Friday, July 25, 2014

Understanding The Drivers of Success with the Business Use of Machine Translation

We have reached a phase where there is a relatively high level of acceptance of the idea that machine translation can deliver value in professional translation settings. But as we all know the idea and the reality can often be far apart. It would be more accurate to say this acceptance of the idea that MT can be valuable, is limited to a select few among large enterprise users and LSPs (the TAUS community) and has yet to reach the broad translator community who continue to point out fundamental deficiencies in the technology or share negative experiences with MT.   So while we see growth in the number of attempts to use  MT, as it has gotten mechanically easier to do, there is also more evidence that many MT initiatives fail in achieving sustainable efficiencies in terms of real translation production value.

In a typical TEP (Translate-Edit-Proof) business translation scenario, A "good" MT system will provide three things to be considered successful:

1) Faster completion of all future translation projects in the same domain
2) Lower cost/word than doing it without the MT system
3) Better consistency on terminology especially for higher volume projects where many translators need to be involved

All of this should happen with a final translation delivered to the customer that is indistinguishable in terms of quality from a traditional approach where MT is not used at all.

It is useful to take a look at what factors underlie success and failure in the business use setting, and thus I present my (somewhat biased) opinions on this as a long-time observer of this technology (largely from a vendor perspective). I think that to a great extent we can already conclude that MT is very useful to the casual internet user, and we see that millions use it on a regular basis to get the gist of multilingual content they run into while traveling across websites and social platforms. (e.g. I use it regularly in Facebook.)

What are the primary causes of failure with MT deployments in business translation settings?

Incompetence with the technology: The most common reason I see for failed deployments is the lack of understanding that the key users have about how the technology works. Do-it-yourself (DIY) tools that promise that all you need to do is upload some data and press play are plentiful, and often promise instant success. But the upload and pray approach does not often work to provide any real satisfaction and business advantage. Unfortunately the state of the technology is such that some expertise and some knowledge are required. The translators and post-editors who have to work with the output results of these lazy Moses efforts, are expected to clean-up and somehow fix this incompetence usually at lower wage rates. And thus resentment grows and many are speaking up frequently in blogs and professional forums about bad MT experiences. Those that have positive MT experiences rarely speak up in these forums since the work is not so different from regular TM-based translation work and MT is often regarded as just another background tool that helps to get a project done faster and more consistently. MT output that does not provide cost and turnaround advantages for translation work cannot be considered to be useful for any professional use. Thus, a minimum requirement for using MT in professional settings is that it should enhance the production process.


Lowering cost is the ONLY motivation: The most na├»ve agencies simply assume that using MT, however incompetently, is a way to reduce the cost of getting a translation project done, or more accurately a way to justify paying translators less. Thus the post-editors are often in a situation where they have to clean up low quality MT output for very low wages. Given that we live in a world where the customers who pay for professional translation are asking for more efficient translation production i.e. faster and cheaper, agencies are being forced to explore how to do this, but this exploration needs to happen from a larger vision of the business.  As Brian Solis points out, using technology without collaboration and vision is unlikely to succeed (emphasis mine).
"That's the irony about digital transformation, it doesn't work when in of itself technology is the solution. Technology has to be an enabler and that enabler needs to be aligned with a bigger mission. We already found that companies that lead digital transformation from a more human center actually bring people together in the organization faster and with greater results," Solis says. “When technology is heralded above all else, there becomes an even greater disconnect between employees (translators)  and the challenges that their business is trying to solve.”
What many LSPs fail to understand is that their customers are asking for ongoing efficiencies, and new production models to handle the new kinds of translation challenges they face in their businesses. They are not just asking for a lower rate for a single project. Agencies focused on the bigger picture are asking questions like how MT can enable them to achieve new things and what's different about their customers needs today versus yesterday. With the right MT strategy in place, technology becomes an enabler, not the answer and enables agencies to build strong long-term relationships with customers who could not get the same price/performance with another agency that does not understand how to leverage technology for these new translation challenges. Agencies must evolve and reimagine their internal process, structure and culture to match this evolution in customer behavior among their own employees and translators.

No engagement with key stakeholders: Many if not all the bad MT experiences I hear about have one thing in common. Very poor communication between the MT engine developers (LSP), the customer and the translators and editors. MT is as much about new collaboration models as it is about effective engine development, and collaboration cannot happen without open and transparent communication, especially during the initial learning phase when there is a great deal of uncertainty for all concerned. If this communication process is in place in the early projects, it enables everybody to rise together in efficiency, and gets easier and more streamlined and more accurately predictable with each successive MT project. The communication issue is quite fundamental and I have tried to address and explore this in a previous post

What are the key drivers of successful deployments of MT?


Expert MT Engine Development: The building of MT engines has gotten progressively easier in terms of raw mechanics, but the development of MT engines that provide long-term competitive advantage remains a matter of deep expertise and experience. If as an LSP, you instantly create an MT engine that any of your competitors could duplicate with little trouble, you have achieved very little. Developing MT systems that provide long-term production advantage and a real competitive advantage is difficult, and requires real expertise and experience. The odds of a developer who has built thousands of engines producing a competitive engine are much higher than someone who uploads some data and hopes for the best. Skillful MT engine development is an iterative process where problems are identified and resolved in very structured development cycles so that the engine can improve continuously with small amounts of corrective feedback. Knowing which levers to pull and adjust to solve different kinds of problems is critical to developing competition beating systems. Really good systems that are refined over time are very difficult to match and will continue to provide price/performance advantages over the long-term that competitors will find difficult to match.


Engaged Project Managers and Key Translators: The most valuable feedback to enhance MT system output will come from engaged PMs and translators who see broad error patterns and and can help develop corrective strategies for these errors. Executives should always strive to ensure that these key people are empowered and encourage them to provide feedback in the engine development process. For most PMs today, MT is new and an unknown and unpredictable element in the translation production process. Thus in initial projects, executives should allow PMs great leeway to develop critical skills necessary to understand and steer both the translators and the MT engine developers. These new skills are very key to success and can help build formidable barriers to competition. While very large amounts of high quality data can sometimes produce excellent MT systems, a scenario where you have a a good project manager steering the MT developers and coordinating with translators to ensure that key elements of an upcoming project are well understood, will almost always result in favorable results other things being equal, especially with challenging situations like very sparse data or when dealing with tough language combinations. 

Communication and collaboration are key to both short and long-term success. The worst MT experiences often tend to be with those LSPs (often the largest ones ) where communication is stilted, disjointed and focused on CYA scenarios rather than getting the job done right. Successful outcomes are highly likely when you combine informed executive sponsorship, expert MT engine development and have empowered PMs who communicate openly and frequently with key translators to ensure that the job characteristics are well understood and that outcomes have a high win-win potentiality. Even really good MT output can fail when the human factors are not in sync. Remember that some translators really don’t want to do this kind of work and forcing them to do it is in nobody’s interest.

Fair & Reasonable Compensation for Post-Editors: I have noted that a blog post I wrote on this issue almost 30 months ago still continues to be amongst the most popular posts I have written. This is an important issue that needs to be properly addressed with a basic guiding principal, pay should be related to the specific difficulty of the work and quality of the output. So low quality output should pay higher per word rates than very high quality output. This means that you have to properly understand how good or bad the output is in as specific and accurate terms as possible since people’s livelihoods are at stake. This accuracy can be gauged in terms of average expected throughput i.e. words per hour or words per day. You may have to experiment at first and be prepared to overpay rather than underpay. Make sure that translators are involved in the rate setting process and that the rate setting process is clearly communicated so that it is trusted rather than resisted. Translators should also ask for samples to determine when a job is worthwhile or not. The worst scenario is where an arbitrary low rate is set without regard for the output quality, and typically in these scenarios incompetent MT practitioners always tend to go too low on the rates, resulting in discontent all around. 


Real Collaboration & Trust Between Stakeholders: This may be the most critical requirement of all as I have seen really excellent MT systems fail when this was missing. Translation is a business that requires lots of interaction between humans with different goals and if these goals are really out of sync with each other it is not possible to achieve success from multiple perspectives. Thus we often see translators feel they are being exploited or agencies feeling they are being squeezed to offer lower rates because an enterprise customer has whipped together some second rate MT system together with lots of noisy data for them to “post-edit”. When the technology is used (actually misused) in this way it can only result in a state of in equilibrium that will try to correct itself or make a lot of noise trying to find balance. This I think is the reason why so many translators protest MT and post-editing work. There are simply too many cases of bad MT systems combined with low rates and thus I have tried to point out how a translator can make an assessment of a post-editing job that is worth doing from an economic perspective at least. 

Perhaps what we are witnessing at this stage of the technology adoption cycle is akin to growing pains, like the clumsy first steps of a baby or the shyster attempts of some agencies to exploit translators as some translators have characterized it. Both cases are true I feel. And so I repeat what I said before about building trusted networks as this seems to be an essential element for success.

The most successful translators and LSPs all seem to be able to build “high trust professional networks”, and I suspect that this will be the way forward i.e. collaboration between Enterprises, MT developers, LSPs and translators who trust each other. Actually quite simple but not so common in the professional translation industry.

There seems no way to discuss the use of MT in professional settings without raising the ire of at least a few translators as you can see from some of the comments below. So I thought it might be worth trying to lighten the general mood of these discussions with music. I chose this song carefully as some might even say the lyrics are quite possibly the result of machine translation or not so different from what MT produces. As far as I know it is just one example of the poetic mind of Bob Dylan. If you can explain the lyrics shown below you are a better interpreter and translator than I am. Musically this is what I would call a great performance and a good vibe. So here you have a rendition of Dylan's My Back Pages on the Empty Pages blog.

Crimson flames tied through my ears
Rollin’ high and mighty traps
Pounced with fire on flaming roads
Using ideas as my maps
“We’ll meet on edges, soon,” said I
Proud ’neath heated brow
Ah, but I was so much older then
I’m younger than that now

Half-wracked prejudice leaped forth
“Rip down all hate,” I screamed
Lies that life is black and white
Spoke from my skull. I dreamed
Romantic facts of musketeers
Foundationed deep, somehow
Ah, but I was so much older then
I’m younger than that now

Girls’ faces formed the forward path
From phony jealousy
To memorizing politics
Of ancient history
Flung down by corpse evangelists
Unthought of, though, somehow
Ah, but I was so much older then
I’m younger than that now

A self-ordained professor’s tongue
Too serious to fool
Spouted out that liberty
Is just equality in school
“Equality,” I spoke the word
As if a wedding vow
Ah, but I was so much older then
I’m younger than that now

In a soldier’s stance, I aimed my hand
At the mongrel dogs who teach
Fearing not that I’d become my enemy
In the instant that I preach
My pathway led by confusion boats
Mutiny from stern to bow
Ah, but I was so much older then
I’m younger than that now

Yes, my guard stood hard when abstract threats
Too noble to neglect
Deceived me into thinking
I had something to protect
Good and bad, I define these terms
Quite clear, no doubt, somehow
Ah, but I was so much older then
I’m younger than that now