Neural Machine Translation

With the rise of neural machine translation, the question of whether or not automated translations will lead to the demise of human translators has become more prominent. Admittedly, neural machine translations have improved the quality and especially the fluidity of texts produced through machine translation programmes such as DeepL significantly compared to previously rule-based and statistical automated translations. The advantages for clients are obvious in respect of translation speed, particularly for high-volume projects, as well as cost savings. However, despite the materially improved machine translation output, it seems almost certain that human translators will still be required for the foreseeable future to improve texts translated by neural programmes so that target texts read like authentic texts written in the target language rather than translations. In terms of “feeling for the language”, human intelligence remains superior to machine translation.

Light Post-Editing vs. Full Post-Editing

Before taking a closer look at MTPE and its focus on the detection of errors in machine translations, it’s important to define the terms ‘light post-editing’ and ‘full post-editing’. Light post-editing is looking to eradicate obvious machine translation errors as well as those that may distort the meaning of the text. The objective here is not to produce a perfectly rendered target text but to make it a comprehensible, “good enough” text that conveys the messages of the source text correctly. By contrast, with full post-editing, the translator aims to render the machine translation output in a way that achieves a naturally flowing text, free of errors, which does not sound like a translation. Given its scope, full post-editing usually takes longer and is more expensive than light post-editing. In my experience, it also drives the vast majority of MTPE projects that human translators are asked to do. The diagram below depicts where in the MTPE process the human translator sits, namely in the editing step of the machine output and the subsequent quality assurance stage, both often being done by the same person.

❝Post-editors should ideally undergo specialised training since their tasks differ from those of a reviser or a proofreader. ❠

Like a reviser, the post-editor must look at both the source and the target text. However, the errors in machine translation output are usually quite different from errors committed by a human translator, for example choosing words that don’t convey the message of the source text correctly (e.g. ‘false friends’) or words that are not in line with the tone and register required for the respective text type. Such lexical errors are in scope of light editing processes. Syntax errors fall into the same category of obvious errors that need to be corrected in light editing. They often occur in machine translations where source text sentences are very long. Such syntax errors, e.g. an incorrect word order, often make the machine translation output sound non-sensical, hence correcting them is crucial to achieve a target text that does not distort the meaning of the source text or leaves the reader feeling confused.

Then there are less obvious errors, which must be picked up in full editing. Such errors can be manifold: spelling errors, orthographic errors such as upper and lower case, wrong grammatical forms, for example mixing up singular and plural or using the wrong case, untranslated strings of text, and punctuation errors. Where a CAT tool is being used, it’s particularly important to confirm that tags have been placed correctly in the target text. Tags are placeholders that determine the format in the native target file, i.e. the target text in the original file format such as Word or Power Point. Another focus in full post-editing are errors in style characteristics, for example tone and register. Here, human intelligence is indispensable since the machine may not necessarily emulate the style required by the target language. A good example are idiomatic expressions where the machine often comes up with literal translations rather than the figurative meaning. Style characteristics are essential elements in the fluency of a text that ultimately determine whether or not a translation sounds like an authentic text written in the target language. Full post-editing aims to achieve a quality matching that delivered by a human translator. These are some examples where human translators provide a clear benefit in terms of translation quality through MTPE.

I approach post-editing tasks in two steps, firstly deciding if a machine translation string makes sense or if a segment needs to be translated from scratch, and secondly, if the string does make sense, eradicating all obvious and less obvious errors. In my view, MTPE is an essential service in an increasingly automated translation world where human translators remain relevant as long as they accept MTPE as an opportunity rather than a threat.