Machine translation (MT) is becoming ubiquitous, so much so that it is to be considered a key element of translation automation.
Enterprise clients with considerable financial resources for the development of a proprietary MT system have started to introduce post-editing in their workflow, while organizations that are still new to this technology could turn to language service providers to request MT post-editing services or to consultants to seek guidance in the implementation of machine translation and/or post-editing.
Within a project of post-editing of machine translation (PEMT), the requisites and expectations will vary depending on the different stakeholders. The enterprise customer that turns to the freelance translator/post-editor will do so for reasons that are different or in contrast with those for which he could turn to an LSP. Likewise, an LSP which turns to freelancers for post-editing does so with different requisites and with a different scope than the one with which the freelancer will see his/her assignment.
An enterprise client might already have implemented its own MT engine, with its own data etc., and therefore already have a clear idea of what to expect for post-editing. Another translation buyer might simply be anxious to implement MT (maybe also because of the NMT hype) to speed up its processes, increase its volumes and – let’s say it – to save money. In the first case, negotiations of post-editing project could be centered around setting up the right criteria for a Service Level Agreement (or at least follow the same path); in the second case, the translation buyer might need advice on the whole MT+PEMT project.
PEMT Projects: Three Essential Elements
There are three main aspects to consider when negotiating a PEMT project.
- The quality of the data on which the engine(s) have been trained. Since many errors in MT outputs can be found in terminology, sentence structure, and punctuation in the source text, data quality is paramount.
- Statistical or neural engine. Although nowadays MT is mostly NMT, just like with RbMT before the advent of SMT, an enterprise might have already invested in the past in a well-trained and tested statistical engine, with a more than acceptable raw output quality.
- General or vertical engine. An enterprise might find that the output of a general MT engine (like Google Translate or DeepL) is “good enough” for its goals for many reasons, while another would go for a vertical engine to run in-house to avoid IP/confidentiality issues. In terms of post-editing and final results, the more a machine translation engine is vertical and well-trained, the better. In the case of a machine translation system that is customized for language pair, domain and text typology, the output will be of reasonably high quality.
- Client’s expectations. Next, you’ll need to consider your client’s expectations, especially in terms of time-to-market and final quality. An MT-savvy client might already have a general idea of the quality of the raw output and plan according to a defined productivity model, while a client who’s new to MT might have unreasonable expectations due to their lack of experience.
- SaaS platform or API integration. Is the MT technology available in SaaS mode (and therefore the source text will be pre-translated and then sent to post-editing for further processing) or through an API connector to the engine within a CAT tool (and, therefore, post-editors will work efficiently with all functionalities of the tool at their disposal)?
- Target group and use of the post-edited text. The level of post-editing will have to be defined not only based on the quality of the raw output, but also on the target group and the use that the client intends to make of the post-edited text (for example, print in a brochure or on a website, and so on).
- Guidelines. As said before, when dealing with a customized engine, the MT output will be of high or good quality. In this case, the PEMT guidelines should be very specific and rigorously based on the types of errors produced by the engine. It will be necessary to indicate the level of PEMT necessary (light or full), and what the purpose of the text and the target group are.
- Terminology. If using an online general MT engine, it’s advisable to treat the machine translation output more like suggestions to be checked not only for accuracy, but also for terminology. In addition, after the post-editing task, wherever possible, subject-matter experts should check the translation output for terminology accuracy.
- Client’s economic objectives. An MT-savvy client knows that economic gains are not immediate. On the other hand, for a novice customer, cost-savings will play a major, if not primary, role.
When it comes to pricing, it’s important to create a compensation model either prior to or after post-editing.
- Before starting. If you want to set a pricing model before starting the post-editing project, you’ll need a clear-cut predictive scheme (for example the translation-memory fuzzy match scheme) and apply a word rate. The disadvantage is that this model may prove largely inadequate. Translation-memory fuzzy matches and MT segments differ significantly. Fuzzy matches over 85% are inherently correct and requires minor changes; on the other hand, machine-translated segments may contain errors and inaccuracies. In many cases, even a light post-editing may prove challenging. This model can be suitable for light post-editing of a very good output when time-to-market is the first requirement.
- After completion of the project. For this compensation model you need to perform an accurate measurement of the actual work performed, i.e. calculate the edit distance and then infer the percentage on an hourly rate. This pricing model is suitable for full post-editing.