For the record.  To get tuning to work for a tree to tstring model
system and the ems do the following:
Under the [TUNING] module in the configuration file insert:
tuning-settings = "--mertdir $moses-src-dir/mert --inputtype 3"

Otherwise inputtype defaults to 0 in mert-moses.pl.

John




On 10/27/10, John Morgan <[email protected]> wrote:
> On 10/27/10, Hieu Hoang <[email protected]> wrote:
>> err, error correction on the iphone is rubbish!
>>
>> i meant to say, 'you may have given the ems the ini file and told it to
>> skip tuning'
>>
>> On 27/10/2010 16:56, Hieu Hoang wrote:
>>> Not sure. YouTube have given ema the ini file&  told it to skip tuning
>>>
>>> Hieu
>>> Sent from my flying horse
>>>
>>> On 27 Oct 2010, at 15:29, John Morgan<[email protected]>
>>> wrote:
>>>
>>>> Hieu,
>>>> Thanks for the references.
>>>> I'm working on an English to Pashto project right now with very little
>>>> training data (~35k segments).
>>>> I'm getting good results with the hierarchical models, but I'd like to
>>>> try all the options available in the ems.
>>>>
>>>> For my first tree2string setup I set the input parser to the collins
>>>> parser.
>>>> I set hierarchical-rule-set to true.
>>>> I ignored tuning.
>>>> The results I got were not as good as a hierarchical system, but
>>>> better than a phrase-based system under similar conditions.  I'm
>>>> rerunning this experiment to verify the results.
>>>> I run into problems when I introduce tuning.  I've modified my
>>>> experiment.meta file and it's possible my problems come from there.
>>>> My guess right now is that if tuning and hierarchical-rule-set are
>>>> both set in the configuration file, there is no route through the
>>>> experiment.meta file.
>>>> But why does a tree2string system work without tuning?
>>>>
>>>> John
>>>>
>>>>
>>>>
>>>> On 10/27/10, Hieu Hoang<[email protected]>  wrote:
>>>>> hey john
>>>>>
>>>>> i'm interested in how your tree-to-string model gets on as i'm looking
>>>>> @
>>>>> these models myself.
>>>>>
>>>>> are you doing english-foreign ? and you're using an english parser?
>>>>>
>>>>> fyi, what you're doing is unlikely to be better than hierarchical
>>>>> model.
>>>>> syntax constraints are usually too strict.
>>>>>
>>>>> Liang Huang's paper that did this used a very small test corpora.
>>>>>     http://www.cis.upenn.edu/~lhuang3/amta06-sdtedl.pdf
>>>>> imho, to get them to work, u need to a forest
>>>>>     http://www.cis.upenn.edu/~lhuang3/forest-based-translation.pdf
>>>>> or to relax the syntax
>>>>>     http://aclweb.org/anthology-new/W/W10/W10-1761.pdf
>>>>>
>>>>> On 26/10/2010 19:54, John Morgan wrote:
>>>>>> Hieu,
>>>>>> According to the TUNING_tune.?.STDERR file it looks like tuning
>>>>>> succeeded and a moses.ini.? file was generated under the tuning
>>>>>> directory.
>>>>>>
>>>>>> I had hierarchical-rule-set = true set in my configuration file.  I
>>>>>> guess this is the cause of my problems, right?
>>>>>> I'm trying to build a tree to string system.
>>>>>> Thanks,
>>>>>> John
>>>>>>
>>>>>>
>>>>>> On 10/26/10, Hieu Hoang<[email protected]>   wrote:
>>>>>>> hi john
>>>>>>>
>>>>>>> do you know what the error is? It's usually found in the log file
>>>>>>>     TUNING_tune.??.STDERR
>>>>>>> can you send me an example of your input&   a few files of the
>>>>>>> translation
>>>>>>> bmodel.
>>>>>>>
>>>>>>> using inputtype=0 (normal sentence) should be ok if it's a hiero
>>>>>>> model
>>>>>>> or
>>>>>>> a
>>>>>>> string-to-tree model. you only need inputtype=3 if you're doing a
>>>>>>> tree-to-string model
>>>>>>>
>>>>>>> On 26 October 2010 15:30, John Morgan<[email protected]>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> --
>>>>>>>> Regards,
>>>>>>>> John J Morgan
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Hello,
>>>>>>>> I'm trying to get an English to Pashto system working with
>>>>>>>> syntax-directed models in the ems framework.
>>>>>>>> Tuning with mert fails.
>>>>>>>> I think I've tracked down the problem to the commandline option
>>>>>>>> inputtype = 0.  inputtype is set to 3 in the moses.ini, but the
>>>>>>>> automatically generated tuning script sets inputtype to 0 for the
>>>>>>>> moses_chart decoder.  I set inputtype = 3 under the [TUNING] module
>>>>>>>> in
>>>>>>>> my configuration file  and mert tuning runs for 8 iterations then
>>>>>>>> crashes.
>>>>>>>> What is the correct setting for inputtype when training
>>>>>>>> syntax-directed
>>>>>>>> models?
>>>>>>>>
>>>>>>>> I'm successful if I stay away from mert tuning.
>>>>>>>> Thanks,
>>>>>>>> _______________________________________________
>>>>>>>> Moses-support mailing list
>>>>>>>> [email protected]
>>>>>>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>>>>>>
>>>>>>>>
>>>>
>>>> --
>>>> Regards,
>>>> John J Morgan
>>
>
>
>
>
>
>
>
> I ran the decoding step by hand using the commandline automatically
> generated by experiment.perl.
> I can get the decoder to translate by deleting the -v 0 option to
> moses_chart.
>


-- 
Regards,
John J Morgan
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to