Post-editing of machine translation output (MT-PE) is the process of revising machine-translated content and editing it in such a way that the final product meets the requirements of the client. In most cases, the requirement is to achieve the same quality level as with fully human translation. The industry is working on solutions for analysing the quality of raw MT output and establishing the editing effort needed (preferably before the post-editing actually takes place), since these obviously affect the pricing of the service and the turnaround time that can be expected.
While looking for savings, it is important to realise that the raw output quality of MT engines varies wildly from Google Translate to highly-tuned and customised domain-specific engines. The latter can give very good results. Unless you have worked on the particular MT engine output before, it is difficult to estimate exactly what an equitable level of compensation for the post-editing should be.
The MT output is also usually combined with traditional TM output. Typically, any matches above 70–75% come from the translation memory and are dealt with by the linguist as in any normal job. Any segment below a 70–75% match gets machine-translated and the linguist post-edits it, recognising that there is a difference between editing a segment you know was produced by a human in the past (a TM match) and a segment that was put together by an engine. Typical errors in MT output include incorrect sentence structures, tenses, articles, inconsistent terminology and incorrect or missing tags. A good rule of thumb for the linguist is to look at an MT segment for two seconds and if they do not think they can easily edit it to produce a good result, to discard it and translate it from scratch, or use a lower fuzzy match from the TM instead.
The speed at which a linguist can carry out post-editing is directly linked to the quality of the raw MT output. With the projects STP has done so far, involving Scandinavian languages and customised MT engines, an experienced linguist would generally be expected to process 20–50% more work than when working from scratch. This productivity enhancement gives an indication of the savings in time and costs that can be expected.
Learn more about machine translation here.
