What is Holding the Wider Adoption of Machine Translation (MT) back?

I have been seriously distracted the last few weeks by the World Cup which is really the only sporting event I really ever connect to anymore. This year I watched more games (in HDTV too) than I have since my childhood and I am glad that the clearly better team won the final. My opinion on the real winners: Germany, Uruguay, Spain and Ghana, the biggest and possibly worst loser in World Cup history: The Netherlands. 

Anyway, there is an interesting discussion in the LinkedIn MT forum started by Lori Thicke that I thought was worth summarizing and highlighting. The original question she poses is: Why is MT adoption so slow, given the clear and real benefits that she has seen in her own experience with MT? I would recommend going through the whole discussion as I am only selectively paraphrasing some of the comments (IMO the best and most informative) and the comments are worth reading in their original state. This is my summary which is in no particular order.

The Reasons Behind Slow Adoption

• Translator Resistance (could it be because many have direct experience with low quality over-hyped MT or insufficient training?).
• Lack of Training on how and when to use MT.
• People who think that MT is a replacement for human translators (especially those who expect the same quality as humans). Consumers need to understand that MT + Human is the construct necessary to get the best quality.
• Too many eMpTY Promises from MT vendors who claim to have the silver bullet and deliver mud. 
• There is a lot of low quality and erroneous advice (mostly opinion) that is outdated and obsolete on translator bulletin boards and forums that malign and present MT as a threat or a crappy technology.
• Very few understand that customized MT engines are quite different from the free online engines at Google, Microsoft and Babelfish.
• MT has been complex and expensive.
• The poor quality of communication between MT vendors (many who feel offended by the quality criticism) and translators (who feel threatened by lower costs and bad quality and reduction of their value and role).  Marco Cevoli has provided a presentation on this conflict here from a slide deck presented at a ProZ conference.

Bob Donaldson points out the disconnect between the vendors and the clients in the translation production chain (Buyer > LSP > Translator) and points to three reasons for this disconnect:

1. Different understanding and expectations of translation quality through the chain
2. Lack of Translator (Vendor) Acceptance because of poor project management and skill mismatches in the editing and cleanup (post-editing) process.
3. Inability to accept risk and manage it in a way to meet customer objectives, and without expecting the translators to bear the burden of poor quality MT systems. LSPs should be willing to share benefits as well as risk with translators as MT systems improve and ensure that they ACTUALLY raise productivity.

• Translators may also not be the best post-editors and users will need to find how to find the right balance here. Sometimes less skilled people would be a much better fit for post-editing tasks. The world where MT is commonplace will have many more linguistic roles for people with linguistic skills.
• I think that better definitions of deliverable MT linguistic quality would also go a long way to making the technology more widely accepted.
• Many people do not realize that MT + Human Post-editing only makes sense when there is an actual measurable productivity advantage. Don’t use MT if this is not true and getting an MT engine to this state requires investment, effort and know-how.

Joachim Van den Bogaert also made some interesting points:

"Using MT and post-editing requires a paradigm shift. We see that, for now, translation jobs primarily come in project packages: the typical cycle is: engineering (pre-processing) -> translation -> engineering (post-processing) -> QA -> corrections -> QA -> final delivery. MT is at its most efficient in continuous projects: add content -> MT (with included automatic processing) -> post-editing -> delivery. Instead of focusing on delivering localized content at a specific point in time, MT allows you to consider having ALL your content localized ALL the time. Indeed, a well implemented MT chain, eliminates initialization costs, which allows you to have minor changes and additions to be translated instantly. But this requires a different mind set from (traditional) localization consumers."

• Open source MT like Moses is difficult to use productively without a set of other tools (segmentation, alignment, corpus cleaning and analysis etc..).
• MT systems have very different requirements and controls and thus it is difficult to integrate with production systems and have common and consistent pre and post processing.
• In terms of mindset, I feel that MT should be considered as moving to a production line where some things can be built more efficiently as opposed to the craftsman approach that many have today.

Just because 90% of the professional translation industry makes disparaging remarks about MT on a regular basis does not make the technology irrelevant or unusable. Change often comes from the few, who solve many of the problems we have listed above and develop the new models for success. There are now a number of firms in the professional translation world who “get it” and they are likely to pave the way and build competitive advantage through informed and intelligent use of MT technology.

Alon Lavie points out that MT is the only way that new high-value dynamic content like customer support and user generated content (UGC) can be done, I believe this can be done at much higher quality levels when done in partnership with LSPs who are informed about the linguistic strategies to improve and refine MT engines. I have written about this type of content that presents a major opportunity for the professional translation industry. (I think so anyway). However, this will require new mindsets and new translation production strategies. We are in the early stages of this today.

So while many of these problems and issues remain unresolved, there are more and more initiatives that are attempting to address at least some of these problems. We are seeing much these days about how the translation production chain needs to evolve to handle the more dynamic and continuous flow of content that is increasingly valuable to translate. MT will be critical here.

It is clear that we should attempt to address all or as many of these problems listed above as possible but a few stand out for me.

•  More education and more information about the current MT technology 
• More information to help users assess where it makes sense and where it does not. 
• Direct contact and hands on experience with a customized MT system. 
• Better integration with the other translation production infrastructure. 
• Development of win-win scenarios in MT + Post-editing work processes and compensation so that all participants share in the benefits of improved production capacity.

So while there are many difficulties there are also many new opportunities that provide new vistas for professional translation market leaders. And while it is clear that MT on it’s own is not the solution for every translation problem, it is also increasingly clear that it is a technology that will have a long-term role in making large volumes of high-value content multilingual at a cost and turnaround time that is viable and furthers business objectives.


And a parting note: for those of you think that MT is not a mainstream technology yet, consider how many people are using free online MT engines each and every day. I don’t have accurate data (please let me know if you do), but I estimate that anywhere from 10 –20 million people translate a web page, a document, a phrase or a word every single day in hundreds of language combinations. (About 2.5B to 5B words daily!) I think we can all safely assume that the thirst for information and knowledge in the growing human online population, has an energy and momentum that is far stronger than the inertia in the professional translation industry. It will drive this technology forward and large volumes of content will get translated at amazingly high quality levels. We are already seeing examples of this.

My advice to you: Wake up and smell the coffee (as they say in America) because often "good enough" is good enough. And I have to again use the quote that June Cohen, Executive Producer and Curator of the amazing TED Talks made at SXSW this year (even though it may seem naive and idealistic to some) when she was asked "What technology would you like invented? Or uninvented?"


"Instantaneous, accurate translation online. Nothing would do more to promote peace on this planet."