A common question faced by newer translators is how slow they should translate, which is a corollary of how fast they should translate. My answer is to suggest translating at a speed that maximizes value for the client, because the main reason a translator is hired is to create value, even if they are paid by the word. Translators who try to maximize quality will nonetheless come up against a soft cap on quality where increased time investments yield only marginal quality improvements. The only way to cover the gap between ability and accuracy is to involve genuine expert reviewers, as additional time investments in research will rapidly become cost-ineffective. This also implies that continuous investment in learning and improving overall expertise to develop an effective base of knowledge is an effective long-term strategy for providing highly accurate translation work.
Newer Translators Should Work Slower
There is a lot of interest among organizations to take new translators and throw them at Google Translate, having them translate at 1,000 words per hour from their first day on the job. This is something I tend to see in law firm environments that have set up their own translation departments, and also in fields like banking AML. The result of having inexperienced translators working this way is that they rarely do more than make a few superficial changes to the machine translation. Quite often, the translation quality is actually significantly worsened compared to the original machine translation, but non-experts nonetheless deem it acceptable or at least have no complaints. From the management perspective, this workflow allows a legal matter involving 300,000 words of translation to be completed for $3,000 total, which is still quite a lot of money.
The situation nonetheless begs an important question: if the translator isn’t adding anything of value beyond what machine translation provides, why pay them a salary? In organizational behavior terms, the answer is that everyone knows machine translation alone is dangerous and unreliable, especially in legal contexts where its accuracy rate is maybe 25%. Adding a human to the mix changes the label applied to the workflow because it’s no longer unsupervised machine translation. The organization is investing substantial amounts of money to make it look like the workflow is being done right. But what would happen if the organization cared about the reliability and quality of the work?
To achieve improvement over machine translation levels, a newer translator will need to slow down significantly from that 1,000 words per hour benchmark. For students, this might mean a rate of 100 words per hour, while apprentice-level workers may manage 200-300 words per hour. During this time, the translator will need to dedicate additional time to learning the vocabulary used in the field and understanding how the terminology corresponds between the two language fields. Difficult items will inevitably come up and, rather than merely skipping them and relying on a machine’s best guess, the translator will spend time doing terminology research. As you can guess, this kind of turtle’s pace can easily lead to runaway cost overruns.
Translation Quality Soft Cap
The litmus test for whether runaway cost overruns are occurring with newer translators can be described by the translation quality soft cap phenomenon. If you are familiar with economics, this is a variant of the diminishing marginal returns theory, which says that for each additional unit of effort invested, returns begin to diminish. In the context of translation, this principle is described by concepts such as Heap’s Law, which states that a lot of vocabulary can be encountered quickly in a text but growth plateaus quickly thereafter.
A translator working on data privacy, for example, can read a short primer on the GDPR and quickly learn the basics of data privacy regulation. About 20% of the material would be context-specific, such as data privacy discussions around blockchain-based data exchanges, which would require some additional learning. Perhaps 1-2% of the vocabulary will be so obscure, ambiguous, or novel that a newer translator could spend 2 hours researching a single word before arriving at a reliable answer that an expert would know off the top of their head.
Herein lies the concept of the translation quality soft cap: as steps are taken to marginally increase translation quality, the cost of each improvement also increases. In an observation of student interns in China working on either high-importance law school academic translations or for the market, students produced work at a cost of about $0.01/Chinese character for low-quality translation agencies. In contrast, the “100% quality” cohort took about 12 times as long, translating at a cost of $0.12/Chinese character. Moreover, many of the final translation decisions were still just guesses or made up. To actually finish the job, these student translators would likely need to spend twice as much time.
Lone Wolf Translators Do Poorly on the Soft Cap
These translators would be much more efficient if they worked with a more senior, more expert translator with deep knowledge in the particular domain. This approach is used in virtually every industry in the world except translation; for example, accountants, lawyers, electricians, and resident physicians all do valuable work with the oversight and supervision of senior experts. In translation, this tends not to be the case for two reasons. First, most translators go their entire careers without developing any subject matter domain expertise, which is a result of the high transactional costs involved in translation—finding and matching the appropriate person for the job. Second, professional project managers are often needed to route jobs to the appropriate person, and protocols require the junior/senior team compositions to be arranged by the project manager. In practice, this is basically impossible. The expert chosen may not actually have the appropriate skills, and the junior person may refuse to listen to the expert or simply cheat by copying and pasting low-quality AI output.
In-house teams offer a much better opportunity to achieve genuine teamwork by matching junior translators with more experienced ones who can provide knowledge and guidance about highly difficult terminology. Nonetheless, there is a challenge here in assessing who has the relevant skills. When compared to law or finance, the top firms in those fields typically only allow the top 5% of performers to remain onboard. In translation, things work a lot differently because translation services consumers tend to be very unsophisticated. Thus, superstar translators are usually rewarded the same as uncommitted, mediocre translators. You can see some evidence of this in translation companies’ marketing materials that argue that ATA or CIOL Certification doesn’t mean anything, and in translation management courses that urge hiring the cheapest translator possible because clients won’t pay attention to quality. There may be competition in most of the translation industry, but if anything, it’s competition to put nice packaging around scams.
My advice to in-house teams is to ignore what the translation headhunters are saying and instead look for association-based qualifications and recognition when looking for genuine experts for this role.
Beating the Diminishing Returns Curve with Long-Term Knowledge and Study
Junior translators may find themselves forever stuck doing slow, low-quality translation work or making a career of passing off machine translations as their own. To do any genuine, high-quality work, they will need to rely on more expert translators for corrections. Moreover, most will never gain the skills necessary to pass a certification exam, despite high interest in doing so—ATA Certified translators in Chinese<>English make 500% of what uncertified ones do. However, there is a way to beat the diminishing marginal returns curve: by separating the translation process, to some extent, from the learning process. Most translators expect 100% of their learning to happen on the job, but this isn’t realistic in this industry. Vocabulary and underlying background knowledge are often only available through engaging with language materials in the same way that professionals in the relevant industry engage with them. That is, looking things up in dictionaries won’t lead to useful linguistic knowledge. For example, if you want knowledge about contracts, then you’ll need to read books about contracts and related substantive law, such as civil procedure or intellectual property.
Anecdotally, I recall a new legal translation professor in China, who came from several years of corporate law practice, speaking at a conference last year about his practice experience. He said that, when in the USA, they constantly discussed and referenced entities as a very broad legal concept. However, in China, the term “entity” is virtually absent from English texts, despite corporate lawyers doing the exact same things in the exact same ways. He raised the possibility that perhaps translators should consider addressing entities more explicitly in their translations.
The knowledge gained from books alone is not effective, rather, it needs to be applied specifically to legal translation tasks. In the above example, we can see that this translator understood very clearly, from many years of legal practice experience, what an “entity” is, and that knowledge is indeed essential to translating business law documents. A next step would be to identify what specifically constitutes an entity in Chinese law, and indeed in which context. What makes this task particularly difficult is that the kinds of things usually considered entities in China have traditionally been highly politicized in each of the various Communist revolutions and reforms.
A good example of this is the word “danwei” in Chinese, much loved by Sinologists, traditionally translated as a “work unit” or “unit” but later on evolved to refer to different types of organizations. When the danwei were dismantled in the 1980s, they were replaced with business entities called qiye, a name picked out of market economics textbooks describing free “enterprise.” Analyzing this particular example is very involved, so much so that we’ve devoted an entire article to it, which you can read here.
In the question addressed by this article, it’s enough to know that a translator’s knowledge needs to improve through continuous learning, which helps bring down error rates. A new translator might start out at a speed of 200 words per hour and gradually move up to 400 or 500 words per hour before hitting the quality soft cap. Working with experienced translators and striving to become one of those experts is essential to improve further.
