[SRILM User List] cashe adaptation

[Andreas Stolcke]

iman emamgholipour wrote:
>  
>    Dear Dr. Stolcke,
>
>      I used SRILM for cache adapting in language models. and I used
>     this command to reach that goal.
>
>         /*ngram -rescore-ngram CacheAdaptedLanguageModel.txt -lm
>         Unigram.txt  -mix-lm BackgroundLM.txt -cache-lambda
>         "lambda_value" -cache "lenght" -write-lm 
>         CacheAdaptedLanguageModel_Output.txt*/
>
>     But unconventionally the results show that cache length has no
>     effect on /Word Error Rate /. I test it till 10000 for length.
>     Would you please tell me where my mistake is?
>
>     Best regards
>      
>
You cannot use -rescore-ngram in conjunction with -cache .  Cache-based 
adaptation only makes sense when processing words sequentially.
The -rescore-ngram function is to reassign ngram conditional 
probabilities in an LM,  without providing more distant context. So a 
cache is useless here.
And -write-lm is only supported to pure ngram and a few other 
specialized LM formats.

ngram -cache only makes sense in combination with options that evaluate 
the LM on input text, such as -ppl,  -rescore, etc.
(nbest rescoring is also problematic since you probably don't want one 
nbest hypothesis for a sentence to affect the cache for another 
alternate hypothesis.)

Andreas