Think of all the services we routinely seek out, either as part of our day-to-day jobs or in our personal lives — everything from integration of business systems to dry cleaning. How often do we really know or care about who touches the process and what exactly is involved? In many cases, it simply doesn’t matter as long as the result meets our expectations.
We don’t always know every service provider’s process, and that’s okay. It’s enough that the quality is there … right?
During my tenure in the language services business, I’ve found this to be a loaded question when it comes to machine translation (MT). Many clients request content translation with quick turnaround without asking how it gets done. For these buyers, as long as there’s a quality process firmly in place, and the final product makes the grade, it’s a successful transaction. Assuming all conditions are right, a solution involving MT can cut costs for the vendor as well as the client while accelerating delivery and ensuring superb quality.
Of course, it bears repeating that the situation has to be just right. Where we start to get into murky waters is when companies promise translate-edit-proof but deliver MT with post-editing. Linguists using MT behind the scenes can likewise cause headaches for vendors and clients alike, especially if and when errors show through in the translation. Aside from the possibility of low quality, there’s the fundamentally problematic matter of going against client expectations.
All of this raises the issue of whether language service providers (LSPs) should introduce MT with post-editing into the language translation equation as a rule. But is this practice fair to unsuspecting clients? It’s a question that has led to many a furrowed brow for leaders of translation companies the world over.
In the constant race for reduced costs, all translation companies are under pressure to maintain what little margin there is. Clients as well as vendors are forced into a position to do everything possible to increase profitability while keeping up with the massive content volumes and real-time expectations of today. While MT is widely gaining more acceptance for dealing with these conditions, divided opinions still persist on whether to work it into the translation workflow as a matter of course.
Few would disagree that MT with post-editing delivers on the speed front. Whereas a human translator completes on average up to 3,000 words per day, a machine can pump out up to 8,000 words in the same period with MT and post-editing. Cost savings potential is similarly compelling. It’s common to see clients come away with 30% to 60% savings and sometimes more, assuming the process involves the use of well-trained engines.
Still, not everyone agrees about how this tool should be used. Opinions vary wildly within a single organization, to say nothing of the language translation industry and world at large.
Buyer and vendor concerns
Among buyers, the fundamental concern around MT use will always be quality. That is inexorably due to the stigma surrounding the technology. Buyers may wonder: if a machine translates the content, how can it possibly produce high-quality output? Fair or not, it’s a question that often floats in the ether. Education can go a long way in addressing this concern. Buyers may appreciate seeing comparison tests that detail how MT with post-editing has produced the same level of quality as human translation. It’s very possible that the other benefits of increased acceleration and cost effectiveness can swing opinions in a positive direction.
Another valid concern for buyers has to do with process disclosure. No one wants to be in the position of discovering too late the use of MT when that wasn’t part of the deal. Vendors that truly value the relationships they have with their clients should take every measure to be transparent. This could mean that the vendor preemptively explains the entire process or simply makes it a rule to be up front with clients when and if they ask about MT use.
From the LSP side of things, one common concern is translation process ownership. Vendors will need to reach a consensus internally about how they will approach the inclusion of MT with post-editing and whether it makes sense to incorporate it as needed without informing clients. Making and enforcing a sweeping operational change like this may prove challenging if not all voices in the organization are in harmony. Further, production staff and anyone else who deals with clients on a daily basis may worry that the use of MT with post-editing could damage the relationship if or when the client learns about the MT step.
Should clients be informed about MT usage? Unfortunately, I have to answer the prevailing question of whether clients should be informed with a diplomatic statement worthy of a politician who’s courting both sides of the aisle: yes and no. There is no broadly accepted, universal answer to whether clients should be informed about MT use as part of the localization workflow. By the way, a fun game you could try sometime is to pose this question to each person you come across in your office and see how many different answers you get.
But let’s start with common ground. Most of us would probably agree that selling and charging a client for translate-edit-proof but delivering MT with post-editing is dishonest and unethical, to say the least. That said, in some situations it’s acceptable to bring MT into the workflow without discussing it with the buyer. For one thing, a client might simply expect fast, high-quality translations with no desire to know how the team gets it done.
Along those lines, some argue that MT is just another tool in a vendor’s toolbox, so the vendor is not under any obligation to spill the contents of said toolbox. Recently I heard advice from an MT service provider that a vendor should avoid revealing MT use during the sales cycle because companies then ask for a discount right off the bat. However, if a vendor makes it a practice to pass on the cost savings to clients, this cautionary advisement is unnecessary.
Whether or not a vendor informs the client of MT use depends on standard operating procedure, the particular situation and sometimes a client’s localization maturity. The more experienced clients broach the subject voluntarily. Tech-savvy organizations such as Google or Microsoft will likely ask for MT with post-editing right from the outset because they understand — and want to capitalize on — the process efficiencies, return on investment potential and other benefits.
For example, we have been translating technical documentation, among other content, for one of the world’s largest technology companies for many years. From the first discovery meetings, this company requested MT with post-editing in the workflow. This works out particularly well since they have several gigabytes of solid translation memory that we have fed into the MT engines specific to their content and domain. Last year we processed over 15.7 million words through the engines. Considering they get over 80% translation memory leverage in addition to the reduced per-word costs with the MT process, they save a staggering amount on translation initiatives each year.
For companies with less localization experience, it might make sense to educate buyers about the advantages of MT with post-editing, which is everything from faster turnaround to lower per-word rates. In certain situations we might advocate for MT with post-editing, if source content testing shows that it is highly structured, consistent and high quality. In addition, particular verticals, content types and language pairs actually lend themselves to MT use. With a sufficiently trained engine, the quality level meets and sometimes exceeds that of human-performed translation.
To support the education effort, a vendor could conduct parallel process testing. This is simply running the same content through the translate-edit-proof process and also through MT with post-editing, side by side, to compare the results and benefits of the MT route. Of course, a downside to doing this is that it’s expensive, and it goes against the whole idea of increasing profit margins.
In any case, for a buyer who is looking for a faster and less costly solution to language translation without sacrificing quality, vendor-produced case studies and parallel testing can prove to be very persuasive and enlightening.
Toward a culture shift
Even though most would likely agree that MT is getting to be as mainstream as translation memory, if it’s not already, there is still a haze of uncertainty around it. The mere mention of MT frequently brings about a knee-jerk reaction, and not a very positive one. A single negative experience can cause people to view the entire prospect unfavorably. Even so, I continue to see some powerful forces at work right now that will change the direction that the wind blows.
First, MT technology is advancing all the time. As our ability to train MT engines becomes increasingly sophisticated, coupled with more consistent, high-quality multilingual resources to feed into those engines, we will see ever-higher MT quality even before it reaches the post-editors. Our industry is so fundamentally tech-forward that the reality of MT 2.0 really isn’t too far away; in fact, we’re in the midst of it now.
Second, the market continues to dictate the use of this technology. As more buyers come to understand how the process can deliver major benefits, greater numbers will ask for the technology until it becomes common practice. MT will keep steadily rising in prominence and general repute. After all, getting the best of all worlds — lower costs, high quality and increased velocity — is pretty hard to ignore. To get to this point, we will need to continue evangelizing the technology capabilities with buyers when it makes sense. Part and parcel of this will be continuing efforts to run parallel tests on MT with post-editing against the traditional translation process and support the idea of change with solid metrics.
Perhaps the biggest shift, though, needs to occur within the LSP’s own walls. It is not uncommon for executives, production teams, heads of technology and others to disagree on all the points I mention here. Only when we all synchronize on how to approach MT as a part of the total quality process will we begin to really see the fruits of its full potential.