The LSPs and Machine Translation Why Not Treat

The LSPs and Machine Translation: Why Not Treat MT as TM? David Canek, Mem. Source Technologies Torben Dahl Jensen, Oversætterhuset

Mem. Source Technologies • Offshoot of a Charles University research project started in 2006 with Sun Microsystems • Develops Translation and Authoring Software: – Mem. Source Translation Server – Mem. Source Translation Cloud – UTMA Authoring Server • Headquartered in Prague

Oversætterhuset / Translation House of Scandinavia • Leading Danish LSP with offices in Århus, Copenhagen and Kolding • Established in 1990 • Covers major European languages • Eager to explore new technologies to make the translation workflow more efficient

Background • The last Loc. World conference in Seattle covered MT deployments in Adobe, Autodesk and Cisco • Last year’s Loc. World in Berlin also covered primarily enterprise case studies on MT • What about LSPs and Machine Translation?

We Will Explore • MT deployment scenarios • MT quality assessment & monetization

MT ADOPTION

Who Got MT Technology First? Enterprises? LSPs? Translators?

Who Got MT Technology First? Enterprises? 1 st: Translators LSPs? Translators?

Who Gets the Latest Technology First?

Translators and MT • MT Deployment: easy – uploading files to Google Translate costs just a little bit of time; and it is free • MT Monetization: trivial – MT simply speeds up their translations, so translators get more work done in less time

Enterprises and MT • MT Deployment: challenging – but have the resources to manage this • MT Monetization: complex – but being on the top of the food chain they have the power to renegotiate rates and drive home the MTgenerated savings

LSPs and MT • MT Deployment: challenging – have limited resources and specific obstacles • MT Monetization: complex – will have to renegotiate translator rates to reflect MT savings

MT Deployment in an LSP

LSP Custom MT Development • Considerable time and money to develop custom MT engine • Can easily end up with MT quality far inferior than the free online MT services • Specific obstacles: multiple domains and language pairs • Google spent millions of USD, has excess of 100 billion words of training data. . .

A Scenario to Avoid • LSP asks translator to post-edit a text machine translated by the LSP’s MT engine. • The quality is poor. Translator deletes the machine translation and instead uses GT, gets much better results. . . • On what basis can the LSP ask the translator to charge a reduced rate?

Can LSPs Succeed with MT? Yes. But do not necessarily start by developing a custom MT engine. Instead: • Begin using a readily available MT service • Measure its benefits • See if/how you are able to monetize the benefits • Only then explore the MT technology options

BUSINESS CASE

Building a Business Case for MT (MT savings) minus (MT costs) = MT Profit

MT Quality Measurement Today Kirti Vashee

Mem. Source MT Quality Measurement • Simple, fast, precise • Extends the established translation memory analysis and discount schemes to machine translation Why not treat MT just as another TM?

How Does It Work Exactly? • Traditional translation memory analysis – Document source segment vs. TM source segment • Mem. Source machine translation analysis – Document target segment vs. MT target segment

Translation Memory Match TM MT Source Europarat Target

Translation Memory Match TM MT Source Europarat Target Council of Europe

Translation Memory Match 100% TM MT Source Europarat Target Council of Europe

Translation Memory Match 100% TM MT Source Europarat Target Council of Europe

Machine Translation Match TM MT Source Europarat Target

Machine Translation Match TM MT Source Europarat Target Council of Europe

Machine Translation Match TM MT Source Europarat Target Council of Europe 100% ?

Machine Translation Match TM MT Source Europarat Target Council of Europe 100% ?

Machine Translation Match TM MT Source Europarat Target Council of Europe 100% ✓

Analyzing MT Matches Simply analyze MT matches and add them to the existing TM matches:

Analyzing MT Matches Simply analyze MT matches and add them to the existing TM matches:

Analyzing MT Matches Simply analyze MT matches and add them to the existing TM matches:

Turning MT Matches into Money Use your own discount scheme, e. g. : TM Match New words 75%-84% 85%-94% 95%-99% 100% % of Rate Paid 100% 50% 33% 25% 10%

Turning MT Matches into Money. . . and add MT matches TM & MT Matches New words 75%-84% 85%-94% 95%-99% 100% % of Rate Paid 100% 50% 33% 25% 10%

Knowing Your MT Savings When you know your MT savings, you can also better decide how much you can afford to pay for the MT service/technology.

CASE STUDY RESULTS

Case Study Overview • Two LSPs participated • January – May 2011 • Domains: – Marketing – Law – EU – Technology

Case Study Overview • Language pairs: – English > Danish – English > Norwegian – English > Czech – Czech > English – English > German • Volume: 1 million words • Two MT engines: GT and a custom MT engine

Case Study Results: Domain MT Match Rate 0%-50% 50%-74% 75%-84% 85%-94% 95%-99% 100% Legal 73% 17% 5% 2% 1% 2% Technology 38% 21% 10% 14% 6% 11%

Case Study Results: Language MT Match Rate 0%-50% 50%-74% 75%-84% 85%-94% 95%-99% 100% EN>CS 72% 14% 6% 3% 2% 3% EN>DE 62% 19% 8% 4% 2% 5%

Case Study Results: LSPs MT Match Rate 0%-50% 50%-74% 75%-84% 85%-94% 95%-99% 100% LSP 1 69% 15% 7% 4% 2% 3% LSP 2 63% 18% 6% 5% 3% 5%

Next Steps • Talking to translators and post-editors about the new approach • Negotiating TM/MT based discount schemes. . .

THANK YOU