Are You Ready for Machine Translation?

By Alan Houser | Fellow

A profound change is underway in the translation industry. Organizations are changing the way they translate their content, using a technology called machine translation. Machine translation offers new ways to reach your global audience, in your customers’ native languages. But knowing what machine translation can provide, and how to use it in your translation workflows, is critical.

Machine translation is not a magic solution. Organizations that attempt to deploy pure machine translation, without careful consideration of their translation requirements, business requirements, and successful machine translation workflows, are likely to be very disappointed. Let’s consider how you can successfully deploy machine translation to increase your organization’s global presence and reach your customers more efficiently and effectively.

What Is Machine Translation?

In the Global English Style Guide, John Kohl defines machine translation as “software that translates sentences from one language (such as English) into one or more other languages (such as French or Japanese).” Sounds good, right?

Machine translation has been around since the 1950s. Translation quality, however, has generally kept machine translation out of mainstream translation processes. But recent developments in machine translation technology, cloud-based services, and machine translation workflows have converged to bring machine translation into the mainstream.

How Does It Work?

Implementers of machine translation software and systems use one of two approaches. Rules-based machine translation uses the rules of grammar, syntax, and meaning for each language to develop a translation.

Statistical machine translation performs translations by identifying the most likely translation for source content, one word or phrase at a time, to produce the most statistically valid translation. Statistical machine translation systems make this assessment after being “trained” with large volumes of previously translated content.

Both approaches have their strengths and limitations. Rules-based machine translation tends to work best with source content that is of high quality and consistency. However, rules-based machine translation may be ineffective with content from a highly specialized subject or domain, and robust rules may not exist for less popular language pairs.

Statistical machine translation can be highly effective with specialized subject matter, if a substantial amount of previously translated content exists. Organizations with large volumes of previously translated content and well-maintained translation memories may be good candidates for statistical machine translation.

What About Quality?

The promise of machine translation is enormous. Translation is expensive. Imagine that you can translate your organization’s content immediately, automatically, and at trivial cost. That is possible, with one catch.

Machine translation quality is generally far lower than that of human translators. You have probably seen examples of laughable or even disastrous machine translations. Machine translation may not approach the quality of human translators for a long time, if ever. Any machine translation initiative will require ways to assess the quality of machine-translated output, and adjust the machine translation workflow accordingly.

If you work with machine translation, you will encounter the BLEU score for measuring machine translation quality. Simplistically, the BLEU score reflects:

  • How many words are correctly translated into the target language?
  • How many machine-translated words are in the correct order in the target language?

To evaluate the quality of a machine translation, you must have a human-generated translation as a reference. For this reason, machine translation is often most effective where a large body of previously translated content already exists, along with a large, well-maintained translation memory.

Achieving Quality Results

You can achieve high-quality results with machine translation, if you engage human translators in the process. In this increasingly popular workflow, human translators post-edit machine translation output to bring it to a desired level of quality. Instead of translating from the original source language, a post-editor can start with a rough translation. When properly deployed, this can dramatically improve the efficiency of the human translator.

Translation professionals have been justifiably skeptical of machine translation post-editing. Original post-editing workflows often included arbitrary or unreasonable constraints on the post-editor (e.g., “change no more than three segments per paragraph”) and compensated post-editors poorly compared to conventional translation workflows.

Newer post-editing workflows leverage an organization’s translation memory; identifying accurate machine-translated segments, close matches, and unusable segments. Human translators “fix” the close matches and unusable segments based on the original source material, and are compensated for their efforts in the same way that they have always been. This provides for a familiar working environment for post-editors, fair compensation, and good translation results.

Consider Your Strategy

It’s important to have a sound strategy for deploying machine translation. Your strategy may include:

  • Human translators for highly visible marketing and website copy.
  • Straight machine translation for technical support and customer-generated content (such as customer-support forums).
  • Machine translation with human post-editing for technical documentation.

You may want to engage several different vendors. Assess quality carefully; you will find that machine translation quality can vary substantially for different language pairs.

Tips for Successful Deployment

Consider the following approach to successfully deploy machine translation:

  • Assess your customer requirements. In many circumstances (social media, customer support forums), your customers may be thrilled to have translated content in their native languages, and may be surprisingly forgiving of subpar translation quality.
  • Develop a machine translation strategy. This includes analyzing your content streams, assessing appropriate quality targets, developing translation workflows for each stream, and measuring the benefits of providing translated content.
  • Measure your results and refine your processes. Machine translation workflows can be complex, especially when high-quality results are required. Be sure to have metrics to measure the efficiency of your processes, quality of results, and business benefits. Be prepared to refine your machine translation processes and workflows.

Selecting a Machine Translation Vendor

Machine translation technologies and workflows are rapidly evolving. Ask your vendor the following questions:

  • How do they assess machine translation quality?
  • How do they train their machine translation system with your content?
  • How do they protect your training data? This is particularly important for hosted or cloud-based services.
  • What is their rate structure for post-editing? Some vendors provide a discounted rate for post-editing machine translation output. Some vendors will identify “unusable” machine translation output, and re-translate only that portion using human translators at customary rates.
  • What mechanisms does the vendor provide for improving translation quality?

Selecting, configuring, training, and maintaining machine translation systems and workflows is a complex task. If your current language service provider is not already doing so, consider engaging a vendor that specializes in machine translation.

Conclusion

The global translation market is growing, and customers are becoming more likely to expect content in their native languages. Although roles will remain for human translators, organizations require more efficient tools and workflows to successfully meet customer demand for localized content.

There is a substantial amount of fear and distrust of machine translation, as exists with any disruptive technology that can replace human effort. However, solutions are quickly emerging for overcoming issues with machine translation quality, and for effectively using machine translation along with human translators to meet the global demand for localized content.

Machine Translation in the Mainstream

Is the Universal Communicator of Star Trek a reality? Not quite, but almost. The ubiquity of smartphones and cloud computing has put machine translation capabilities in the hands of everybody. Consider the following recent developments:

  • Make Magazine featured a do-it-yourself project for a spoken-language translator. The project is based on the $35 Raspberry Pi micro-computer, an open-source speech-recognition API, and Google’s cloud-based translation service.
  • The Google’s Translate smartphone app accepts text, voice, and image input. For example, you can take a picture of a restaurant menu, and Google will provide a translation in the language of your choice. Google claims the app supports 90 languages, has 500 million monthly users, and delivers a billion translations a day.
  • Microsoft recently announced Skype Translator, which provides real-time audio translation and text transcription of Skype conversations between English- and Spanish-speakers.
  • Google, Microsoft (Bing), IBM (Watson), and others offer API-based services for organizations that wish to develop machine translation applications.

Resources

Swisher, Val. Writing for Translation, Even If Your Company Does Not Translate. Intercom, January 2015.

Kohl, John R. The Global English Style Guide, SAS Press Series, 2008.

Make Magazine, Universal Translator, http://makezine.com/projects/make-42/universal-translator.

LoPresti, Michael. Preparing Your Content for Machine Translation. EContent Magazine, October 2014.

By Alan Houser, © 2015 Intercom, Volume 62 Issue 3, March 2015