Hi again Hieu,

I didn't mention that filtering within mert and filtering test set for  
translation was different in one thing: the latter used -Binarizer  
option. The value I passed for it was  
"$MOSES_ROOTDIR/dist/bin/CreateOnDisk 1 1 5 100 2" according to  
recommendation found.

However, I just found that the problem was caused by this very option.  
I removed it and did only pure filtering and it was solved. Did you  
mean using this option by creating on-disk rule table? Fortunately  
(and also unfortunately) the rule table without relaxtion was so small  
that none of these steps were required: 13 tuning steps in less than  
an hour!

Best Regards,
Rasul.

Quoting Hieu Hoang <[email protected]>:

> yep, i'm not sure if the filtering script has been tested with
> tree-to-string models so it may not work properly.
>
> if you're trying to minimize memory usage, you could use the on-disk
> rule-table. That works with tree-to-string
>
> On 03/02/2012 18:04, [email protected] wrote:
>> Hi Hieu,
>>
>> Thanks for reply. I used the option for glue rules in training and   
>> -Hierarchical for filtering inside mert, and it solved the problem.
>>
>> However, I faced some puzzling results: the BLEU score shown in   
>> tuning/moses.ini was normal. But when I filtered the rule table   
>> separately and passed it to moses-chart, the BLEU score was   
>> strangely low. Therefore, I decided to test with original   
>> (non-filtered) rule table, and it worked. What is surprising for me  
>>  here is that both decoding procedures, inside mert and final   
>> translation, used filtered tables, but produced very different   
>> results. Although filtering for each of them was based on different  
>>  input (development and test sets), these inputs were similar and   
>> randomly chosen from a single data set. I have not yet been able to  
>>  figure it out.
>>
>> Thanks again.
>>
>> Best Regards,
>> Rasul.
>>
>> Quoting Hieu Hoang <[email protected]>:
>>
>>> hi razul
>>>
>>> 1. yes, you need the glue rules. For tree-to-string models, they are the
>>> same as for hierachical models. If the script produces anything different,
>>> then it's probably wrong
>>> 2. Probably yes. -Hierarchical is for syntactic models generally but no-one
>>> has had time to change the name of the argument.
>>>
>>> Not a lot of people are investigating tree-to-string models so some of the
>>> scripts may work as intended. Double check your data at each step. If you
>>> manage to fix scripts to work with tree-to-string models, please recommit
>>> them to github
>>>
>>>
>>> On Fri, Feb 3, 2012 at 8:59 AM, <[email protected]> wrote:
>>>
>>>> Hi all,
>>>>
>>>> I'm training a syntax-based tree-to-string model. I appreciate your
>>>> answers to these questions:
>>>>
>>>> - Do I need to use --glue-grammar option when training as in
>>>> hierarchical models?
>>>> - Do I need to use -Hierarchical option in passing --filtercmd option
>>>> to mert? The name implies that it should be used for hierarchical
>>>> models. However, when dropping it, segmentation fault occurs.
>>>>
>>>> Thank you very much in advance.
>>>>
>>>> Best Regards,
>>>> Rasul.
>>>> _______________________________________________
>>>> Moses-support mailing list
>>>> [email protected]
>>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>>
>>>
>>
>>
>>



_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to