Re: modernmt

[John Hewitt]

I found the reference for that 1,000,000 number a bit too late -- according
to this more recent paper from Koehn, it's more like 15,000,000 tokens for
NMT to meet phrase-based MT, and they omit syntax-based.

https://arxiv.org/pdf/1706.03872.pdf

-John

On Sun, Jul 2, 2017 at 12:38 PM, John Hewitt  wrote:

> I've talked with the ModernMT people; they're well aware that they're in a
> neural MT world, and they also know that there's a sizable market for
> non-neural MT solutions.
> To back this up -- Philipp Koehn gave a talk in March on comparing
> phrase-based, syntax-based, and neural MT in low-resource settings, that
> is, when the amount of bilingual text to train on is small.
>
> Neural MT needs (if I remember correctly) about 1,000,000 tokens of
> training data to outpace syntax-based MT.
> Many language pairs (and, for that matter, domains within a single
> language pair) do not meet that requirement, and in those cases
> syntax-based MT performs best.
>
> That being said, there are some cool opportunities to combine neural and
> syntax-based MT. I can't commit the work hours right now, necessarily, but
> I've worked with xnmt , an MIT-licensed
> neural MT library that is purpose-built to be highly modular. It may offer
> some good opportunities to make an ensemble system.
>
> On Sun, Jul 2, 2017 at 4:22 AM, Tommaso Teofili  > wrote:
>
>> I think it's interesting as it extends some features that also Joshua has,
>> it's open source and has good results compared with NMT.
>>
>> Tommaso
>>
>> Il giorno sab 1 lug 2017 alle ore 18:56 Suneel Marthi <
>> suneel.mar...@gmail.com> ha scritto:
>>
>> > Is this the latest/greatest paper around MT @tommaso ??
>> >
>> > On Sat, Jul 1, 2017 at 7:55 AM, Tommaso Teofili <
>> tommaso.teof...@gmail.com
>> > >
>> > wrote:
>> >
>> > > I accidentally found the paper about mmt [1]
>> > >
>> > > [1] :
>> > > https://ufal.mff.cuni.cz/eamt2017/user-project-product-
>> > > papers/papers/user/EAMT2017_paper_88.pdf
>> > >
>> > > Il giorno gio 1 dic 2016 alle ore 22:19 Mattmann, Chris A (3010) <
>> > > chris.a.mattm...@jpl.nasa.gov> ha scritto:
>> > >
>> > > > Guys I want to point you at the DARPA D3M program:
>> > > >
>> > > > http://www.darpa.mil/program/data-driven-discovery-of-models
>> > > >
>> > > > I’m part of the Government Team for the program. This will be a good
>> > > > connection
>> > > > to have b/c it’s focused on automatically doing model and code
>> building
>> > > > for ML based
>> > > > approaches.
>> > > >
>> > > >
>> > > > ++
>> > > > Chris Mattmann, Ph.D.
>> > > > Principal Data Scientist, Engineering Administrative Office (3010)
>> > > > Manager, Open Source Projects Formulation and Development Office
>> (8212)
>> > > > NASA Jet Propulsion Laboratory Pasadena, CA 91109 USA
>> > > > Office: 180-503E, Mailstop: 180-503
>> > > > Email: chris.a.mattm...@nasa.gov
>> > > > WWW:  http://sunset.usc.edu/~mattmann/
>> > > > ++
>> > > > Director, Information Retrieval and Data Science Group (IRDS)
>> > > > Adjunct Associate Professor, Computer Science Department
>> > > > University of Southern California, Los Angeles, CA 90089 USA
>> > > > WWW: http://irds.usc.edu/
>> > > > ++
>> > > >
>> > > >
>> > > > On 12/1/16, 1:15 PM, "Matt Post"  wrote:
>> > > >
>> > > > John,
>> > > >
>> > > > Thanks for sharing, this is really helpful. I didn't realize
>> that
>> > > > Marcello was involved.
>> > > >
>> > > > I think we can identify with the NMT danger. I still think there
>> > is a
>> > > > big niche that deep learning approaches won't reach for a few years,
>> > > until
>> > > > GPUs become super prevalent. Which is why I like ModernMT's
>> approaches,
>> > > > which overlap with many of the things I've been thinking. One thing
>> I
>> > > > really like is there automatic context-switching approach. This is a
>> > > great
>> > > > way to build general-purpose models, and I'd like to mimic it. I
>> have
>> > > some
>> > > > general ideas about how this should be implemented but am also
>> looking
>> > > into
>> > > > the literature here.
>> > > >
>> > > > matt
>> > > >
>> > > >
>> > > > > On Dec 1, 2016, at 1:46 PM, John Hewitt <
>> john...@seas.upenn.edu>
>> > > > wrote:
>> > > > >
>> > > > > I had a few good conversations over dinner with this team at
>> AMTA
>> > > in
>> > > > Austin
>> > > > > in October.
>> > > > > They seem to be in the interesting position where their work
>> is
>> > > > good, but
>> > > > > is in danger of being superseded by neural MT as they come
>> out of
>> > > > the gate.
>> > > > > Clearly, it has benefits over NMT, and is easier to adopt, but
>> > may
>> > > > not be
>> > > > > the winner over the long 

Re: modernmt

[John Hewitt]

I've talked with the ModernMT people; they're well aware that they're in a
neural MT world, and they also know that there's a sizable market for
non-neural MT solutions.
To back this up -- Philipp Koehn gave a talk in March on comparing
phrase-based, syntax-based, and neural MT in low-resource settings, that
is, when the amount of bilingual text to train on is small.

Neural MT needs (if I remember correctly) about 1,000,000 tokens of
training data to outpace syntax-based MT.
Many language pairs (and, for that matter, domains within a single language
pair) do not meet that requirement, and in those cases syntax-based MT
performs best.

That being said, there are some cool opportunities to combine neural and
syntax-based MT. I can't commit the work hours right now, necessarily, but
I've worked with xnmt , an MIT-licensed
neural MT library that is purpose-built to be highly modular. It may offer
some good opportunities to make an ensemble system.

On Sun, Jul 2, 2017 at 4:22 AM, Tommaso Teofili 
wrote:

> I think it's interesting as it extends some features that also Joshua has,
> it's open source and has good results compared with NMT.
>
> Tommaso
>
> Il giorno sab 1 lug 2017 alle ore 18:56 Suneel Marthi <
> suneel.mar...@gmail.com> ha scritto:
>
> > Is this the latest/greatest paper around MT @tommaso ??
> >
> > On Sat, Jul 1, 2017 at 7:55 AM, Tommaso Teofili <
> tommaso.teof...@gmail.com
> > >
> > wrote:
> >
> > > I accidentally found the paper about mmt [1]
> > >
> > > [1] :
> > > https://ufal.mff.cuni.cz/eamt2017/user-project-product-
> > > papers/papers/user/EAMT2017_paper_88.pdf
> > >
> > > Il giorno gio 1 dic 2016 alle ore 22:19 Mattmann, Chris A (3010) <
> > > chris.a.mattm...@jpl.nasa.gov> ha scritto:
> > >
> > > > Guys I want to point you at the DARPA D3M program:
> > > >
> > > > http://www.darpa.mil/program/data-driven-discovery-of-models
> > > >
> > > > I’m part of the Government Team for the program. This will be a good
> > > > connection
> > > > to have b/c it’s focused on automatically doing model and code
> building
> > > > for ML based
> > > > approaches.
> > > >
> > > >
> > > > ++
> > > > Chris Mattmann, Ph.D.
> > > > Principal Data Scientist, Engineering Administrative Office (3010)
> > > > Manager, Open Source Projects Formulation and Development Office
> (8212)
> > > > NASA Jet Propulsion Laboratory Pasadena, CA 91109 USA
> > > > Office: 180-503E, Mailstop: 180-503
> > > > Email: chris.a.mattm...@nasa.gov
> > > > WWW:  http://sunset.usc.edu/~mattmann/
> > > > ++
> > > > Director, Information Retrieval and Data Science Group (IRDS)
> > > > Adjunct Associate Professor, Computer Science Department
> > > > University of Southern California, Los Angeles, CA 90089 USA
> > > > WWW: http://irds.usc.edu/
> > > > ++
> > > >
> > > >
> > > > On 12/1/16, 1:15 PM, "Matt Post"  wrote:
> > > >
> > > > John,
> > > >
> > > > Thanks for sharing, this is really helpful. I didn't realize that
> > > > Marcello was involved.
> > > >
> > > > I think we can identify with the NMT danger. I still think there
> > is a
> > > > big niche that deep learning approaches won't reach for a few years,
> > > until
> > > > GPUs become super prevalent. Which is why I like ModernMT's
> approaches,
> > > > which overlap with many of the things I've been thinking. One thing I
> > > > really like is there automatic context-switching approach. This is a
> > > great
> > > > way to build general-purpose models, and I'd like to mimic it. I have
> > > some
> > > > general ideas about how this should be implemented but am also
> looking
> > > into
> > > > the literature here.
> > > >
> > > > matt
> > > >
> > > >
> > > > > On Dec 1, 2016, at 1:46 PM, John Hewitt <
> john...@seas.upenn.edu>
> > > > wrote:
> > > > >
> > > > > I had a few good conversations over dinner with this team at
> AMTA
> > > in
> > > > Austin
> > > > > in October.
> > > > > They seem to be in the interesting position where their work is
> > > > good, but
> > > > > is in danger of being superseded by neural MT as they come out
> of
> > > > the gate.
> > > > > Clearly, it has benefits over NMT, and is easier to adopt, but
> > may
> > > > not be
> > > > > the winner over the long run.
> > > > >
> > > > > Here's the link
> > > > > <
> > > > https://amtaweb.org/wp-content/uploads/2016/11/MMT_
> > > Tutorial_FedericoTrombetti_wide-cover.pdf
> > > > >
> > > > > to their AMTA tutorial.
> > > > >
> > > > > -John
> > > > >
> > > > > On Thu, Dec 1, 2016 at 10:17 AM, Mattmann, Chris A (3010) <
> > > > > chris.a.mattm...@jpl.nasa.gov> wrote:
> > > > >
> > > > >> Wow seems like this kind of overlaps with BigTranslate as
> 

Re: modernmt

[Tommaso Teofili]

I think it's interesting as it extends some features that also Joshua has,
it's open source and has good results compared with NMT.

Tommaso

Il giorno sab 1 lug 2017 alle ore 18:56 Suneel Marthi <
suneel.mar...@gmail.com> ha scritto:

> Is this the latest/greatest paper around MT @tommaso ??
>
> On Sat, Jul 1, 2017 at 7:55 AM, Tommaso Teofili  >
> wrote:
>
> > I accidentally found the paper about mmt [1]
> >
> > [1] :
> > https://ufal.mff.cuni.cz/eamt2017/user-project-product-
> > papers/papers/user/EAMT2017_paper_88.pdf
> >
> > Il giorno gio 1 dic 2016 alle ore 22:19 Mattmann, Chris A (3010) <
> > chris.a.mattm...@jpl.nasa.gov> ha scritto:
> >
> > > Guys I want to point you at the DARPA D3M program:
> > >
> > > http://www.darpa.mil/program/data-driven-discovery-of-models
> > >
> > > I’m part of the Government Team for the program. This will be a good
> > > connection
> > > to have b/c it’s focused on automatically doing model and code building
> > > for ML based
> > > approaches.
> > >
> > >
> > > ++
> > > Chris Mattmann, Ph.D.
> > > Principal Data Scientist, Engineering Administrative Office (3010)
> > > Manager, Open Source Projects Formulation and Development Office (8212)
> > > NASA Jet Propulsion Laboratory Pasadena, CA 91109 USA
> > > Office: 180-503E, Mailstop: 180-503
> > > Email: chris.a.mattm...@nasa.gov
> > > WWW:  http://sunset.usc.edu/~mattmann/
> > > ++
> > > Director, Information Retrieval and Data Science Group (IRDS)
> > > Adjunct Associate Professor, Computer Science Department
> > > University of Southern California, Los Angeles, CA 90089 USA
> > > WWW: http://irds.usc.edu/
> > > ++
> > >
> > >
> > > On 12/1/16, 1:15 PM, "Matt Post"  wrote:
> > >
> > > John,
> > >
> > > Thanks for sharing, this is really helpful. I didn't realize that
> > > Marcello was involved.
> > >
> > > I think we can identify with the NMT danger. I still think there
> is a
> > > big niche that deep learning approaches won't reach for a few years,
> > until
> > > GPUs become super prevalent. Which is why I like ModernMT's approaches,
> > > which overlap with many of the things I've been thinking. One thing I
> > > really like is there automatic context-switching approach. This is a
> > great
> > > way to build general-purpose models, and I'd like to mimic it. I have
> > some
> > > general ideas about how this should be implemented but am also looking
> > into
> > > the literature here.
> > >
> > > matt
> > >
> > >
> > > > On Dec 1, 2016, at 1:46 PM, John Hewitt 
> > > wrote:
> > > >
> > > > I had a few good conversations over dinner with this team at AMTA
> > in
> > > Austin
> > > > in October.
> > > > They seem to be in the interesting position where their work is
> > > good, but
> > > > is in danger of being superseded by neural MT as they come out of
> > > the gate.
> > > > Clearly, it has benefits over NMT, and is easier to adopt, but
> may
> > > not be
> > > > the winner over the long run.
> > > >
> > > > Here's the link
> > > > <
> > > https://amtaweb.org/wp-content/uploads/2016/11/MMT_
> > Tutorial_FedericoTrombetti_wide-cover.pdf
> > > >
> > > > to their AMTA tutorial.
> > > >
> > > > -John
> > > >
> > > > On Thu, Dec 1, 2016 at 10:17 AM, Mattmann, Chris A (3010) <
> > > > chris.a.mattm...@jpl.nasa.gov> wrote:
> > > >
> > > >> Wow seems like this kind of overlaps with BigTranslate as well..
> > > thanks
> > > >> for passing
> > > >> along Matt
> > > >>
> > > >> 
> > ++
> > > >> Chris Mattmann, Ph.D.
> > > >> Principal Data Scientist, Engineering Administrative Office
> (3010)
> > > >> Manager, Open Source Projects Formulation and Development Office
> > > (8212)
> > > >> NASA Jet Propulsion Laboratory Pasadena, CA 91109 USA
> > > >> Office: 180-503E, Mailstop: 180-503
> > > >> Email: chris.a.mattm...@nasa.gov
> > > >> WWW:  http://sunset.usc.edu/~mattmann/
> > > >> 
> > ++
> > > >> Director, Information Retrieval and Data Science Group (IRDS)
> > > >> Adjunct Associate Professor, Computer Science Department
> > > >> University of Southern California, Los Angeles, CA 90089 USA
> > > >> WWW: http://irds.usc.edu/
> > > >> 
> > ++
> > > >>
> > > >>
> > > >> On 12/1/16, 4:47 AM, "Matt Post"  wrote:
> > > >>
> > > >>Just came across this, and it's really cool:
> > > >>
> > > >>https://github.com/ModernMT/MMT
> > > >>
> > > >>See the README for