Earlier this year, I had a different alignment problem but with -print-alignment-info. I never had time to troubleshoot or share my model. The model was the same sample data we have used for years. There were two differences from our traditional usage. 1) I changed to the compact phrase & reordering tables, and 2) I added the -print-alignment-info flag to the Moses command line. Here's the thread from mid-March 2015.

Tom


From: Tom Hoar <[email protected]<http://gmane.org/get-address.php?address=tahoar%2dcB%2fihCqopeG7AKbOTpHwwjx6VEFkrI8ampATvIKMPHk%40public.gmane.org>> Subject: Re: -print-alignment-info option<http://news.gmane.org/find-root.php?message_id=55090DAB.8000500%40precisiontranslationtools.com>
To: [email protected]
Date: 2015-03-18 05:31:23 GMT

Hieu,

Yes. Without the "-print-alignment-info" argument, moses outputs 5 expected translations from a small input file. With the "-print-alignment-info" argument, the 1st, 3rd and 5th expected translations are fine. The 2nd and 4th are zero-length strings. Unexpectedly strange results, indeed. Since "-print-alignment-info" was not in R-1, this is my first time using it.

The model is very small from a 15,000 pair NL-EN example corpus we've used for years for testing. The model is small. Phrase & reordering tables are compact format. The LM is trained by lmplz, binarized in KenLM format and loaded with "9" option.

Let me put together a package with examples and get back to you. The package will include the .gz files, compact & binary files, the moses.ini file, the input file, the command lines I use and the corresponding outputs. I'll have that ready early next week.

Tom


On 03/16/2015 11:58 PM, Hieu Hoang wrote:
You're saying it gives a different translation depending on whether -print-alignment-info is used or not? That would be a very strange behaviour. Which phrase-table are you using?

If you can share your model files, i'll take a look at it.

Hieu Hoang
Research Associate (until March 2015)
University of Edinburgh
http://www.hoang.co.uk/hieu

On 16 March 2015 at 02:53, Tom Hoar <tahoar-cB/[email protected] <mailto:tahoar-cB/[email protected]>> wrote:

    This is my first time using the '-print-alignment-info' option and
    I'm getting some unexpected results.

    In a set of five sentences, when using no option or
    '-report-segmentation' or '-report-segmentation-enriched', all
    five sentence translations are output with respective alignment
    info. With the '-print-alignment-info' option, two sentence
    translations are output as zero-length strings.

    Is this what is meant when the help prompt says: "Word-to-word
    alignments are taken from the phrase table if any"?

     When no word-to-word alignment info is in the phrase table, I
    would expect the translated output followed by " ||| " and then
    nothing.

    Thanks.

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

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
Moses-support@... <http://gmane.org/get-address.php?address=Moses%2dsupport%2d3s7WtUTddSA%40public.gmane.org>
http://mailman.mit.edu/mailman/listinfo/moses-support


On 11/1/2015 6:41 PM, [email protected] wrote:
Message: 3
Date: Sun, 1 Nov 2015 15:11:23 +0330
From: Arefeh Kazemi<[email protected]>
Subject: Re: [Moses-support] mismatch in alignment file when using
        -alignment-output-file
To: Hieu Hoang<[email protected]>,[email protected]
Cc:[email protected]

Hi Hieu

Thanks for the reply.
my original files are so huge so I attached a toy model which the mismatch
happens for it too.

Thanks again.
Arefeh?
  toyCorpus.zip
<https://drive.google.com/a/dcu.ie/file/d/0B37sY2C6IhcjS0dSSnA2NjVBTnc/view?usp=drive_web>

  toyLM.zip
<https://drive.google.com/a/dcu.ie/file/d/0B37sY2C6IhcjM2xiVkVYaDY3ZFk/view?usp=drive_web>

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

Reply via email to