This was a presentation given at the conference of the Association of Machine Translation in the Americas (AMTA) in Austin, Texas on October 31st, 2016. This is a predominantly academic event, and this presentation was a condensed version of our "MT Success Blog Series" on our website where we aimed to give the community and idea as to the practical considerations around commercial machine translation.
http://iconictranslation.com/2016/07/8-steps-to-mt-success-series-introduction/
What? Why? How? Factors that impact the success of commercial MT projects
1. What? Why? How?
Factors that impact the success of
commercial MT projects
John Tinsley
Iconic Translation Machines
AMTA – Austin, Texas – October 2016
2. Why MT?
• Speed
• Cost savings
• Time to market
• Your competitors are
doing it!
“Why would I need MT?”
Why Now?
• volume of content is
growing
• demand, more words less
time
• growth facilitator
• #FOMO – you’re missing
out on business
What’s the MT value proposition?
4. • The goal of the MT here is to be good enough so that - on the
whole – with TMs, translators are faster post-editing some
segments
• Challenges
– development has to focus on reducing needs for edits, not
necessarily anything else
– translator acceptance always a big barrier
– evaluation can take time and has many factors
– pricing models
How are companies using MT?
What are the use cases for MT?
Translator productivity through post-editing
5. • The goal is to produce MT that’s fit for a particular purpose as
is
• Arguably easier from an MT development perspective
• Often high-volumes = more achievable
How are companies using MT?
What are the use cases for MT?
MT for information
10. Content Type
The bed was two twin beds put together and
me and my girlfriend kept fallin in the middle
(since we like to cuddle) and that was iritating
Late nite room service was awesome
Social Media
User Generated Content
Highly Technical
Marketing, Nuanced
12. MT experience
Little experience A lot of experience
Har
d
“Easy”
LSP/vendor
experience with MT
Ease of
adoption
The more experience the LSP has
with onboarding/training vendors,
and the more experience the vendor
has with MT, the more feasible the
adoption of MT will be
13. Integration requirements
Standard vs Custom Integration
“instant” solution costs rise
proportionality with the number of
languages and the throughput needs
14. TM Leverage
High TM
Leverage
Low MT
Effectiveness
Matches # words
Context 403,803
100% 585,459
95-99% 50,366
85-94% 41,604
75-84% 32,319
50-74% 18,972
No Match 81,119
Total 1,213,643
Only 8% of
all words go
to MT
15. Quality requirements
• Fully automatic human quality
• 300% post-editing productivity
• French to Spanish == English to Korean
• Best performance out of the box
17. • “What volume of words do you estimate for the project?”
• “Do we have translation memories, glossaries that are
relevant? Can we create them?”
• “If so, what leverage are we getting?”
• “To we have post-editors? Access to a supply chain?”
– “what experience do they have?”
• “Where will MT fit in the workflow (depending on the use
case)?”
• “What variety is there in the content that the MT will be
processing?”
• “Why aren’t you using Google Translate?”
• “Is there sufficient budget for this project?”
What questions should YOU be asking?
18. “How much
training data do I
need?”
“How frequently
can I retrain the
engine?”
“What happens
to my data?”“Do you do
language X?”
“How good is
the quality?”
“How do you
measure
performance
over time?”
john@iconictranslation.com
www.iconictranslation.com
@iconictrans
Editor's Notes
Let’s look at some of the initial basic questions you might have to field you use to convince clients, to convince yourself (especially if you’re PROACTIVELY making a decision)
It’s a common question to be asked “What type of machine translation engine is it?”
It’s often asked but not necessarily because the buyer is looking for something specific. A lot of it is marketing and doesn’t actually mean much. Whoever is asking probably just wants to make sure you have something cutting edge that suits their needs.
That’s why “CUSTOM” is a buzzword but it’s important because it mean’s it’s designed for purpose.
Other things (enterprise, cloud, etc.) have nothing to do with the type of MT but rather how they’re delivered. And others are just random adjectives
Blurb about statistical and rules and most things being or claiming to be hybrid when really they’re just predominantly one with a sprinkling of the other.
They’re all more or less used for the same thing anyway, which is the following…
Butterfly effect. One small change can have an impact down the line….
Same with MT, many factors at play and one small factor can have an impact on the outcome of a project
This is really the crux of the matter…
Not in any strict order necessarily because they’re not mutually exclusive but maybe some have a bigger impact or are less obvious than others.
These effect projects broadly in 2 ways: FEASILITY i.e. whether MT will actually work at all, and COST: some have no impact on whether MT will work or not, but it could impact whether MT with work for YOUR needs given the cost associated.
**EFFECT ON FEASIBILITY**
Basically, some languages are easier for MT that others.
General rule, closer two languages are to one another in terms of word order, grammatical structure, the easier.
Here’s some rules of thumb (with English)
**EFFECT ON FEASIBILITY AND COST**
sometimes volume too low can effect the feasibility of a custom MT solution, Making the economics of investment not worth it…
but obviously, volume impacts cost so if someone has a project with billions of words, you need to manage their expectation that it’s going to cost…
**EFFECT ON FEASIBILITY **
Related to the language, just adds a little to the aspect of feasibility… could impact on cost in terms of development time
**EFFECT ON FEASIBILITY**
Without sufficient relevant training data, we may not be able to build a good enough engine.
How much data is required?! How long is a piece of string!
**EFFECT ON FEASIBILITY**
Experience also has a lot to do with acceptance…
**EFFECT ON COST**
Integration and delivery methods can drive the cost up….
Via a plugin, probably ok. Some bespoke integration, depending on the complexity, can carry professional services costs.
24/7 instant translation across multiple languages…
Cloud is beautiful
**EFFECT ON FEASIBILITY**
8% - not that there’s anything wrong with that. That’s great. It just means MT is less likely to have a positive impact here.
These things are often things that are difficult for MT too perhaps?
**EFFECT ON FEASIBILITY**
These CAN be the case depending on the conditions, but they would be exceptions rather than the rule
Actually, we’re planning a series of articles on our blog deep diving into each of these topics so stay tuned…
Questions you should be asking yourself!Treat this almost like a checklist!