[hts-users:02806] Re: About state clustering

[那兴宇 <nxy-yzqs@xxxxxxx>]

Hi, Kei

It turns out that I only changed the -q of duration other than the composed features...
I checked the resulting duration tree and it is the expected format.
I rechecked the perl script and remake voice, the training process goes well till now. The first clustering has been finished and the trees are correct. I think it will work.

Thank Kei and Oura for your help.
 

At 2011-04-21 23:36:23，"Kei Hashimoto" <bonanza@xxxxxxxxxxxxxxx> wrote:

>Hi,
>
>I think your procedures are correct.
>Please confirm your edit file "question-all.hed" and options.
>
>
>(2011/04/21 12:44), 那兴宇 wrote:
>> Hi, Kei&  Oura
>>
>> I tested -q option with value 0 and 1, but it didn't work. Let me discribe my purpose and may you check if my procedure is correct please.
>>
>> My request:
>> In my question set, there is a question that i always wanted to set as the root node question for the tree of every state of every feature. I call it the "root-node question". My expected tree is like this:
>> ========this is the expected tree-mgc.inf for example========
>> QS  ROOT_QUEST      { "*+x#*" }
>> QS  other_quests........
>> .......
>>
>> {*}[2]
>> {
>>     0  ROOT_QUEST       -2          -1
>>    -1  some_quest            -3          -7
>>   ...............
>> }
>>
>>
>> {*}[3]
>> {
>>     0  ROOT_QUEST       -2          -1
>>    -1  some_quest            -4          -6
>>   ...............
>> }
>> ......
>>
>> ================== end of the tree file ==================
>>
>> What I did:
>> I created two question set files. One is with the name "question-top.hed" which includes only one question that I selected. The other one is with the name "question-all.hed" which has all the questions for training including the "root-node question"
>>
>> First, I use HHEd with "question-top.hed" to do an initial clustering. The resulting tree is saved in top-mgc.inf. All the trees has only two leafs, eg." 0   TOP_QUEST   "mgc_s2_2"   "mgc_s2_1"  ". Second, I used HHEd with "question-all.hed" and -q option. In this step, I added "LT /tree/cmp/top-mgc.inf" before the TB command, which is my reference tree.
>>
>> What I got:
>> The expected result is that the cluster start from the specified root question and continues as it always did. However, the resulting tree-mgc.inf is the same as the reference/initial tree, no matter  what -q value I set.
>>
>> I think my usage is not correct, but don't know why. Please teach me.
>>
>> At 2011-04-21 03:51:30，"Kei Hashimoto"<bonanza@xxxxxxxxxxxxxxx>  wrote:
>>
>>> Hi,
>>>
>>> The -q option is used to set the question in clustering.
>>> When you set the -q value to 0 or 1, you can create larger tree than the reference tree.
>>> When the -q value is set to 1, the standard clustering will be performed after the same tree as the reference tree is created.
>>> When the -q value is set to 0, the question which the gain of likelihood is less than threshold will be not used even if the reference tree is set.
>>>
>>> When you set the -q value to 2 or 3, you cannot create larger tree than the reference tree.
>>> And, if you want to create the same tree as the reference tree even though the gain of likelihood is less than threshold, you should set the -q value to 3.
>>> It will work as like IT command in HHEd.
>>> When the -q value is set to 2, it will work as like back-off according to the threshold.
>>>
>>>
>>> Regards,
>>> Kei
>>>
>>>
>>> (2011/04/21 1:22), 那兴宇 wrote:
>>>> Hi,
>>>>
>>>>
>>>> Thanks for answering, Kei.
>>>>
>>>> I don't understand the -q values very clearly. I can tell that they stop splitting according to distance threshold or occ. But I cannot understand the explanation of "standard cluster will not be performed". And why the occ threshold is said to be 0 other than the values commonly used?
>>>>
>>>> And what is the relationship between the reference tree and the -q values. I will try 0 and 1 tomorrow and see what happens.
>>>>
>>>>
>>>>
>>>> At 2011-04-21 00:09:18，"Kei Hashimoto"<bonanza@xxxxxxxxxxxxxxx>   wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I think you should set the -q value to 0 or 1.
>>>>> The difference between 0 and 1 is the stopping criterion when reference trees are used.
>>>>> When the value is set to 0, clustering is stopped even though you use reference trees.
>>>>>
>>>>> Regards,
>>>>> Kei
>>>>>
>>>>>
>>>>> (2011/04/21 0:47), 那兴宇 wrote:
>>>>>> Hi,
>>>>>>
>>>>>>
>>>>>> I tried HHEd in HTS-2.2beta in this way:
>>>>>>
>>>>>> 1. Use a question set including only one question that I wanted to set as the root node to cluster a initial tree, which I got:
>>>>>>
>>>>>> QS A0_0_TYPE=EN { ...... }
>>>>>>
>>>>>> {*}[2]
>>>>>>
>>>>>> {
>>>>>>
>>>>>>      0  A0_0_TYPE=EN   "mgc_s2_2"   "mgc_s2_1"
>>>>>>
>>>>>> }
>>>>>>
>>>>>> ......
>>>>>>
>>>>>> This is the resulting tree file.
>>>>>>
>>>>>> 2. Then I tried HHEd with -q option and a full version of question set. I added a LT command before TB to load the initial tree. And the -q value is set to 2. Then I make voice.
>>>>>>
>>>>>> However, the resulting tree was exactly the same as the initial tree above. Do I used the -q option in a wrong way or the value is not appropriate?
>>>>>> The top node question is not included in the full question set.
>>>>>>
>>>>>> Please inform me about how to use the -q option of HHEd. Thanks!
>>>>>>
>>>>>>
>>>>>> At 2011-04-19 23:03:24，"Keiichiro Oura"<uratec@xxxxxxxxxxxxxxx>    wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> HHEd in HTS-2.2 beta can do it with '-q' option.
>>>>>>> Before TB command, LT command can be used.
>>>>>>>
>>>>>>> Regards,
>>>>>>> Keiichiro Oura
>>>>>>>
>>>>>>>
>>>>>>> 2011/4/19 那兴宇<nxy-yzqs@xxxxxxx>:
>>>>>>>> Hi, all
>>>>>>>>
>>>>>>>> I want to set a specified question to be used in the root node of the state
>>>>>>>> decision tree. Which edit command of HHEd should I use? If there is none,
>>>>>>>> how should I do with the current version to specify a root node question?
>>>>>>>> Please give some suggestions, thanks!
>>>>>>>> --
>>>>>>>> Na Xingyu
>>>>>>>> Beijing Institute of Technology
>>>>>>>>
>>>>>>>> ________________________________
>>>>>>>> 体验网易邮箱2G超大附件，轻松发优质大电影、大照片，提速3倍!
>>>>>>>
>>>>>
>>>>>
>>>>>
>>>
>>> --
>>> -------------------------------
>>> Nagoya Institute of Technology
>>> Tokuda and Lee lab.
>>> Kei Hashimoto
>>> bonanza@xxxxxxxxxxxxxxx
>>> -------------------------------
>>>
>>
>>
>>
>
>-- 
>-------------------------------
>Nagoya Institute of Technology
>Tokuda and Lee lab.
>Kei Hashimoto
>bonanza@xxxxxxxxxxxxxxx
>-------------------------------
>