The other day we introduced a blog post about myths some translation users believe—and are led to believe by translation providers—about the translation business. Today we will focus on the translation process, particularly as it affects translation quality, pointing out some misunderstandings about the process. Some of these misconceptions are related to the highly vaunted MT (machine translation) capability sold by some companies, but others apply as well to translation done in the conventional way, using human translators.
Misunderstanding One: Machine Translation Has Taken over from Humans.
I recall a day in the mid-1980s when a company developing MT here in Japan (Bravice International) apparently convinced Mainichi, one of the three largest daily newspapers here, to run a substantial article announcing that automatic translation had finally been developed. Well, translation of a sort.
The article was very similar to a press release from the company, and at least people in the translation business realized that MT was not even close to being poised to take over from humans just yet. But shortly thereafter I did have some people comment to me that I might be out of work shortly, what with machine translation taking over. That was 35 years ago. Not much has changed, MT has not taken over, and we don’t yet feel threatened. We are amused, however, by some misconceptions about MT harbored by a small portion of potential clients, but those are easily overcome. Our existing clients are a bit wiser.
Misunderstanding Two: Well, at least machine translation has eliminated the need for human translators to do the actual translation; humans can just fix what the MT got wrong later.
The prospect of eliminating expensive and slow human translators has long been implied by people selling MT. Alas, that has not yet been achieved. Humans (and more importantly, human translators) are still required to rescue the low-quality output of machine translation systems (and to rescue the people who thought they could rely on MT). More on this below.
Misunderstanding Three: A translation that is 95% accurate, be it from a human translator or a machine translation system, only requires a small additional investment to make it a 100% accurate translation.
People who believe this should try machine translation and tell their human editing/rescue team that, to find and correct the translation problems, they have been allotted only 5% of the time that they would have needed to translate the Japanese-language source text from scratch, without the “aid” of machine translation. That clearly will not work. In fact, it probably will fail even if you give the human rescue team 50% of the from-scratch translation time to fix the MT output.
Often the rescue team (made up of people who are optimistically referred to as post-editors) will need to spend nearly as much time fixing a poor translation from an MT system (or a low-skilled human translator) as they would have spent in producing a translation from scratch. And even then, because of the desire on the part of the human repair team to give the benefit of the doubt to the machine translation (or poor human translator), the finished product is still likely to fall short of the quality of a translation produced from the ground up by a qualified human translator. There are several problems at work here that might not be obvious to a first-time user of MT.
The above misunderstanding has three aspects, related, respectively, to translation quality metrics (or, more precisely, the lack thereof), the ability of a human translator to find errors, and the questionable advisability of using a translator—human or machine—capable of only 95% accuracy.
Metrics. How does one quantitatively assess the accuracy of a translation to start out with? And what does 95% accuracy mean when even five percent of the words mistranslated can make a translation totally wrong and unusable.
What’s more, finding even small parts of a translation that make it a bad or unusable translation is a decidedly non-trivial task. There are no automatic methods that obviate human intervention in either assessing a percentage of mistranslation or evaluating the usability of a translation, regardless of whether the translation was done by software or by a human, or to find the N% that was incorrectly translated, let along correct the errors.
The most commonly cited standard that has been established specifically for translation, ISO 17100, addresses only the translation process, dealing with the administrative processes of executing and providing translation. It lacks guidance regarding either how to assess the quality of a translation or the definition of a good translation. Addressing those tasks continues to require a human translator. To be sure, the above-noted process-directed standard apparently calls for a translator’s work to be checked by another translator, but the issue of the abilities of those translators remains unaddressed, as does the definition of a good translation.
Finding errors. It is highly unlikely that a non-translator (i.e., inexpensive) “post-editor” (a distinct misnomer in most cases) is going to easily find the “mistranslated 5%.” That is a task for a translator. And even a human translator will have to go through both the source document and the translated document carefully to find errors. This is certainly the case when using machine translation, but it is just as true when checking and correcting a translation done by a poor human translator.
The result of that effort might very well be that the errors found would be of no consequence to some particular purpose of the translation. But learning even that ultimate outcome will require a full check through the source text and the translated document by a qualified human translator. Ultimately, if you need a high-quality translation, you are very unlikely to gain much in speed or economy by having a poor translation done first, whether the translator is made of flesh and bones or of software code.
Should you even consider using a 95% accurate translator? Is a translator, human or machine, only capable of producing 95% accuracy usable? People in the translation field know the answer is a resounding no, and I know no human translators who boast of only 95% accuracy or who think that 95% accuracy (whatever that might be taken to mean) would be sufficient. Regardless of the oohs and aahs that an MT system gets when it reaches that level, being impressed with an MT system reaching 95% accuracy merely demonstrates that the bar has been lowered to allow or promote use of technology. The lowering of standards to make use of wonderful modern technology is not unique to machine translation, of course, but that is a topic for another day and another article.
Because the next two misunderstandings are somewhat interrelated, I will discuss them together.
Misunderstanding Three: If a document is only something like an email, it can be translated satisfactorily by an MT system.
Misunderstanding Four: MT systems cannot translate literature, but can handle non-creative documents such as found in industry.
Over decades of observing the marketing of MT, we have repeatedly heard that MT can’t, of course, translate a novel, but (sometimes left as an unspoken implication) it can translate non-literary texts.
It is certainly correct that you should not attempt to use MT to translate a novel. But that in no way should lead you to conclude that MT can be generally used without ill effects to translate texts that are not literary or creative in nature, and it is just such non-literary texts that make up the overwhelming portion of commercial Japanese-to-English translation.
The notion that Japanese-language texts produced in the business and technical/industrial domains are not informed and influenced by Japanese culture is far from being correct. The very act of writing a message to someone, even in a business or technical/industrial domain, is influenced by the context and the culture shared by the writer and the reader. Neither a machine translation system nor a human translator who has not lived in one of the source- and target-language cultures (commonly the case these days with discovery document translation) is up to the task.
Take, for example, the case of email exchanges that need to be translated in an antitrust matter. Some common expressions used in everyday business and normal life in Japan, when placed in the context of people accused of antitrust behavior, can take on (or be purported to take on) a significance far beyond their common use as boilerplate formalities in everyday email exchanges. The translation of such communications without taking context into consideration could be misleading at best.
In such cases, even humans have difficulty translating the source texts. Machine translation systems, totally unable to understand culture and not caring about context, are breathtakingly bad at handling such translation tasks. I have seen this time and time again, including with documents presented to witnesses in depositions. And a non-translator human rescue squad is also in serious danger of failing if they do not both know the languages and understand and apply the appropriate contextual and cultural clues. The bottom line is that these texts should be translated from the start by qualified human translators aware of and able to understand the context of such communications. The need to know and consider context has been with us since people started translating foreign languages; the appearance of machine translation has not changed that. The blinding speed of MT and ostensibly low cost of MT should not blind translation consumers to the pitfalls.
As noted in a recent article, a human translator with neither the source language nor the target language as their native language is almost as limited as a machine translation system when confronted by texts written in the real world.
Misunderstanding Five: Translators are Linguists.
Translators are sometimes characterized as linguists. Some translators are indeed linguists. But many good translators are not linguists, and many linguists are not translators and would be poor translators, even if they wanted to be translators, and many do not. Additionally, merely being fluent in two languages does not give one the ability to translate well. Translation skills need to be acquired separately from (and can be acquired without) formal training in the field of linguistics, although linguistic knowledge is often acquired in the process of learning a language and learning translation, these being two processes that are essentially separate from the process of learning the formal discipline of linguistics.
Misunderstanding Six: Machine Translation is Very Fast.
Here, finally, is a misunderstanding with a bit of truth in it. Fast, yes. The questions remain as to how much quality you need, how tolerant you are of serious errors, and whether you need accountability for the accuracy and appropriateness of the translation.
Quality. If the level of quality produced by machine translation is “good enough” for your application (for example, a quick gist translation, followed by a proper translation done from scratch by a responsible and accountable human translator afterward if the content is judged to be important), then it is perhaps “good enough.” But it is foolishly optimistic to hope that quickly produced MT output can be brought in a short amount of time up to the quality needed to trust its accuracy. And who could attest to even 95% accuracy of a machine-produced translation, when we don’t even have a definition of 95% accuracy?
Risks. Anything worth spending money to have translated warrants translation to some expected level of accuracy and suitability.
The largest Japanese-to-English translation demand in the US is for civil litigation involving Japanese entities. Litigators need to assess just how bad a translation could be and still be acceptable as a translation of a Japanese discovery document, and should remember that both machine translation systems and low-quality human translators are quite capable of producing unbelievably bad translations. If we extend the discussion to translations in the medical and pharmaceutical fields, the risks extend to the health and wellbeing of many persons not involved in the translation process.
Accountability. Placing a machine translation system in the translation loop, even with an after-the-fact human damage control team, blurs the chain of accountability. And, of course, it is folly to think that anybody would be foolish enough to declare that a raw machine-produced translation is reliable and usable to N% (where N is even a low number) accuracy.
Although a human translator that is identified can be deposed and examined about a translation they did, you are not going to be able to depose a collection of computer code, and the people using the MT system cannot but trust as an act of faith that the translation it produces is usable. That is a misplaced belief in almost all cases, based on observing a large amount of machine-translated texts.
Summing up, unless you have a high tolerance for serious errors and the damage they can cause, a human translator is still very much an essential part of the process of producing the translations you need. And the humans executing translations need to be highly qualified, with subject expertise and real-world knowledge, two areas in which machine translation has yet to make a sizable dent. In a future article, I will discuss the making of a translator, something about which there are also some misunderstandings, including misunderstandings by some people aiming to be translators or already working in the field.