Post-editing of statistical and neural machine translation
Let’s start with the good news. Machine translation systems based on a neural approach are becoming more reliable and accurate all the time, and sometimes they cannot be distinguished from the work of a human translator. However, language and translation are too complex to rely fully on artificially produced MT output. The problem tends not to be the obvious mistakes or stylistic errors that we are all too familiar with from statistical machine translation (SMT). Rather, it’s the mistakes that we don’t notice at first glance – because on a superficial reading they sound perfectly correct – that cause problems.
Examples without post-editing
Original: Select the name of the person you want to review the file you are sending from your computer, enter comments for the reviewer, and then click Send.
SMT (statistical machine translation): Wählen Sie den Namen der Person ein, für die Sie die Datei überprüfen möchten, die Sie von Ihrem Computer, geben Sie Kommentare für den Überprüfer senden, und klicken Sie dann auf „Senden.“
NMT (neural machine translation): Wählen Sie den Namen der Person, die Sie von Ihrem Computer senden möchten, geben Sie Kommentare für den Überprüfer ein und klicken Sie dann auf „Senden.“
While the SMT sentence is awkward enough to make you check the original to be sure of the meaning, the NMT version reads fluently and might be assumed to be correct – if you didn’t know that people can’t be sent by computer! At first glance you also might not even notice that something important is missing: “you want to review the file”. If you imagine a similar result for a specialised text, the content of which you are not an expert in, you get an idea of how easily translation errors in machine translation output can be overlooked.
So even with neural machine translation (NMT), rework by an experienced post-editor is not superfluous but an absolutely essential step.
Feasibility analysis, objectives and benefits of post-editing
At the beginning of each post-editing process, a feasibility analysis is performed. This determines the estimated degree of usability of the machine output. Based on this, a decision is made as to whether the project is suitable for post-editing or whether other steps (pre-editing and subsequent re-translation by machine, human translation) should be introduced.
Finally, the objective of all post-editing is to ensure that
- the target text is understandable to the reader,
- that the content of the translation corresponds with the content of the source text and
- that the customer’s requirements and specifications are met.
In order to achieve these objectives, we at oneword have certain quality requirements for the target text:
- Consistency of terminology and language register
- Use of domain-specific terminology
- Use of standard syntax, orthography, punctuation, diacritics, special characters, and abbreviations
- Compliance with guidelines
- Correct formatting
- Suitability for the target group
- Suitability for the intended purpose
- Compliance with customer requirements
The post-editor’s important task: keeping an eye on quality and efficiency
Post-editing is neither a traditional translation task nor a revision, which places special demands on post-editors.
Post-editing requires the post-editor to have a high level of understanding of the functions and objectives of machine translation, as well as linguistic and technical competence, and to be able to decide quickly whether the machine output is suitable for further processing or whether a segment would be better re-translated. This means that, efficiency is the main focus, as the entire process of MTPE in many use cases is mainly adopted to achieve higher degrees of efficiency.
Skills that our post-editors bring to your projects:
- Translation competence
- Linguistic/textual competence
- Research and information gathering competence
- Cultural competence
- Technical competence
- Specialist competence
- Detailed knowledge of MT systems and typical sources of errors in machine translation
- Competence in considering efficiency and cost-benefit analyses
The more experienced a post-editor is, the easier it is for them to recognise and eliminate the typical errors and peculiarities seen in machine translation. You can therefore rely on experienced teams that will find a happy medium for you and, in the interests of high productivity, decide what can>/em> remain as a blemish and what must be corrected as a fundamental error.