[SRILM User List] sausage edges merging

[Andreas Stolcke]

Tim Kempton wrote:
>
> Hi
>
>  
>
> I'm trying to prevent the merging of sausage (i.e. confusion network) 
> edges. I thought the option "-no-merge" might work e.g. I was hoping 
> to see two edges labelled "hello" in the following sausage but there 
> is just one:
>
>  
>
> -bash-3.2$ echo -e "-5 -1 2 hello world\n-5 -2 2 hello mould" 
> |nbest-lattice -use-mesh -nbest - -write - -no-merge
>
>  
>
> hello world
>
> name -
>
> numaligns 2
>
> posterior 1
>
> align 0 hello 1
>
> align 1 world 0.909091 mould 0.0909091
>
>  
>
> I get the same result whether I use the "-no-merge" option or not. 
> Maybe I've got the wrong end of the stick and this option is for 
> something else. I am using SRILM version 1.5.8, but I don't believe 
> there's been any relevant changes to nbest-lattice since then.
>
>  
>
> The reason I want to do this is to preserve timing information from an 
> NBestList2.0 list; when the edges get merged there is also a loss of 
> backtrace information (when using -nbest-backtrace).
>
Sorry, but -no-align has no effect with -use-mesh because word confusion 
networks as implemented only support unique word labels per alignment 
position (all the information is hashed on the word type).

However, you can work around this by
(1) making word labels unique -- hello-1, hello-2, etc.
(2) use the -dictionary option to specify an alignment cost based on 
dictionary pronunciations. The pronunciations could be real ones (so all 
the hello-1, hello-2 etc. have the same pronunciation and hence align) 
or you could even use dummy pronunciations that just consist of the 
"real" word labels:   hello-1 -> hello, hello-2 -> hello, etc.
That was the alignment cost will exactly mimic the usual word identity.

Note I haven't tried this, but it should work.

Andreas

>  
>
> Thanks,
>
> Tim
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> SRILM-User site list
> SRILM-User at speech.sri.com
> http://www.speech.sri.com/mailman/listinfo/srilm-user