In the first installment of our Machine Translation series, “Introduction to Machine Translation,” we provided a brief history of MT, presented an overview of the technology and touched on its use in business today.
Machine Translation has endured a questionable reputation over the years, and it is understandable. For a profession/craft (translation) that touts accuracy and cultural appropriateness as the two main indicators of quality for final output, relying on a tool that only gets you a percentage of the way there, a percentage of the time seems the antithesis to what you are trying to accomplish.
However, each year it seems MT is improving – when in the right hands and applied to the right content with the right process.
In our second installment of the MT series, we will discuss three myths of MT in effort to dispel some common misconceptions and help provide further clarification of its applicability.
Myth #1: “Machine Translation will be useful in the future, but we’re not there yet.”
No need to climb into your DeLorean and fire up the flux capacitor – the future is here today! But the “current” future with Machine Translation is not quite as advanced as science fiction movies would have us hope.
Today we do see an impressive sample list of well-known corporations and organizations currently investigating and/or using MT technology in some form. Companies include:
- Harley Davidson
- John Deere
What is impressive is that many of these companies and other organizations report that using MT has helped them to cut turnaround times and costs for MT-related projects. Please keep in mind that none of these companies use MT exclusively and some do employ post-editing effort to reach a final product.
Are these companies using MT for the home page of their websites or for the advertisements you see on the subway? Most likely not, but you would find it surprising to hear how companies are applying MT and the feedback they have about it, which leads us to Myth #2…
Myth #2: “Machine Translation quality is so bad, that any translations done by machine are useless.”
We have all heard appalling translations that may have been done by machines. And we all know it’s a bad idea to use Google Translate to do your French homework. However, it would be unfair or misguided to say that Machine Translation quality is so bad that it cannot possibly have any practical use. Let’s look at some real-world examples that should help lay this myth to rest.
Catherine Dove, Quality Lead at PayPal Inc. . probably raised more than a few eyebrows at the 2011 TAUS User Conference when she said, “Human quality is not good enough for PayPal.” (See video from the conference at the TAUS YouTube Channel). PayPal faces a number of challenges when localizing their content into 24 languages. For example, their content does not consist of complete sentences and contains a lot of tags. It is challenging for human translators to handle the tags correctly and they often make mistakes. On top of that, PayPal translators must work within very tight deadline constraints, but applying multiple linguists to the same project leads to inconsistencies. PayPal is getting help from MT, using a process they call “Machine Aided Human Translation.” Catherine Dove explained that, for their content, MT is better at Human Translation at handling tags and the “MAHT” process helps PayPal improve consistency, allows human post editors to focus on style, fluency, voice, brand & tone, and drives internal rework of translation content down by 20%.
At the 2012 TAUS User Conference, a panel of MT users from companies like Microsoft, eBay, Cisco and Dell was asked how they get MT non-believers at their respective organizations to buy into the process and believe that MT quality can be good enough. Their answers were very insightful. (See video from the conference at the TAUS YouTube Channel). Tim Young, Sr. Operations Manager at Cisco, explained that Cisco feels confident enough in the quality of their MT output due to the post-editing process and relationships with their post-editing service providers, that those conversations around quality are giving way to more meaningful conversations about things like time-to-market and the authoring & content management process.
Wayne Bourland, Director of Translation at Dell likened the application of MT in their process to the application of Translation Memory technology; each tool is just a piece of the localization puzzle and the human touch is still there in the form of post-editors. He also brought up an interesting point that we will explore in later blogs: the idea that end-users may not expect “perfect” translation quality and that “good enough” quality can satisfy the language needs of end-users while allowing Dell to focus on delivering the other things their customers need, such as fewer clicks on Dell.com to get to the content they seek.
When you hear it from the lips of companies like these, it is impossible to deny that MT is useful & applicable today and the quality can be acceptable (or above par) in certain scenarios.
Myth #3: “Machine Translation will replace human translators.”
Never say “never”, but it is highly unlikely Machine Translation will replace human translators in our lifetime. Machine Translation is not a disruptive innovation; human translators will not suffer the same fate as film cameras, cassette tapes and floppy disks! There will always be a call for human translators as long as machines are not able to pick up on context, idioms, slang, style, tone or cultural nuance as effectively as the human brain. In fact, who better to run MT systems, than qualified translators?
Let’s take a closer look at “context” – content that proceeds and follows a segment, which influences the meaning of that segment. The word “lie” for example can have many different meanings depending upon the context in which it is used. Recall that a Statistical MT engine works on the basis of probabilities: for each segment of source text, there are a number of possible target segments with a varying degree of probability of being the correct translation. The software will select the segment with the highest statistical probability, independent of contextual information found in nearby segments. Most Statistical MT engines cannot pick up on contextual details.
It appears that translators have nothing to fear, however, most of the pushback against MT seems to come from the professional human translator crowd. Self-preservation is a strong instinct and surely a sizable percentage of the world’s 700,000+ translators feel threatened by this technology. Some may even do their best to discourage its use by proliferating some of these myths we have discussed.
Machine Translation and Human Translation are not mutually exclusive; as we have seen, companies like PayPal, Dell and Cisco use MT as only a step in their localization processes. Perhaps one day more translators will be willing and able to use MT as a tool in their work, just as they use Translation Memory now after years of reluctance.
By now, it should be clear that Machine Translation (1) is in use today and (2) is part of the process that provides corporate users with good or “good enough” quality, a process which also includes (3) human translators, who are and will always be an essential piece of the localization puzzle.
Hopefully this helps to dispel some of the common myths around Machine Translation. Can you think of any other common Machine Translation myths or misconceptions? If so, we invite you to leave a comment!