Moses-support Digest, Vol 107, Issue 35

Send Moses-support mailing list submissions to
moses-support@mit.edu

To subscribe or unsubscribe via the World Wide Web, visit
http://mailman.mit.edu/mailman/listinfo/moses-support
or, via email, send a message with subject or body 'help' to
moses-support-request@mit.edu

You can reach the person managing the list at
moses-support-owner@mit.edu

When replying, please edit your Subject line so it is more specific
than "Re: Contents of Moses-support digest..."


Today's Topics:

1. Re: Performance issue with Neural LM for English-Hindi SMT
(Rajnath Patel)
2. Re: Performance issue with Neural LM for English-Hindi SMT
(Raj Dabre)
3. Re: Performance issue with Neural LM for English-Hindi SMT
(Rajnath Patel)


----------------------------------------------------------------------

Message: 1
Date: Mon, 14 Sep 2015 10:24:52 +0530
From: Rajnath Patel <patelrajnath@gmail.com>
Subject: Re: [Moses-support] Performance issue with Neural LM for
English-Hindi SMT
To: moses-support <moses-support@mit.edu>
Message-ID:
<CAE-r4u=bqjWqKG5+46v1RRyYPfSioa5OQ19q+TDek-FjFLY_ug@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Thanks for quick response.

@Raj Dabre
Corpus statistics as follows-
Approx -65k sentences, 1200k words, 50k vocab.
Please suggest, what size of corpus is enough for neural LM training?

@Riko
I will try with development set and more epochs as you suggested. Back-off
LM you mean fall back to neural LM if its not found in n-gram model(Please
correct if I got it wrong). If so, could you please suggest how to
configure the same with moses.

Thanks.



> Message: 1
> Date: Mon, 14 Sep 2015 01:56:14 +0900
> From: Raj Dabre <prajdabre@gmail.com>
> Subject: Re: [Moses-support] Performance issue with Neural LM for
> English-Hindi SMT
> To: Rajnath Patel <patelrajnath@gmail.com>
> Cc: moses-support <moses-support@mit.edu>
> Message-ID:
> <CAB3gfjCGapWtYTheh6mKHhica7v7d=
> q81iA5L7jiZS4kKjudfA@mail.gmail.com>
> Content-Type: text/plain; charset="utf-8"
>
> Hi,
> I have had a similar experience with NPLM.
> Do you perhaps have a small corpus?
>
> On Sun, Sep 13, 2015 at 6:51 PM, Rajnath Patel <patelrajnath@gmail.com>
> wrote:
>
> > Hi all,
> >
> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
> > translation quality is kind of not good as compared to n-gram LM(scores
> are
> > given below). I have trained LM for 3-gram and 5-gram with default
> > setting(as mentioned on statmt.org/moses). Kindly suggest, If some one
> > has tried the same English-Hindi SMT and got improved results. What may
> be
> > probable cause of degraded results?
> >
> > BLEU scores:
> > n-gram(5-gram)=24.40
> > neural-lm(5-gram)=11.30
> > neural-lm(3-gram)=12.10
> >
> > Thank you.
> >
> > --
> > Regards:
> > Raj Nath Patel
> >
> > _______________________________________________
> > Moses-support mailing list
> > Moses-support@mit.edu
> > http://mailman.mit.edu/mailman/listinfo/moses-support
> >
> >
>
>
> --
> Raj Dabre.
> Doctoral Student,
> Graduate School of Informatics,
> Kyoto University.
> CSE MTech, IITB., 2011-2014
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL:
> http://mailman.mit.edu/mailman/private/moses-support/attachments/20150913/7fa15fdd/attachment-0001.html
>
> ------------------------------
>
> Message: 2
> Date: Sun, 13 Sep 2015 23:19:19 +0100
> From: Rico Sennrich <rico.sennrich@gmx.ch>
> Subject: Re: [Moses-support] Performance issue with Neural LM for
> English-Hindi SMT
> To: moses-support@mit.edu
> Message-ID: <55F5F667.9030704@gmx.ch>
> Content-Type: text/plain; charset="windows-1252"
>
> Hello Raj,
>
> Usually, nplm is used in addition to a back-off LM for best results.
> That being said, your results indicate that nplm is performing poorly.
> If you have little training data, a smaller vocabulary size and more
> training epochs may be appropriate. I would advise to provide a
> development set to the nplm training program so that you can track the
> training progress, and compare perplexity with back-off models.
>
> best wishes,
> Rico
>
> On 13/09/15 10:51, Rajnath Patel wrote:
> > Hi all,
> >
> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
> > translation quality is kind of not good as compared to n-gram
> > LM(scores are given below). I have trained LM for 3-gram and 5-gram
> > with default setting(as mentioned on statmt.org/moses
> > <http://statmt.org/moses>). Kindly suggest, If some one has tried the
> > same English-Hindi SMT and got improved results. What may be probable
> > cause of degraded results?
> >
> > BLEU scores:
> > n-gram(5-gram)=24.40
> > neural-lm(5-gram)=11.30
> > neural-lm(3-gram)=12.10
> >
> > Thank you.
> >
> > --
> > Regards:
> > Raj Nath Patel
>
>


--
Regards:
??? ??? ????/Raj Nath Patel
KBCS dept.
CDAC Mumbai.
http://kbcs.in/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20150914/b5c14648/attachment-0001.html

------------------------------

Message: 2
Date: Mon, 14 Sep 2015 14:53:47 +0900
From: Raj Dabre <prajdabre@gmail.com>
Subject: Re: [Moses-support] Performance issue with Neural LM for
English-Hindi SMT
To: Rajnath Patel <patelrajnath@gmail.com>
Cc: moses-support <moses-support@mit.edu>
Message-ID:
<CAB3gfjAWMg-MiueqU69oQqVGeGM1Ojh5TnHtU3Y1cJJswwwA-A@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Hi,

1. 65k is quite small. You might need many (Read: MANY) iterations till the
perplexity stops dropping by significant amounts.

2. In Moses, I think you can try this--- Add 2 lines as below:

Under *feature* add this: NeuralLM factor=0 name=LM1 order=5
path=<path/to/neural/lm/file>

Under *weight *add this: LM1=0.5

I am not 100% sure but it should work.




On Mon, Sep 14, 2015 at 1:54 PM, Rajnath Patel <patelrajnath@gmail.com>
wrote:

> Thanks for quick response.
>
> @Raj Dabre
> Corpus statistics as follows-
> Approx -65k sentences, 1200k words, 50k vocab.
> Please suggest, what size of corpus is enough for neural LM training?
>
> @Riko
> I will try with development set and more epochs as you suggested. Back-off
> LM you mean fall back to neural LM if its not found in n-gram model(Please
> correct if I got it wrong). If so, could you please suggest how to
> configure the same with moses.
>
> Thanks.
>
>
>
>> Message: 1
>> Date: Mon, 14 Sep 2015 01:56:14 +0900
>> From: Raj Dabre <prajdabre@gmail.com>
>> Subject: Re: [Moses-support] Performance issue with Neural LM for
>> English-Hindi SMT
>> To: Rajnath Patel <patelrajnath@gmail.com>
>> Cc: moses-support <moses-support@mit.edu>
>> Message-ID:
>> <CAB3gfjCGapWtYTheh6mKHhica7v7d=
>> q81iA5L7jiZS4kKjudfA@mail.gmail.com>
>> Content-Type: text/plain; charset="utf-8"
>>
>> Hi,
>> I have had a similar experience with NPLM.
>> Do you perhaps have a small corpus?
>>
>> On Sun, Sep 13, 2015 at 6:51 PM, Rajnath Patel <patelrajnath@gmail.com>
>> wrote:
>>
>> > Hi all,
>> >
>> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
>> > translation quality is kind of not good as compared to n-gram LM(scores
>> are
>> > given below). I have trained LM for 3-gram and 5-gram with default
>> > setting(as mentioned on statmt.org/moses). Kindly suggest, If some one
>> > has tried the same English-Hindi SMT and got improved results. What may
>> be
>> > probable cause of degraded results?
>> >
>> > BLEU scores:
>> > n-gram(5-gram)=24.40
>> > neural-lm(5-gram)=11.30
>> > neural-lm(3-gram)=12.10
>> >
>> > Thank you.
>> >
>> > --
>> > Regards:
>> > Raj Nath Patel
>> >
>> > _______________________________________________
>> > Moses-support mailing list
>> > Moses-support@mit.edu
>> > http://mailman.mit.edu/mailman/listinfo/moses-support
>> >
>> >
>>
>>
>> --
>> Raj Dabre.
>> Doctoral Student,
>> Graduate School of Informatics,
>> Kyoto University.
>> CSE MTech, IITB., 2011-2014
>> -------------- next part --------------
>> An HTML attachment was scrubbed...
>> URL:
>> http://mailman.mit.edu/mailman/private/moses-support/attachments/20150913/7fa15fdd/attachment-0001.html
>>
>> ------------------------------
>>
>> Message: 2
>> Date: Sun, 13 Sep 2015 23:19:19 +0100
>> From: Rico Sennrich <rico.sennrich@gmx.ch>
>> Subject: Re: [Moses-support] Performance issue with Neural LM for
>> English-Hindi SMT
>> To: moses-support@mit.edu
>> Message-ID: <55F5F667.9030704@gmx.ch>
>> Content-Type: text/plain; charset="windows-1252"
>>
>> Hello Raj,
>>
>> Usually, nplm is used in addition to a back-off LM for best results.
>> That being said, your results indicate that nplm is performing poorly.
>> If you have little training data, a smaller vocabulary size and more
>> training epochs may be appropriate. I would advise to provide a
>> development set to the nplm training program so that you can track the
>> training progress, and compare perplexity with back-off models.
>>
>> best wishes,
>> Rico
>>
>> On 13/09/15 10:51, Rajnath Patel wrote:
>> > Hi all,
>> >
>> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
>> > translation quality is kind of not good as compared to n-gram
>> > LM(scores are given below). I have trained LM for 3-gram and 5-gram
>> > with default setting(as mentioned on statmt.org/moses
>> > <http://statmt.org/moses>). Kindly suggest, If some one has tried the
>> > same English-Hindi SMT and got improved results. What may be probable
>> > cause of degraded results?
>> >
>> > BLEU scores:
>> > n-gram(5-gram)=24.40
>> > neural-lm(5-gram)=11.30
>> > neural-lm(3-gram)=12.10
>> >
>> > Thank you.
>> >
>> > --
>> > Regards:
>> > Raj Nath Patel
>>
>>
>
>
> --
> Regards:
> ??? ??? ????/Raj Nath Patel
> KBCS dept.
> CDAC Mumbai.
> http://kbcs.in/
>



--
Raj Dabre.
Doctoral Student,
Graduate School of Informatics,
Kyoto University.
CSE MTech, IITB., 2011-2014
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20150914/d43c24a4/attachment-0001.html

------------------------------

Message: 3
Date: Mon, 14 Sep 2015 11:38:14 +0530
From: Rajnath Patel <patelrajnath@gmail.com>
Subject: Re: [Moses-support] Performance issue with Neural LM for
English-Hindi SMT
To: Raj Dabre <prajdabre@gmail.com>
Cc: moses-support <moses-support@mit.edu>
Message-ID:
<CAE-r4un=3eA=WcNui0hatXbdm7Ppr0FMgvAkPbekV6SbvcPRBQ@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Thanks,

Yes configuration for neural LM is as you mentioned, but I was asking for
back-off Language model decoding, mentioned by Rico in his response.

--
Regards:
Raj Nath


On Mon, Sep 14, 2015 at 11:23 AM, Raj Dabre <prajdabre@gmail.com> wrote:

> Hi,
>
> 1. 65k is quite small. You might need many (Read: MANY) iterations till
> the perplexity stops dropping by significant amounts.
>
> 2. In Moses, I think you can try this--- Add 2 lines as below:
>
> Under *feature* add this: NeuralLM factor=0 name=LM1 order=5
> path=<path/to/neural/lm/file>
>
> Under *weight *add this: LM1=0.5
>
> I am not 100% sure but it should work.
>
>
>
>
> On Mon, Sep 14, 2015 at 1:54 PM, Rajnath Patel <patelrajnath@gmail.com>
> wrote:
>
>> Thanks for quick response.
>>
>> @Raj Dabre
>> Corpus statistics as follows-
>> Approx -65k sentences, 1200k words, 50k vocab.
>> Please suggest, what size of corpus is enough for neural LM training?
>>
>> @Riko
>> I will try with development set and more epochs as you suggested.
>> Back-off LM you mean fall back to neural LM if its not found in n-gram
>> model(Please correct if I got it wrong). If so, could you please suggest
>> how to configure the same with moses.
>>
>> Thanks.
>>
>>
>>
>>> Message: 1
>>> Date: Mon, 14 Sep 2015 01:56:14 +0900
>>> From: Raj Dabre <prajdabre@gmail.com>
>>> Subject: Re: [Moses-support] Performance issue with Neural LM for
>>> English-Hindi SMT
>>> To: Rajnath Patel <patelrajnath@gmail.com>
>>> Cc: moses-support <moses-support@mit.edu>
>>> Message-ID:
>>> <CAB3gfjCGapWtYTheh6mKHhica7v7d=
>>> q81iA5L7jiZS4kKjudfA@mail.gmail.com>
>>> Content-Type: text/plain; charset="utf-8"
>>>
>>> Hi,
>>> I have had a similar experience with NPLM.
>>> Do you perhaps have a small corpus?
>>>
>>> On Sun, Sep 13, 2015 at 6:51 PM, Rajnath Patel <patelrajnath@gmail.com>
>>> wrote:
>>>
>>> > Hi all,
>>> >
>>> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
>>> > translation quality is kind of not good as compared to n-gram
>>> LM(scores are
>>> > given below). I have trained LM for 3-gram and 5-gram with default
>>> > setting(as mentioned on statmt.org/moses). Kindly suggest, If some one
>>> > has tried the same English-Hindi SMT and got improved results. What
>>> may be
>>> > probable cause of degraded results?
>>> >
>>> > BLEU scores:
>>> > n-gram(5-gram)=24.40
>>> > neural-lm(5-gram)=11.30
>>> > neural-lm(3-gram)=12.10
>>> >
>>> > Thank you.
>>> >
>>> > --
>>> > Regards:
>>> > Raj Nath Patel
>>> >
>>> > _______________________________________________
>>> > Moses-support mailing list
>>> > Moses-support@mit.edu
>>> > http://mailman.mit.edu/mailman/listinfo/moses-support
>>> >
>>> >
>>>
>>>
>>> --
>>> Raj Dabre.
>>> Doctoral Student,
>>> Graduate School of Informatics,
>>> Kyoto University.
>>> CSE MTech, IITB., 2011-2014
>>> -------------- next part --------------
>>> An HTML attachment was scrubbed...
>>> URL:
>>> http://mailman.mit.edu/mailman/private/moses-support/attachments/20150913/7fa15fdd/attachment-0001.html
>>>
>>> ------------------------------
>>>
>>> Message: 2
>>> Date: Sun, 13 Sep 2015 23:19:19 +0100
>>> From: Rico Sennrich <rico.sennrich@gmx.ch>
>>> Subject: Re: [Moses-support] Performance issue with Neural LM for
>>> English-Hindi SMT
>>> To: moses-support@mit.edu
>>> Message-ID: <55F5F667.9030704@gmx.ch>
>>> Content-Type: text/plain; charset="windows-1252"
>>>
>>> Hello Raj,
>>>
>>> Usually, nplm is used in addition to a back-off LM for best results.
>>> That being said, your results indicate that nplm is performing poorly.
>>> If you have little training data, a smaller vocabulary size and more
>>> training epochs may be appropriate. I would advise to provide a
>>> development set to the nplm training program so that you can track the
>>> training progress, and compare perplexity with back-off models.
>>>
>>> best wishes,
>>> Rico
>>>
>>> On 13/09/15 10:51, Rajnath Patel wrote:
>>> > Hi all,
>>> >
>>> > I have tried Neural LM(nplm) with phrase based English-Hindi SMT, but
>>> > translation quality is kind of not good as compared to n-gram
>>> > LM(scores are given below). I have trained LM for 3-gram and 5-gram
>>> > with default setting(as mentioned on statmt.org/moses
>>> > <http://statmt.org/moses>). Kindly suggest, If some one has tried the
>>> > same English-Hindi SMT and got improved results. What may be probable
>>> > cause of degraded results?
>>> >
>>> > BLEU scores:
>>> > n-gram(5-gram)=24.40
>>> > neural-lm(5-gram)=11.30
>>> > neural-lm(3-gram)=12.10
>>> >
>>> > Thank you.
>>> >
>>> > --
>>> > Regards:
>>> > Raj Nath Patel
>>>
>>>
>>
>>
>> --
>> Regards:
>> Raj Nath Patel
>>
>
>
>
> --
> Raj Dabre.
> Doctoral Student,
> Graduate School of Informatics,
> Kyoto University.
> CSE MTech, IITB., 2011-2014
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20150914/1bb13263/attachment.html

------------------------------

_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support


End of Moses-support Digest, Vol 107, Issue 35
**********************************************

0 Response to "Moses-support Digest, Vol 107, Issue 35"

Post a Comment