[hts-users:01427] Re: Regarding the new IT command of HHEd of HTS-2.1RC2

["Heiga ZEN (Byung Ha CHUN)" <zen@xxxxxxxxxxxxxxx>]

Hi,

Tienshyang Huei Liang wrote (2008/06/16 0:25):

> According to its brief description, I presume that "IT" calculates means, 
> variances and weights of all the leaf nodes of a given decision tree, 
> from data of a given full-context model. 
> Namely, in the case below,
>
>     +---fullcontext.mmf---+     +------------+     +---clustered.mmf(A)
>     |         Questions---+-->--+ TB of HHEd +-->--+
>     |   statistic files---+     +------------+     +---decision tree
>     |   ---+-----------                                ------+------
>     |      |                                                 |
>     V      |    +--------------------<-----------------------+
>     |      |    |
>     |      |    +---->----+     +------------+
>     |      +--------->----+-->--+ IT of HHEd +-->--clustered.mmf(B)
>     +---------------->----+     +------------+
>
> clustered.mmf(A) should be exactly the same as clustered.mmf(B).
> Is my presumption correct? If so, I find something strange.
>
> Here are a tied stream of clustered.mmf(A),
> ~p "spectrum_s6_22"
> <STREAM> 1
> <MEAN> 123
>  1.047655e-002 1.504280e-002 2.348183e-002  ....
> <VARIANCE> 123
>  7.492220e-007 8.066245e-007 5.884825e-006  ....
> <GCONST> -1.209281e+003
>
> and a tied stream of clustered.mmf(B),
> ~p "spectrum_s6_22"
> <STREAM> 1
> <MEAN> 123
>  1.047094e-002 1.496241e-002 2.337582e-002 ....
> <VARIANCE> 123
>  8.766731e-007 1.213980e-006 7.213330e-006 ....
> <GCONST> -1.199125e+003
>
> Means and variances of the two tied streams are not equal. 
> Especially, variance differences of some dimensions are not negligible. 
> I don't know why this happens.

I checked it on my environment but I get the exactly the same mmfs for clustered.mmf(A) and (B).
Did you see un-reestimated clustered.mmf(A)  (just after TB clustering) ?
I doubt that you compared clustered.mmf(A) reestimated by HERest with un-reestimated clustered.mmf(B).

Best regards,

Heiga ZEN (Byung Ha CHUN)

--
------------------------------------------------
Heiga ZEN     (in Japanese pronunciation)
Byung Ha CHUN (in Korean pronunciation)

Department of Computer Science and Engineering
Nagoya Institute of Technology
Gokiso-cho, Showa-ku, Nagoya 466-8555 Japan

http://www.sp.nitech.ac.jp/~zen
------------------------------------------------