Isn’t It Time to Embrace Machine Translation Post-Editing? The Localization Use Case for MT

Many of us will already be familiar with the Project Management Iron Triangle:

It’s also known as the “Triple Constraint,” and for just reasons. Anyone who’s managed a localization project knows the constraints are real, and their impact on quality is certain.

Uber Constraint: Human and Mortal

Let’s not forget that localization is by nature a highly human task. The nuances perceived and achieved by humans will – most likely – never be replicated by machines, at least not in all content types. It takes a poet to translate poetry.

Humans will always be involved in the localization effort in one way or another, and we must embrace both the beauty and the hazards they bring.

That said, let’s consider each of the three triangle’s constraints. Suppose you have a simple (that is, not high-tech such as a multimedia project) 20,000-word localization project, fairly small and manageable.

Constraint 1: Scope

Want to speak to the entire world? According to Quora, a popular Q&A site, just to reach all of India and China, which alone have about 37-40% of the entire world’s population, you’d need to communicate in 450 distinct languages (dialects not included).

Let’s say you want to stick to the United Nations official languages, so you’d need six languages in total: English, Russian, Spanish, Arabic, Chinese and French. This means the scope of your simple project is:

20,000 words x 5 UN languages* = 100,000 words

*Not counting the source language

Constraint 2: Time

Industry standards dictate that the average translator with good knowledge of the subject matter can translate 2,000 words per day. Easy math:

20,000 words ÷ 2,000 words/day = 10 days …per language!

Constraint 3: Cost

The price of your translation will be affected by different factors:

  • Operational model: If you engage linguists directly, you won’t incur additional PM costs. If you have in-house linguists, per-word costs will depend on their salaries and productivity
  • Language: Each language has its own average price per word
  • Resource: Each linguist can charge a different price unless the LSP has fixed prices per language
  • Turnaround: As we know, the shorter the turnaround, the higher the price per word. Plus, other rush fees may apply

Considering all major languages and practiced price levels, we can set reasonable average per word prices for regular turnaround translation and review to calculate the final project costs:

Translation 100,000 words x $0.15 / word** = $15,000
Review 100,000 words x $0.05 / word** = $5,000
Total*     $0.20 / word** = $20,000

*Not including any PM costs
** Cost per word taken on industry average

Let’s Get Real

We all know that few projects are “regular,” and usually we need to choose which tip of the triangle to break – never compromising quality, of course.

  • Break “Scope”: Translate fewer words
  • Break “Cost”: Pay a higher price
  • Break “Time”: Wait longer

So, something’s gotta give. But what really are the choices?

 Solution 1: Wait under the Waterfall for a “Regular” Project

The traditional waterfall localization project management has been around for ages for a reason: it’s solid and works. In a perfect world, we would have time to comfortably accommodate all phases of the multilingual content supply chain: create, translate, deliver.

This is how your 20,000-word project into 5 languages would look in the traditional scenario:

Translation 100,000 words @ $0.15 / word = $15,000 @ 2,000 words/day = 10 days + 1 PM day
Review 100,000 words @ $0.05 / word = $5,000 @ 500 words/hour = 5 days + 1 PM day
Totals     $0.20 / word   $20,000       17 days

Solution 2: Open Your Wallet for a “Rush” Project

Of course, to get faster results, you can use multiple linguists at rush rates. To have the absolutely shortest turnaround still observing best practices, you will need:

Translation 20,000 words @ 2,000 words/day = 10 translators per language
Review 20,000 words @ 500 words/hour = 5 reviewers per language
Harmonization 20,000 words @ 2,500 words/hour = 1 reviewer per language

With all the handovers, content splitting and rejoining, just think about the administrative burden of managing 60 translators and 30 reviewers!… But here are the numbers:

Translation 100,000 words x $0.25 / word = $25,000 1 day + 1 PM day
Review 100,000 words x $0.10 / word = $10,000 1 day + 1 PM day
Harmonization 100,000 words x $0.10 / word = $10,000 1 day + 1 PM day
Totals     $0.45 / word   $45,000 6 days

Note: About Splitting the Baby

Administrative burden aside, if well managed, a team of qualified translators and reviewers can produce excellent results. But more often than not, the main issue with using multiple translators will be consistency and the integrity of the project. No two translators have the same style or use exactly the same words, and with such a complex project, it’s so easy for something to fall through the cracks!

Let’s remember that in a split project scenario, a solid terminology base and a comprehensive style guide will be vital for success.

Solution 3: Cut off Your Nose to Spite Your Face

Umm… Reducing the scope of your project is just not an acceptable solution, is it?…

Come on, can’t we have it all?

Yes, actually. With Neural Machine Translation and Machine Translation Post-Editing (MTPE) you can! It won’t do magic, but…

MTPE Benefits

  • Instant translation regardless of word count
  • Similar per-word rate across languages
  • Option to post-edit only part of your content

MTPE Drawback

  • Upfront investment in engine acquisition and training

Core goal: Quality

In the center of the Project Management Iron Triangle is quality. That’s the one core goal we cannot compromise.

When considering MTPE, however, quality is a constraint in itself. If the MT output is good, it will take the Post-Editor X hours to review it; if it’s bad, it will take the PE X+N, which will obviously affect the per-word price.

Speaking of which, a Post-Editor could charge either by the word or by the hour. Charging by the hour benefits the Post-Editor because he/she can take the necessary time to really perfect the MT output. Charging by the word benefits you because you would always know what you will pay.

A linguist can also charge MT quality-based rates. If that is the agreement, to calculate the MTPE per-word rate, the linguist would:

  • Post-edit for an hour
  • Note down the total number of post-edited words (source text):
    • e.g. 500, which is double the “regular” productivity
  • Divide their hourly rate by the number of edited words:
    • e.g. $40 ÷ 500 = $0.08/word

How to Train Your MT Engine

Easier than training your dragon! And better news: over time, it gets even easier.

One of the benefits of automation is the lower number of touches and therefore, fewer points of failure. Poor-quality MT output washes away the benefits of time, cost and final quality as it turns the project into a high-touch one. Therefore, training your engine well is not optional.

The road to great MT output ends in a loop:

 

Putting It All Together

So what’s the bottom line? How much would you save on that 20,000-word project if your MT were up, trained and running? Let’s put the numbers together for a “regular” MTPE project applying best-practices, that is, using one single Post-Editor per language:

Translation 100,000 words x $0.00 / word = $0 1 day
MTPE 100,000 words x $0.08 / word = $8,000 5 days + 1 PM day
Totals     $0.08 / word   $8,000 7 days

And… Compare it to the fully human rush project:

Human Rush 100,000 words x $0.20 / word = $45,000 6 days
MTPE Regular 100,000 words x $0.08 / word = $8,000 7 days
Savings     $0.12 / word   $37,000 -1 day

Need I say more? With MTPE, you get practically the same turnaround as a rush human translation (+1 day), save 82% ($37,000) and obtain human-level quality.
Project Management Iron Triangle, consider yourself broken.