Has machine translation reached parity with human translation?
There is no doubt that machine translation had improved dramatically. Neural modelling has facilitated impressive advances. But can machine translation now rival the work of the best human translators?
Several studies have been conducted in recent years to assess the relative quality of machine and human translation. Some of this research, notably Hassan et al.’s 2018 investigation, has suggested that machine translation can now be at least as good as human work. However, a set of recommendations published March 2020 in the Journal of Artificial Intelligence Research has cast doubt over such conclusions.
Recommendations for assessing translation quality
L ̈aubli, Castilho, Neubig, Sennrich, Shen, & Toralfor looked at how the quality of translations are assessed. They suspected that the efficacy of machine translation may have been overstated by researchers due to flaws in the methods utilised to assess it.
The research team found that the quality of machine translation had, indeed, been overstated by previous studies. Three principle issues were identified:
The choice of assessors
The relative efficacy of machine and human translations were often assessed by students or inexperienced translators. These assessors may not have been the best-qualified people to judge the quality of the outputs.
The importance of fluency
It is possible for a translation to be correct but to lack the fluency or nuances which characterise a superior translation. L ̈aubli, Castilho, Neubig, Sennrich, Shen, & Toralfor found that human translations boasted greater fluency than machine translation. They concluded that the importance of fluency has not been properly considered when judging the efficacy of machine outputs.
The human translations used in the studies
It is only possible to judge the quality of a machine translation when it is compared to the finest human translation. Recent studies have not presented assessors with superior human translations for consideration. Questions were also raised over the nature of the source material used for comparison.
In addition to identifying the above mentioned flaws in recent studies, L ̈aubli, Castilho, Neubig, Sennrich, Shen, & Toralfor have also suggested that there are problems with the way that machine translations are tackled by the software. Translations are typically performed sentence by sentence. This practice can mean that the outputs lack cohesion. They may feature contradictions and errors which would be readily identifiable, were an entire text considered during translation.
Conclusions
L ̈aubli, Castilho, Neubig, Sennrich, Shen, & Toralfor’s investigation suggests that the best human translations remain superior to machine translations. The team demonstrated that human work features fewer errors and offers greater fluency. The finest human translators are able to feature the nuances which are only possible if a text is tackled as a complete entity rather than sentence by sentence.
Professional human translators deliver superior results
Due to the evident flaws in machine translation, there remains no substitute for talented humans. Professional translators can assess an entire document to deliver fluent, cohesive results. Their work offers the consistency throughout a document or project that a machine simply cannot rival. Native speakers of any language possess the ability to comprehend nuances in meaning and context which are currently beyond the scope of artificial intelligence.
Machine translations must be reviewed and edited by professionals. If a project involves the translation of various texts, these must be compared to achieve uniformity of vocabulary, character and tone. It is, therefore, so costly, complex and time-consuming to effectively utilise machine translation, that it is far better to defer to the best human translators.
At Word Connection, our translators deliver consistently outstanding results. Their language skills, technical knowledge and creativity combine to ensure that no machine translation can rival their work. For this reason, our clients return to us again and again.
