You're right, there is a difference is model scores and output when using
the binary or memory phrase-table.

This happens when I'm using Moses v1, but not if I'm using the current
Moses in github.

Your lexical reordering models are fairly strange because they depend on a
source factor that only 1 phrase-table has.

So you want to use only 1 lex. reordering model for each phrase-table? I'm
not sure if Moses is working as you intend. I'm not sure what it's doing
either!

However, I think it's an interesting experiment so if you want to pursue
this and you're willing to do some coding, I can help you implement this.





On 12 February 2014 14:28, Mahmoud Ghoneim <[email protected]> wrote:

> Thanks Hieu for your reply.
> I changed my email address and I needed to register the new one. I did.
>
> I ran it on my machines without LMs and without all the command line
> arguments and again I got two different outputs!
> The attached file contains the two outputs.
>
> Also if "both lexical reordering models should be used whenever a
> hypothesis is created, by either translation model", then the text-based
> model is not working as supposed. Would you please review the verbose files
> I sent with the previous email. (I mentioned the exact lines that shows
> different behaviors).
>
> Is there a possibility that I have a corrupt setup on my machine? How to
> double check?
>
> Thanks,
> Mahmoud Ghoneim, PhD
> Post Doctoral Research Scientist
> Computer Science Department
> School of Engineering and Applied Science
> The George Washington University
>
>
> On Wed, Feb 12, 2014 at 7:30 AM, Hieu Hoang <[email protected]> wrote:
>
>>  Hi Mahmoud
>>
>> To post to the mailing list, please subscribe to it first. You can
>> subscribe here
>>    http://mailman.mit.edu/mailman/listinfo/moses-support
>> Also, if you need to post large datafiles, you should place the files
>> somewhere for people to download it, and give people the URL. Don't post
>> very large datafiles. I use google drive
>>
>> I've looked at your problem. I ran it without LMs and without all the
>> command line arguments.
>>    $MOSES_DIR/bin/moses -config Sent15-FilteredModel/moses.ini
>> -input-file sent151.txt
>>    $MOSES_DIR/bin/moses -config
>> Sent15-FilteredModel/BinaryTables/moses.bin.ini -input-file sent151.txt
>>
>> They gave me exactly the same answer and model scores. So I'm not sure
>> what the problem is.
>>
>> Remember - both lexical reordering models are used whenever a hypothesis
>> is created, by either translation model.
>>
>>
>>
>> On 11 February 2014 19:15, <[email protected]> wrote:
>>
>>> As list administrator, your authorization is requested for the
>>> following mailing list posting:
>>>
>>>     List:    [email protected]
>>>     From:    [email protected]
>>>     Subject: I think moses has a bug when dealing with binarized
>>> factored models
>>>     Reason:  Post by non-member to a members-only list
>>>
>>> At your convenience, visit:
>>>
>>>     http://mailman.mit.edu/mailman/admindb/moses-support
>>>
>>> to approve or deny the request.
>>>
>>>
>>> ---------- Forwarded message ----------
>>> From: Mahmoud Ghoneim <[email protected]>
>>> To: [email protected]
>>> Cc:
>>> Date: Tue, 11 Feb 2014 14:15:17 -0500
>>> Subject: I think moses has a bug when dealing with binarized factored
>>> models
>>> Hi moses-support team,
>>> I have an Ar-En factored translation model (Ar: lexeme|lemma|POS to En:
>>> lexeme|lemma|POS) with two translation paths (0-0 and 1,2-0). When I use
>>> the text version of the model and the binarized version of the same exact
>>> model, I get two totally different outputs!!
>>>
>>> I chose an input sentence and I generated the detailed verbose (v=3)
>>> logging and compared the results for the two cases and it seems that the
>>> decoder is not calculating the hypothesis's scores in a correct way when it
>>> deals with binarized version of a factored model !! The problem is
>>> generated due to considering all factors (rather than only the specific
>>> decoding-path factors) when calculating the lexical-reordering scores
>>> (explained in details below)
>>>
>>> Would you please review the attached files and confirm on my conclusions
>>> or give me the proper explanation of this behavior?
>>>
>>> I am attaching a 'tgz' file
>>>
>>
>>
>>> containing the following:
>>>  1- the input sentence (sent151.txt)
>>> 2- the filtered text model (in subDirectory 'Sent15-FilteredModel')
>>> 3- the binarized version of the filtered text model (in subDirectory
>>> 'Sent15-FilteredModel/BinaryTables')
>>> 4- the output translation using the text model (sent151.txt.out-mosesR1)
>>> 5- the output translation using the binarized model
>>> (sent151.txt.out-mosesR1-filtered-bin)
>>> 6- the verbose logging files for both cases. ()
>>> 7- the commands used to run moses. (command.note)
>>>
>>> Examples of the problem:
>>> - If you review the creation of hypothesis 6 from 0 (at line 2747 in
>>> text model verbose and at line 2743 in the binarized model verbose), you
>>> find moses considers the three factors while calculating the reordering
>>> scores, (while it is supposed to consider factor0 only as this hypothesis
>>> is form the first decoding-path 0-0)
>>> - Also, if you review the creation of hypothesis 61 from 0 for the text
>>> model which corresponds to hypothesis 64 for the binarized model, you will
>>> find again that the lexical-reordering scores for decoding-path 0-0 is
>>> included while this hypothesis is generated from the decoding-path 1,2-0!!
>>>
>>> Thanks in advance for your help,
>>> Mahmoud Ghoneim, PhD
>>> Post Doctoral Research Scientist
>>> Computer Science Department
>>> School of Engineering and Applied Science
>>> The George Washington University
>>>
>>>
>>> ---------- Forwarded message ----------
>>> From: [email protected]
>>> To:
>>> Cc:
>>> Date:
>>> Subject: confirm c4f798e58a24737bbfd9f76a5540c9d2c4b39cdd
>>> If you reply to this message, keeping the Subject: header intact,
>>> Mailman will discard the held message.  Do this if the message is
>>> spam.  If you reply to this message and include an Approved: header
>>> with the list password in it, the message will be approved for posting
>>> to the list.  The Approved: header can also appear in the first line
>>> of the body of the reply.
>>>
>>
>>
>>
>> --
>> Hieu Hoang
>> Research Associate
>> University of Edinburgh
>> http://www.hoang.co.uk/hieu
>>
>>
>


-- 
Hieu Hoang
Research Associate
University of Edinburgh
http://www.hoang.co.uk/hieu
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to