given on the command line.  the first three weights
refer to the "standard" scores: acoustic model, language model, and
word count.  (In your case you didn't give any extra scores.)

"normed=" shows the weights normalized so that the acoustic model weight
is 1.  "scale=" is the normalizing denominator to achieve this.

The use of the normalized score is for programs that assume that 
the AM weight is implicitly 1 (such as the rescore-reweight script).
Decoding based on posterior probabilities depends on the scale value
as well (see the description of nbest-rover).  More info can be found
in the "nbest-scripts" man page.

--Andreas