SimulTrans Localization Blog: SimulTips

Myth #3: Editing machine translation is a waste of time.

[fa icon="calendar"] June 11, 2014 / by the SimulTrans Team

Are there still people among us in the translation and localization industry who support this view? Based on the sample comments below from a professional translator networking site, it seems there are. Many translators remain leery of machine translation technology: 

"I would prefer to start from scratch..."

"Very often these projects are not fit for further processing."

"Post-editing MT is not very gratifying."

Editing machine translation can be a pointless exercise – as the translators quoted here have no doubt experienced – if the entire machine translation process is not expertly managed from start to finish. When asked to edit (or ‘post-edit’) gobbledygook from an uncustomized machine translation engine and offer a discount on their rates, it’s no wonder some translators run for the hills. 

At SimulTrans, where we’ve handled machine translation and post-editing for selected projects since 2001, we’ve learned that post-editing machine translation can produce excellent results and save money. But it’s not a simple matter of throwing texts into Google Translate. The five steps below help to guarantee a successful outcome:

1. Analyze the content for Machine Translation suitability

Despite remarkable progress in both rules-based and statistical machine translation, the number and type of projects that are suitable for machine translation remains limited. With this in mind, our machine translation team conducts a linguistic and technical analysis of the source files to identify if the text type, language pairs, linguistic structures and file types lend themselves to processing by our in-house machine translation systems.

2. Invest effort in preparing and maintaining the Machine Translation engine

Whether a rules-based or SMT system is applied, a lot of upfront investment is required to compile integrated glossaries and, depending on the system used, either customize rules or prepare corpuses of clean data for training an machine translation engine. If a rules-based system is not adequately customized, or if inappropriate or insufficient data is fed into an SMT system, the raw output will simply support the myth that editing machine translation is a waste of time.

3. Train and support the post-editors

The machine translation post-editing guidelines developed by TAUS are a useful starting point. Our in-house team also provides ongoing training and feedback to post-editors to help them contribute to continuous improvement of the machine translation system and thus save more time.

4. Define quality expectations

Both the customer and post-editors need to be involved in defining a satisfactory quality level, with reference to sample target texts. Using machine translation can actually improve overall quality in some respects, such as consistency and terminology. However, there’s little point post-editing machine translation output if it takes longer than the standard translation process, or if the customer is not happy with the results.

5. Define productivity expectations

Expectations within the industry vary hugely, with some machine translation developers reporting daily post-editing throughputs of up to 20,000 words – more than 10 times standard translation output. Most practicing translators would contend that it’s barely possible to proofread that volume of human translation, let alone check machine translation. At SimulTrans, we focus on achieving reasonable and sustainable productivity levels and cost savings with machine translation. Savings for customers can be in the region of 20-30% if ‘human translation quality’ is required, and even higher for some text types and quality expectations.

Greg Hellmann, Translation and Machine Translation Manager at SimulTrans comments: 

"It's ironic that the people who may (or may not be) most wary of machine translation technology (i.e. professional translators) seem to be the ones who can actually build the best machine translation engines.  At SimulTrans, our in-house translators fully embrace the linguistic challenge and benefit of optimizing a machine translation system.  The reason is simple: Unlike many providers of machine translation services, we put our linguists in the driver's seat and allow them to fully control and develop the machines.  After all, they are the ones post-editing the engine's output and have the best sense for the error patterns of particular engines.  There is a massive element of hype accompanying machine translation in the localization industry with lots of empty promises.  The important thing for me is this:  When SimulTrans translators are working, and they need machine translation to be more efficient and deliver projects faster...when our clients need it... SimulTrans knows how to handle it.  And it works."

Topics: Localization Technology, Translation Services

the SimulTrans Team

Written by the SimulTrans Team

SimulTrans provides software, document, and website localization services, translating text into over 100 languages. Established in 1984, SimulTrans has enabled thousands of businesses to provide high-quality content to their international customers. Management ownership allows an exclusive focus on customers and quality, as exemplified by ISO 9001 and ISO 17100 certifications. In addition to its headquarters in Mountain View, California, SimulTrans has offices in Boston, Dublin, London, Paris, Bonn, and Tokyo.