TAUS USER CONFERENCE 2010, Sony, Pangeanic - moving on with mt - building open source mt with your vendor and serving it
1. TAUS USER CONFERENCE 2010
LANGUAGE BUSINESS INNOVATION
4 – 6 OCTOBER / PORTLAND (OR), USA
TUESDAY 5 OCTOBER / 15.25
MOVING ON WITH MT: BUILDING OPEN-SOURCE
MT WITH YOUR VENDOR AND SERVING IT
Salomé López-Lavado, Sony
Elia Yuste, PangeaMT
2. Agenda
Toward a customized PangeaMT Solution for
Sony Professional Solutions Europe (PSE)
Going beyond MT engines
Solution’s benefits and latest features
PangeaMT4SonyE in action
Future work
Q&A
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 2
3. Toward a customized PangeaMT solution
for Sony PSE
Customized engines to benefit from client-specific
data and domain-specific data (TAUS)
Client understands concept and potential usefulness of TAUS
TDA for own MT development and for the GILT community
( data donor)
Main challenge: Heavily formatted, marked-up content
Typical Moses-only SMT implementation wouldn’t be enough!
Necessary to develop peripheral Modules to overcome this
limitation
Inliner – to tackle inline formatting
TMX and XLIFF filters – to go ahead of SMT operating in plain text
only
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 3
4. Route to PangeaMT4SonyE solution
Multi-genre within
Automatic evaluation same domain;
metrics (PangeaMT gradual development
team). Now get the & intake; separate &
user/client involved in Target domains combinatory engine
“experimenting & and MT goals approach
measuring”
Measuring Training /
results Building / Testing
Inliner and
TMX/XLIFF I/O
Service through user- filters – system’s
protected web panel Implementation
most innovative
and MT output delivery assets of
facilitates ease-of-use customized
and interactivity implementation
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 4
5. Going beyond the MT engine:
Solution delivery options
Cygwin (1.7.5-1): Password-protected web
PangeaMT engine, panel
models, Moses and
Interface is scalable
instructions
(more users, language
Zipped pack (300-500 Mb); pairs and engines
installed (1 Gb HD)
make a case for MT
Pre-tested functionality in across corporation)
MS-Windows versions (XP,
Vista and 7) Hosted in-house or by
Command line-based Pangeanic
translation request Easy tracking of MT
requests
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 5
6. Going beyond the MT engine:
Solution delivery options
Cygwin (1.7.5-1): Password-protected web
PangeaMT engine, panel
models, Moses and
Interface is scalable
instructions
(more users, language
Zipped pack (300-500 Mb); pairs and engines
installed (1 Gb HD)
make a case for MT
Pre-tested functionality in across corporation)
MS-Windows versions (XP,
Vista and 7) Hosted in-house or by
Command line-based Pangeanic
translation request Easy tracking of MT
requests
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 6
7. Solution’s Benefits
Proposed workflow based primarily on MT + Post-
Editing (PE)
Post-edited MT output then stored for retraining purposes
(SMT paradigm)
A TM-only-based workflow is outdated and entails undesired
lock-in effect
Gradual intake of XLIFF-based flow
Interoperability
Unaltered <source> and the <target> available:
Through the “state” attribute, possible to revert a <trans-unit> to a
non terminal state (“needs-translation”, or “needs-review-
translation”)
Possible to support translation quality assurance/control activities
(e.g. by sending out XLIFF to reviewers)
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 7
8. Latest Advancement
Inliner gets interactive!
The Tag-Optimal or Trans-Optimal? feature in
web interface
Particularly useful in Sony Europe tag-rich content
Still in beta version, first comparative tests based on
inputting same file and select the two options
consecutively show adequate performance of Tag-
Optimal option:
It restores inlines remarkably well while hardly sacrificing
translation quality
Need to perform very few PE operations/segment
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 8
10. S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 10
11. S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 11
12. S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 12
13. Envisaged work
Gradual customized development expansion
An extensive, monitored usage will allow for
measuring results systematically and carry
out client-specific studies (& improvements):
Productivity
Over a period of time - after engine retraining, related to
better Quality
Genre- (rather than domain-)specific error typology – in
connection with MTPE
Determine ROI and cost-efectiveness (yearly,
cross-annually, per languages, across corp depts)
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 13
14. Thank you !
Salome.Lopez-Lavado@eu.sony.com
eyuste@pangea.com.MT
S. López-Lavado & E. Yuste TAUS User Conference, Portland, 2010/10/5 14