The document summarizes a TAUS Machine Translation Showcase event held in Vancouver, Canada on October 29, 2014. It includes an agenda for presentations on machine translation applications at eBay, getting started with SMT, seamless globalization with crowd posting editing, and an introduction to the Matecat open-source CAT tool. The document also provides an overview of the machine translation market trends presented by TAUS, including growing market size, opportunities and challenges in the industry, and predictions for the future of machine translation and post-editing.
1. TAUS
MACHINE
TRANSLATION
SHOWCASE
Vancouver,
Canada
TAUS Introduction and MT Market Overview
Wednesday, 29 October 2014
Jaap van der Meer & Achim Ruopp,TAUS
The
research
within
the
project
MosesCore
leading
to
these
results
has
received
funding
from
the
European
Union
7th
Framework
Programme,
grant
agreement
no
288487
2. TAUS Introduction and MT Market
Overview
Jaap
van
der
Meer,
TAUS
Achim
Ruopp,
TAUS
Localiza)on
World
Vancouver
29-‐Oct-‐2014
3. This slide may not be used or copied without permission from TAUS
TAUS
Machine
TranslaHon
Showcase
13:30
/
TAUS
Introduc7on
and
MT
market
overview,
Achim
Ruopp
(TAUS)
14:00
/
Machine
Transla7on
at
eBay,
Saša
Hassan
(eBay)
14:30
/
The
Simplified
Guide
to
GeGng
Started
in
SMT,
Tom
Hoar
(Precision
Transla)on
Tools)
15:00
/
Coffee
Break
15:30
/
Seamless
Globaliza7on
with
distributed
crowd
post
edi7ng,
Vasco
Pedro
(Unbabel)
16:00
/
Introduc7on
to
Matecat,
the
open-‐source
CAT
tool
for
post-‐edi7ng,
Marco
TrombeM
(Translated)
16:30
/
Podium
Discussion
17:00
/
Adjourn
MosesCore
is
supported
by
the
European
Commission
Grant
Number
288487
under
the
7th
Framework
Programme.
4. This slide may not be used or copied without permission from TAUS
The
Changing
Nature
of
the
MT
Market
o ExecuHve
Summary
and
Mega
Trends
o Past,
Present
and
Future
of
MT
Research
o Different
Usages
of
Machine
TranslaHon
o Types
of
Players
o Types
of
Offerings
o Defining
the
Market
–
the
Numbers
o Market
OpportuniHes
and
Challenges
o Market
Drivers
and
Inhibitors
o PredicHons
5. This slide may not be used or copied without permission from TAUS
TAUS
Machine
TranslaHon
Market
Report
Execu)ve
Summary
o Market
size:
$250
Million,
growing
16.9%
per
year
o “Perfect
storm
condiHons”
for
MT
o Key
trends:
§ GlobalizaHon,
IntegraHon,
Convergence,
VerHcalizaHon,
Immediacy
of
communicaHon,
Privacy
–
security,
High
quality
translaHon
o OpportuniHes:
§ Business
expansion,
IntegraHon
of
MT,
ProducHvity
gains,
MT
as
enabler
for
new
services,
Narrow
domain
applicaHons,
Customer
support
self-‐service
o Challenges:
§ False
expectaHons
–
false
starts,
Quality
of
MT,
Language
coverage,
Available
training
data,
Specialist
skills,
Vendor
lock-‐in,
CompeHHon
from
free
MT,
Quality
measurement
&
esHmaHon
o PredicHons:
§ Post-‐ediHng
MT
will
grow
very
quickly
and
become
the
primary
producHon
process
in
translaHon
within
five
years.
§ MT
technology
itself
is
on
its
way
to
become
a
commodity,
shijing
the
Holy
Grail
to
the
data
6. This slide may not be used or copied without permission from TAUS
“Perfect
Storm
CondiHons”
1. Ease of communications
2. Hyperglobalization
3. Democratization of knowledge
4. Linguistic diversity
7. This slide may not be used or copied without permission from TAUS
Entering
the
Convergence
Era
8. This slide may not be used or copied without permission from TAUS
Past,
Present
and
Future
of
MT
Research
History
of
Machine
TranslaHon
Research
o Many
ups
and
downs
since
the
1950s
o Strong
compeHHon
between
vastly
different
approaches
o Sudden
leaps
of
improvement
o Ojen
parallel
development
in
academia,
government
and
industry
o Moved
from
ridicule
to
acceptance
for
many
uses
over
the
last
couple
of
years
§ Cynic’s
view
that
FAHQMT
“fully
automated
high
quality
machine
translaHon”
is
always
five
years
away
misses
the
point
o Lately
academic
research
shared
as
open
source
9. This slide may not be used or copied without permission from TAUS
Past,
Present
and
Future
of
MT
Research
Current
Trends
-‐
Hybrid
and
Other
Approaches
o Combine
the
best
features
of
the
linguisHc
approach
and
the
more
modern
staHsHcal
approach
§ Ojen
leads
to
higher
output
quality
§ Lower
customizaHon
costs
o Leads
to
bewildering
range
of
opHons
for
building
the
best
MT
system
for
a
specific
language
pair
and
use
case
§ Common
pracHce
of
picking
single/few
opHons
has
been
likened
to
“alchemy”
by
leading
MT
researcher
o Further
adopHon
of
modern
AI
techniques
§ Deep
learning
with
neural
networks
is
hot
research
topic
10. This slide may not be used or copied without permission from TAUS
Different
Usages
of
Machine
TranslaHon
GisHng
(AssimilaHon)
o Understanding
the
gist
or
central
point
of
a
text
or
conversaHon
in
a
foreign
language
o Conveying
the
semanHc
meaning
more
important
than
syntacHc/grammaHcal
correctness
o Highest
volume
use
of
machine
translaHon
currently
o Examples
§ “Translate
this
page”
links
in
Google
search
results
§ “Translate”
links
for
Facebook
posts
§ Hotel
reviews
on
TripAdvisor
§ Augmented
reality
sign
translaHons
in
Wordlens
app
11. This slide may not be used or copied without permission from TAUS
Different
Usages
of
Machine
TranslaHon
Search
and
Discovery
o Discovery
of
foreign
language
content
of
relevance
to
the
searcher
§ Previously
ojen
not
discoverable
o Closely
related
to
gisHng
o Huge
opportunity
for
human
translaHon
§ Follow-‐up
human
translaHon
of
discovered
content
o Examples
§ eDiscovery
–
finding
relevant
documents
for
legal
cases
§ Patent
translaHon
§ News
translaHon/monitoring
12. This slide may not be used or copied without permission from TAUS
Different
Usages
of
Machine
TranslaHon
SenHment
Analysis
o AutomaHc
detecHon
of
senHment,
ojen
negaHve
or
posiHve
senHment,
in
foreign
language
content
o Basically:
discovery
and
gisHng
for
machines
o Difficult
in
mulHlingual
content
as
ojen
two
imprecise
staHsHcal
systems
are
involved
§ Machine
translaHon
§ SenHment
analyzer
o Example
§ Stock
trading
based
on
senHment
analysis
13. This slide may not be used or copied without permission from TAUS
Different
Usages
of
Machine
TranslaHon
Post-‐ediHng
(DisseminaHon)
o Human
ediHng
of
machine
translated
content
to
a
desired
quality
level
o Quickly
becoming
part
of
the
tool
set
in
the
translaHon
industry
o Various
studies:
30-‐40%
producHvity
gain
o More
important:
Faster
turn-‐around
Hmes
o Useful
for
many,
but
not
all
translaHon
jobs
o AdopHon
challenges
§ IntegraHon
into
workflows
§ Difficult
customizaHon
and
evaluaHon
§ Translator
concerns
o Research
into
Hghter
MT
–
ediHng
integraHon
to
aid
editor
in
best
possible
way
14. This slide may not be used or copied without permission from TAUS
Different
Usages
of
Machine
TranslaHon
Speech
TranslaHon
o Speech-‐to-‐speech
translaHon
requires
combinaHon
of
three
systems
§ AutomaHc
Speech
RecogniHon
(ASR)
§ Machine
TranslaHon
§ Text-‐to-‐Speech
(TTS)
o CombinaHon
of
three
staHsHcal
systems!
o Spoken
language
more
difficult
to
machine
translate
than
well
formed
text
o Despite
difficulty
many
system/apps
in
this
intriguing
area
§ Promises
immersion
into
foreign
language
environment
15. This slide may not be used or copied without permission from TAUS
Types
of
Players
in
the
Machine
TranslaHon
Market
o MT
suppliers
§ Long
established
players
o Ojen
started
out
with
strong
economic
basis
of
government/
insHtuHonal
buyer
§ New
players
o Using
opportunity
of
increasing
MT
awareness/adopHon
o Ojen
using
available
open
source
soluHons
as
a
basis
§ Has
commodizaHon
started?
o Value-‐added
resellers
§ Using
machine
translaHon
to
enhance/complement
an
exisHng
service
§ See
different
uses
of
machine
translaHon
§ More
unexpected
innovaHve
uses
expected
§ Most
important
value
proposiHon
of
MT?
16. This slide may not be used or copied without permission from TAUS
Types
of
Players
in
the
Machine
TranslaHon
Market
o Free
online
machine
translaHon
services
§ Google/Microsoj/Yandex/Baidu
§ Cross-‐subsidized
by
uses
that
generate
revenue
e.g.
adverHsing,
platorm
use
§ Paid
API
use
o In-‐house
users
of
machine
translaHon
§ Governments,
mulHnaHonal
organizaHons
and
mulHnaHonal
companies
§ Strategic
importance
warrants
costs
of
developing/
maintaining
MT
systems
in-‐house
§ Most
flexibility
17. This slide may not be used or copied without permission from TAUS
Types
of
Offerings
in
the
Machine
TranslaHon
Market
Licenses
and
MTaaS
o Licenses
§ TradiHonal
model
of
sojware
distribuHon
o SHll
important
for
server,
not
desktop
§ Provides
a
lot
of
flexibility
and
opHons
for
customizaHon
o OperaHonal
know-‐how
required
§ Provides
highest
degree
of
privacy
§ Allows
translaHon
of
unlimited
number
of
words
o Machine
TranslaHon
as
a
Service
(MTaaS)
§ MT
running
on
MT
provider
infrastructure
§ Ojen
with
subscripHon
pricing
o In
many
cases
preferable
for
supplier
and
buyer
over
high
up-‐front
licensing
fees
§ Web-‐based
user
interfaces
for
MT
training/operaHon
o Some
loss
of
flexibility/control
o Presets
not
always
a
negaHve
18. This slide may not be used or copied without permission from TAUS
Types
of
Offerings
in
the
Machine
TranslaHon
Market
Volume-‐Based
Machine
TranslaHon
Services
o Online
machine
translaHon
services
aim
to
provide
machine
translaHon
§ In
many
language
pairs
§ Worldwide
via
the
internet
o General
domain
§ CustomizaHon
only
via
Microsoj
Translator
Hub
o Very
affordable
o Cross-‐subsidizaHon
puts
long-‐term
availability
of
services
into
quesHon
19. This slide may not be used or copied without permission from TAUS
Types
of
Offerings
in
the
Machine
TranslaHon
Market
Professional
Services
o CustomizaHon
§ Ojen
in
combinaHon
with
license/MTaaS
offerings
§ Data
preparaHon
of
customer-‐owned
training
data
§ MT
engine
training
o Business
consulHng
§ OpportuniHes
to
streamline
processes
§ OpportuniHes
to
generate
new
business
§ Business
consultants
offer
industry
experience
and
shared
industry
knowledge
how
the
new
technology
can
be
applied
20. This slide may not be used or copied without permission from TAUS
Defining
the
Machine
TranslaHon
Market
o Re-‐convergence
of
TM
and
MT
o MT
technology
as
an
enabler
for
other
business
benefits
or
revenue
generaHon
o Paradox
of
a
vibrant
MT
market
and
a
relaHve
small
size
o Facebook,
Baidu,
Google,
Microsoj,
Yandex,
eBay
are
strongest
MT
operators
without
a
goal
of
generaHng
revenue
from
pure
MT
o Focus
on
MT
has
changed
from
FAHQT
to
a
tool
to
support
global
communicaHons
21. This slide may not be used or copied without permission from TAUS
Market
AdopHon
and
Usage
o IdenHfied
65
MT
operators
o Largest
MT
providers
in
alphabeHcal
order:
§ CSLi,
Google,
IBM,
Lionbridge,
Microsoj,
PROMT,
Raytheon
BBN,
SDL,
Smart
CommunicaHons,
SYSTRAN.
22. This slide may not be used or copied without permission from TAUS
Market
AdopHon
and
Usage
Supplier
Revenue
Percentages
Server licenses
16% Desktop licenses
3%
SaaS
17%
Free
0%
Word/volume
27%
Consultancy
9%
Customization
28%
Other
0%
Revenue percentage per offering type
Excluded revenue from post-editing services
23. This slide may not be used or copied without permission from TAUS
Market
AdopHon
and
Usage
Supplier
Revenue
Percentages
Geographical
DistribuHon
North America
46%
Europe
32%
South America
2%
Asia
17%
Rest of World
3%
Revenue percentages per geography
24. This slide may not be used or copied without permission from TAUS
Market
Trends
o GlobalizaHon
o IntegraHon
o Convergence
o VerHcalizaHon
o Immediacy
of
communicaHon
o Privacy
–
security
o High-‐quality
translaHon
25. This slide may not be used or copied without permission from TAUS
Market
OpportuniHes
o Business
expansion
o IntegraHon
of
MT
o ProducHvity
gains
o MT
as
an
enabler
for
new
services
o Narrow
domain
applicaHons
o Customer
support
self-‐service
26. This slide may not be used or copied without permission from TAUS
Market
Challenges
o False
expectaHons
–
False
starts
o Quality
of
MT
o Language
coverage
o Available
training
data
o Specialist
skills
o Vendor
lock-‐in
o CompeHHon
from
free
MT
o Quality
measurement
and
esHmaHon
27. This slide may not be used or copied without permission from TAUS
Drivers
and
inhibitors
28. This slide may not be used or copied without permission from TAUS
Seven
PredicHons
1. Post-‐ediHng
MT
will
grow
very
quickly
and
become
the
primary
producHon
process
in
translaHon
within
five
years.
2. VerHcalizaHon
of
MT
will
conHnue.
Innovators
will
offer
MT
embedded
in
apps
and
hardware
to
run
a
specific
task.
TranslaHon
operators
will
differenHate
themselves
from
free
or
cheap
generic
MT
systems
by
developing
domain
and
customer-‐specific
engines.
3. Training
and
customizing
MT
engines
will
become
much
simpler
in
the
next
five
years,
making
it
possible
for
translators
and
project
managers
to
train
a
new
engine
by
uploading
reference
documents.
4. Spoken
translaHon
(convergence
of
MT
with
speech
technology)
will
become
widely
available
in
the
next
five
years.
5. MT
will
start
playing
a
crucial
role
in
Big
Data,
business
intelligence
and
the
Internet-‐of-‐Things.
6. The
translaHon
industry
will
start
to
agree
on
best
pracHces,
metrics
and
benchmarks
for
automated
translaHon.
7. Access
to
training
data
becomes
a
bigger
challenge
than
access
to
MT
technology.