GALA 2011: Machine Translation in Action

The Globalization and Localization Association (GALA) puts on its third annual conference in Lisbon on March 28-30, 2011.

MT in Action: Deployment Strategies from the Experts

Panelists (from the left): Smith Yewell (Welocalize), Olga Beregovaya (PROMT Americas), Catherine Dove (PayPal), Wayne Bourland (Dell)

This panel of experts talks about ins and outs of how an enterprise-scale MT product, coupled with a skilled post-editing team and a GMS solution is used by Fortune 100 clients to help get content to market quickly and efficiently, while addressing complex production workflows and interoperability requirements.

Client perspective:

  • Talk to other clients, vendors, CSA, TDA/TAUS, etc.
  • Know what you are trying to accomplish before you start the conversation
  • Test different vendor/technology combinations and understand the delta between each scenario and current HT output

MT post-editing – LSP view:

  • Often outside our control: source quality and suitability; MT engine “raw output” quality; client quality and throughput requirements
  • Within our control/influence: quality of post-edited output; capability/efficiency of post-editor
  • What do we need to be able to measure: MT raw output quality; post-editing effort required; post-editing effort efficiency

Client, MT provider and LSP working together:

  • Client formulates the program requirements
  • MT provider and LSP help the client select the most suitable pilot project
  • MT provider, LSP and client define the solution architecture
  • MT provider trains the engine
  • Several feedback loops with automated scores, human PE measurement, human quality assessment
  • LSP calculates PE (post-editing) metrics
  • MT-PE projects go live

What is the best content to be translated with MT?

  • User guides and manuals
  • Customer support sites and knowledge databases
  • Online help
  • E-Learning and training
  • User-generated content (blogs, e-mails)
  • Company websites

Case Study I: PayPal MT Deployment

  • MT Engine: PROMT
  • Post-editing: Welocalize
  • 1.5 million words translated to get started (per language)
  • 25K words translated weekly
  • Nine language pairs and 13 locales deployment
  • Content types: UI, Online help, FAQ, Error messages, Customer support, User guides
  • Workflow: SDL Idiom WorldServer <> Idiom-PROMT XLIFF connector <> PROMT PTS DE server
  • Post-editing environment: Idiom WorldServer Desktop Workbench
  • Rationale for using MT: costs savings, decreased time to market, better quality control, helping other teams with content writing, detecting coding errors, UI design, etc.
  • Planning to use MT for translation of customer e-mails
  • Provides virtual style guide and dictionary
  • MT performs better than human translation for variables (piped-in text)
  • Next steps: migrate to hybrid MT engine, add new languages, evaluate external hosting model

Case Study II: Dell MT Deployment

  • Very large corpus and good leverage (40%)
  • 22 million pages at dell.com
  • Right time to move beyond support content
  • Need to scale cost at a fraction of the content growth
  • Continue working with current preferred vendor (Welocalize)
  • Use PROMT hybrid technology
  • Time-to-market: 200% average productivity gain
  • Translation quality rated “good”
  • BLEU/Meteor scores: 200% improvement
  • Dell glossary and TM prevail over baseline engine
  • Not quite sure about cost savings at this point
  • Use of MT training shared documents (all stakeholders involved)
  • Workflow similar to Case Study I (PayPal)
  • Anticipated time-to-market gains: throughput currently at 4,000 to 8,000 words per day; further MT training will lead to higher throughput

Q&A:

  • Q: Comment on what did not work.  A: PayPal: Unrealistic expectations about the timeline; Dell: initial lack of understanding about proper use of MT; Welocalize: could not sell the solution in the beginning to the clients and translators did not like to post-edit at first
  • Q: Why do you use Idiom WorldServer? A: It’s the client’s choice
  • Q: How do you pay for the services? A: Depends on what the client wants. PayPal wants everything in-house, trained their own staff and have perpetual license. Dell was initially interested in hosted solution. Unlimited use of MT, not based on volume.