Human Post-Editing of Machine Translations
Machine translation technology has made very impressive advances in recent years. There is some dispute as to whether “human parity” has been reached. At Trusted Translations we believe that, for some situations, it has. However, in most cases, we recognize that it cannot yet replace the discerning judgment of a skilled human translator. A human touch will add more or less value, depending on a few factors like language combination, content domain, complexity and technical level, whether content is prosaic or creative, how well-written the source content is, input format, etc.
Due to these infinite variations, machine translations are not (and probably never will be) able to detect and implement the correct translation in every instance. Hence, to have a highly reliable translation output when utilizing machine translation, human post-editing will be indispensable in most instances. Also, to improve the performance of your Custom Neural Machine Translation Engine (CMTE), feedback from a human linguist is necessary for retraining.
At Trusted Translations, we have expert linguists trained specifically to post-edit machine-generated translations. Our expert post-editors utilize their experience with machine or automated translation output to not only correct the output, but also to improve future machine translation output.
With the right technology tools in place, the post-editing process will improve the quality of the output of a current translation project, allow you to store that quality data in your translation memory (TM), and will also improve the efficiency and accuracy of the machine translation engine for future projects.
Translating and Post-Editing Processes
The processes of translating and post-editing machine-generated content are quite different; each calls for distinct skills and training. Not all linguists are comfortable doing post-editing. And, even if they venture to sign on for post-editing, they may not know the ropes and will likely require some training. Academia quite recently started developing curricula to address the growing demand for human post-editors.
It all starts with the mental process used when reviewing bilingual content. When translating content from scratch, faced with the fact that the “target” (the would-be translation) is a blank canvas, translators normally follow this process:
- Carefully read the source content.
- Formulate the translation internally (i.e., in their brain).
- And finally, type their version in the target language (translation) in a software program or interface.
Depending on the content and project, the translator may edit the text immediately, conduct additional research, or come back to it later when the entire text has been translated.
The process when translating either a fuzzy or a full match coming from a translation memory (TM) may vary slightly, but normally three steps are involved. If they find content to leverage:
- They carefully read the source content.
- Then they carefully read the target content offered by the TM.
- They type their adapted version for the target language (translation) in a software program or interface.
In post-editing machine translated content, the process is different and involves the following steps:
- A sentence or segment of the MT output is read (that is, the process begins at the target).
- The MT output is compared to the source text.
- The Post-editors make some snap quality judgment of the MT output, based on explicit instructions.
- If the MT output is decent enough, they work on that to improve the text.
- If the MT output is of poor quality, they delete whatever the MT is offering and re-translate it (as in the translation process listed above).
Human Post-Editing and Human Translation
There is a significant difference in the skills and approach involved when reviewing human translations versus machine-generated translations. In general, the types of errors found in human translations differ greatly from those found in machine-translated text. For example, machine translations have a tendency to be more accurate when transposing numerical values. Machine engines are also less prone to leaving content behind (e.g., omitting a clause within a long legal paragraph). However, machine translations also tend to produce more awkward-sounding phrases. Some MT engines may also perform very poorly with terminology.
Human translations, on the other hand, tend to be better at creating a natural-sounding output, with fewer errors in context and meaning, while being more prone to errors when translating numerical values. A linguist is also more prone to making omissions. Some may even be prone to adding content when purposely trying to disambiguate. But that certainly strays from the source. At Trusted Translations, we train our machine translation post-editors to recognize these differences and to utilize our proprietary tools to improve the output. Furthermore, our tools help the editor re-translate awkward-sounding phrases in a manner that “trains” the machine translation engine to improve on future output.
Light and Full Post-Editing
At Trusted Translations we adhere to the principle that there is no one-size-fits-all solution, and that we must adapt every solution to our clients’ needs.
If quality is the driver, we evaluate not only the prospect of the engine’s performance, but also the level of post-editing needed according to the usability of the output. In this case, a full (ready-to-publish) post-editing is requested as the output for this task. That is to say, if the bilingual content were to go through an additional review step (editing, for instance), the editor should not detect any differences between an MT post-edited segment and a 100% HT (human-translated) segment.
If cost is the driver, or if time is of the essence—and the quality of the MT output is decent enough—we may recommend only light post-editing to make the text flow better, correct major grammatical errors with the help of a spell-checker, and even improve terminology usage.
Post-Editing and Pre-Editing
Post-editing should not be confused with pre-editing. Pre-editing refers to the preparation involved prior to running a particular text through an automatic translation tool. It involves reviewing the content for basic errors, tagging certain content to be translated a certain way (or to not be translated at all), and optimizing the formatting. This preparation prior to utilizing the translation tool can greatly improve output quality and is considered a vital step in most machine or automated translation projects.
Incorporating an efficient pre-editing process can vastly improve the output, making it easier for the post-editors to do their job and produce a quality translation. As machine translation technology continues to improve, it will be vital to have the necessary human skills in both the pre-editing and post-editing processes in order to produce high-quality translations.