In the framework of my Industrial PhD, and after some experience in two Language Service Providers (LSP), I wanted to understand why so many language professionals refuse MTPE projects, or when they accept them, why the LSP or end customer is ultimately dissatisfied with the quality of the product. To address these issues and clarify MTPE needs in the globalization market, I designed an online questionnaire using Jotform. ‘Machine Translation & Post-Editing in the Industry’ was aimed at European LSPs and received 66 valid responses. It was completed in February 2019.
I identified a set of key results from this survey: who decides, who is informed, how it is done, what is expected, and how it is paid.
Surprisingly, it is not the LSP but the end customer that usually triggers the decision to use MT! Why is this?? Is it only due to time frames or financial constraints? I cannot help but wonder how far an LSP would advise against the use of MT+PE when this is not suited to the content or text type. “The customer is always right”, OK? And if we don’t take this project, someone else will. So be it.
One delicate question to ask these 58 people who decide within their own LSP is whether they inform the customer or not that MT will be used. There seem to be very few studies of preferred practice on this matter. The responses are illustrated in the graph.
How is PE usually carried out? I have read many studies on post-editing guidelines, or on PE operations (addition, deletion, shift, or replacement), but there seems to be no clear, definitive or industry-wide method. Maybe it depends on the language pair, content domain, text type, MT system, or final usage of the translation.
Before entering into such detailed discussions, let’s first decide if the MT output will be post-edited without comparing it to the Source Text (ST) (i.e. monolingual PE) or if the linguist will have access to the ST. As my survey shows, monolingual MTPE appears to be very rare if not completely absent; all our respondents have access to the ST!
In regard to quality expectations, I use the traditional dichotomy of light and full PE. Mostly LSPs aim at publishable quality, which explains why so far there has been considerable dissatisfaction in the translation industry: MT engines have not been ready (and we shall see how NMT is changing things), the industry did not benefit from best practices until TAUS published its Post-editing Guidelines, and other publications appeared (such as the ISO 18587 standard), and professionals were never trained in these new skills.
I then wondered how many participants who picked either light or full PE actually carried out one or the other. This is because it is agreed that there is a range of different shades of quality, rather than a clear dichotomy established by an exact science. So I asked the respondents about their typical work scenario, and this time I proposed three possible answers. Among the respondents who originally voted for light post-editing, 3% could not decide when asked for a more detailed answer, so they now chose “I don’t know” (9%). On the other hand, the 73% who voted for full PE remained consistent, and said they work on medium quality MT output to deliver publishable quality. The most interesting group is the 9% who state that they work with poor quality raw MT and who still post-edit up to publishable quality! I will keep in touch with them…
Finally, let me briefly discuss one more delicate issue: payment! Many professionals and LSPs are still thinking about the fairest method to apply: either per source word or per hour. Here we can see which method has been applied until now by the 56 respondents who have a pool of freelance post-editors:
I consider Bammel’s method a good proposal (2019 Revising/Post-Editing Service). It combines a fixed rate per word and a variable component according to the editing distance per segment to build a grid for PE assignments. Challenge accepted!
Assessing whether a candidate is a good fit for an MTPE job, evaluating the quality of translations, and training students to post-edit machine translation output are all issues currently under discussion in the globalization market.
I believe that after the metamorphosis phase whereby translators become post-editors, we shall now move towards the hybridization of many language job profiles, which could in future be entitled “paralinguists”, “digital linguists”, or “language consultants”, for instance. Insofar these profiles are transdisciplinary, we can hope that the industry and academia will collaborate to ensure a better balanced distribution of training content in MTPE courses that are more closely adapted to the way in which this post-editing task is actually carried out.