Pricing in the Translation-Memory Workflow
By: Mike Karpa (Freelance)
06 March 2015
Translation memory (TM) and machine translation (MT) continue to change the translation industry. Pricing is evolving, too, as clients, translators, and agencies offer and accept consideration for translation services. But what goes through the minds of the players as we haggle? One way to puzzle this out is to consider how the players view the most extreme TM result: the 100% match.
The value of the 100% match
A 100% match, just to be clear, is when a translation-memory program examines a translation unit or segment in a source text and finds an exact match for it within the translation memory loaded into the program. The program then fills in the translation previously created for that segment into the translation text. (A repetition draws its translation from elsewhere in the current translation.) A novice client may well look at this process and conclude that the value of the 100% match is the conventional per-word rate times the word count: translation that used to cost X dollars can now be produced by simply running the source segment through the program, so the price should drop by X dollars.
Sadly for the client, it has become obvious that a 100% match does not mean a translation has been procured effortlessly. Far from it. Translating these matches involves real work. Meaning is driven by context, for starters. Even when the sentence itself does not change from previous to current source, the meaning might, so translators have to consciously ignore and look beyond segment boundaries (“Translation Memory: Concept, products, impact and prospects,” Gerald Dennett, p. 34). A verb that in one case translates as process, for example, might become treat in another case or handle in a third as surrounding segments change. And grammar is inflected. A noun occurring in a sentence fragment might be shorthand for its related verb. And if a verb, the word may be best translated as the gerund or in a different tense from the match. Gender may also need to be inferred. In fact, a TM program may offer multiple translations for a 100% match, from which the translator must intelligently choose.
Context may also extend through a text. Tables, lists and equations frequently occur in a multi-sentence construct that introduces them, presents them, and makes a conclusion about them. In such cases, the framing construct needs to be translated as a whole. In Japanese-to-English, for one example, this may mean flipping the parts before and after the object, so that the source and translation bear no resemblance to each other. And context extends not just linearly through the text, but also non-linearly (say, vertical relationships between table cells) and discontinuously (between widely separated segments). Ultimately, a translator must apply an intelligent global understanding to the text as a whole, which means taking time to read and comprehend even no-edit 100% matches. And time, of course, is money.
Putting a number on the gain
Translator workflows generally fall on a spectrum between two tasks: producing a draft and revising it into a deliverable. The drafting/revising balance might be 60%/40% for a quick-and-dirty first draft that needs heavy revising, or 80%/20% for a thoughtfully crafted first draft that needs a quick polish.
Joss Moorkens notes that, up until 2004, translationzone.com (run by the makers of Trados) featured a guideline whereby words in fully matched sentences commanded 30% of the normal rate (“Measuring Consistency in Translation Memories: A Mixed‐Methods Case Study,” Dublin City University, May 2012). Nancy Matis of Nancy Matis SPRL uses a 75%/25% guideline for revising a translation produced by a trusted translator. In my experience, a two-thirds drafting/one-third revising breakdown is a reliable guideline for the time I require to produce finished work, so I will use that here.
(Note that the quality issues of MT blur the drafting/revising boundary. One top search engine recently produced the following translation for me from Arabic: “O uterine poor and God will not lose anything to you said Amen like emoticon Admin: Hafid utsav.” This requires more than revising.)
So with a 100% match, the TM program essentially produces the draft (66% of the work). The work of the translator in revising that draft would therefore be the remaining 33%. So a rate of 33% of the no-match rate is, on its face, what the market has been paying translators for this work. The value of the efficiency gain would seem to be the word count multiplied by 66% of the conventional per-word rate. That’s one theory, anyway.
Divvying up the gain
So how would the value of the efficiency gain be divided? When TM was novel, clients sometimes offered nothing for a 100% match. In the same vein, some translators gave no discount for TM. Agencies sometimes did both: charging the client full price while asking translators to provide 100% matches for free.
That was then. The atomized nature of the translation market may impede the flow of information, but word does get out. Translators and agencies compete by offering a share of those efficiency gains to clients, and arbitrage drives prices to clients down. Over- or under-discounting freelance translators notice drops in their bottom line. And agencies who send texts to translators with 100% matches removed and non-matching segments arrayed with the underlying flow missing find translators demanding higher rates.
In a market without other disruptions, such as oversupply or undersupply of translators, these efficiency gains might be shared equally, each party getting a third. The translator should then be charging 33%—compensation for the translator’s revision task—plus its third of the 66% bonus (22%), or 55% of full rate for a 100% match. Or perhaps the client might share equally with an agency/translator combination that splits its share. A translator working with an agency should then be charging 33% plus 16.5% (half of 33%), or 50% of full rate.
That sounds good to translators. In two ProZ discussions “Zero rate for 100% matches” (May 3-14, 2010) and “Rates for Repetitions Match, Fuzzy Match and No Match” (Aug 3-14, 2005), translators shared the rates they have been offered or accepted: 0%, 5%, 10%, 15%, 20-30%, 25%, 30%, and 35%. One outlier cited 50%. No one mentioned 55% (although some never discount). Since the translator is revising and has made the investments of buying the software, training themselves to use it effectively, working with tech support, and creating TMs, why are these rates so low?
Part of the answer is that costs are not too relevant. Agencies also pay for software, hardware, tech support, training, and creation of TMs. And TM complexity means greater complexity of managerial tasks: estimating, bidding, delivering ancillary materials. And it’s no secret that the clients most likely to benefit from TM generally have internal translation capabilities. Such clients develop, supply and control their own TMs. They are savvy about needs and costs.
Another factor is that client TMs have been edited. Yes, changes may be needed, but the segments have at a minimum been vetted for accuracy, grammar and terminology. This reduces the translator’s revision contribution, shrinking the 33%.
The translation task may also differ in nature. A translator working on a complete, non-repetitive text without TM—I’ll call this bespoke translation—translates differently. Rotely translating the same sentence the same way twice (or more) degrades readability, encouraging readers to skip text, so a bespoke translator is aware of rhythm, pacing, and logic. TM pushes in the opposite direction: it loves one-to-one correspondence and handling of segments in isolation. But many texts—procedure-heavy manuals are an obvious example—benefit from describing the same task identically each time. Indeed, such language is not intended to be linearly read; it is to be consumed when needed.
Another issue is the transfer of tasks. TM may allow the step of applying global understanding to be performed by a client editor, in which case the translator or agency cannot expect to be compensated for it. Clients may be content to have translators do only multi-language revising, entrusting a monolingual editor with true editing tasks. The shrunken 33% is now subdivided again into bilingually sensitive revising and monolingual editing.
More of the task may also be transferred to the TM program. TM programs have increasing awareness of context. A group of 100% matches occurring in the same sequence as a previous source is likely to be usable as is. In this case, the value of the translator’s revising task drops further because the value of the drafting performed by the TM has risen beyond 66%. The percentage hits bottom when a client genuinely wants the translator or agency to do absolutely nothing with a 100% match. The compensation to the translator is now a nuisance charge to visually process and make translate/no-translate decisions for every such segment, tasks that make demands on working memory, a limited resource. (See Neil Betteridge’s 2009 translation of Torkel Klingberg’s The Overflowing Brain: Information Overload and the Limits of Working Memory (Oxford: Oxford University Press).)
The real world
So maybe instead of there being a formula for fair pricing, the opposite is true: pricing is telling us the value of what we do. A client with a professionally developed TM is able to retain proficient translators while offering a high discount because its TM is useful. Conversely, if translators refuse all jobs after the first, a client has a good indicator that the discount is too severe for the quality of the TM.
And yet. Good clients actually pay more, not less. As localization professional Anna Schuster reports, firms that provide the most useful, extensive TMs—developed over years by double-checking high-quality translations contributed by skilled translators—discount the least. They need top quality, so they value experienced professional translators. Such clients may produce style guides, references, glossaries. They know translators read closely and encourage translators to report issues. They believe bilingually sensitive revising has an edge over monolingual editing that can justify the cost.
At his Google Faculty Summit 2009 presentation “Statistical Machine Translation,” Franz Josef Och compared improvements in MT accuracy for seven languages as parallel corpus size grows (https://www.youtube.com/watch?v=y_PzPDRPwlA, 13:00 minute mark). Google’s Portuguese worked the best. Why? Although Google tried to get comparable corpora, the Portuguese translators who produced the first corpus may just have been better translators, and that advantage has been maintained as the corpora move from 1M to 10M to 100M and even 1B words.
So it may be once again simply a matter of getting what you pay for. A high discount tells the translator to not waste too much mental energy revising 100% matches; a low discount signals the translator to be thorough and conscientious with all text. Ultimately, an awareness and honest assessment of all the relevant factors can help all parties identify their needs, get to a fair price faster, and be more satisfied with what they get.
Mike Karpa is a veteran freelance JA>EN translator based in San Francisco CA. He was visiting professor of Japanese translation at the Monterey Institute of International Studies in 2002-2003 and completed an MA in International Policy Studies at Stanford University. His fiction has appeared in a variety of literary magazines and his recent articles on the neurophysiology of translation and bilingualism can be found in the ATA Chronicle, SWET Newsletter, and Traduire 224: revue française de la traduction.