Menu ▾ ▴

FSG to N-gram

Help
Pranav Jawale
2012-05-24
2012-09-22

  • Pranav Jawale

    Pranav Jawale - 2012-05-24

    Hello,

    It seems to me that in principle one can write a finite state grammar in (say)
    bi-gram fashion. One can create a bigram LM, and then just edit the
    probabilities to reflect FSG logic.

    For example, let the FSG be like

     hello (alpha | beta ) how are you?

    corresponding bigram LM will have only following probabilities non-zero
    (disable backoff)

    P(alpha | hello)
    P(alpha | hello)
    P(alpha | hello)

    P(beta | hello)
    P(beta | hello)
    P(beta | hello)

    P(how | alpha)
    P(how | beta)

    P(are | how)
    P(you | are)

    P(|you)

    I am not sure if this can be generalized though (counter-examples please!). If
    it can be, then we can use n-gram code which can give nbest list for FSG
    (currently unsupported feature, see https://sourceforge.net/projects/cmusphin
    x/forums/forum/5471/topic/4473108).

     

  • Pranav Jawale

    Pranav Jawale - 2012-05-24

    Nevermind, Found counter-example

    <sentence> = w1 w2 w1 (w2 | w3) w2 w1 (w2 | w3 | w4) </sentence>

    here one needs a 4-gram representation to convert FSG -> n-gram. So the method
    may work only for some special case FSGs (like in prev post).

     

  • Nickolay V. Shmyrev

    Nickolay V. Shmyrev - 2012-05-28

    If it can be, then we can use n-gram code which can give nbest list for FSG
    (currently unsupported feature, see https://sourceforge.net/projects/cmusphin
    x/forums/forum/5471/topic/4473108).

    The easiest way to support n-best is actually to implement n-best.

     

  • Nickolay V. Shmyrev

    Nickolay V. Shmyrev - 2012-05-28

    It has been some time since I looked at it but the nBest code for FSG
    already exists

    The code exists but it does not work properly. It doesn't take fsg
    probabilities into account, only the acoustic scores.

     

Log in to post a comment.