Machine translation – Future oriented or a thing of the past?
Machine translation (MT) refers to computerised systems responsible for the production of translations from one natural language into another, with or without human assistance. These systems have been developed since the 1960s following rule-based (RBMT) and statistical (SMT) approaches. In the early years after the inception of MT, the translation quality achieved by such systems advanced so rapidly that the notion of developing a fully automated high-quality machine translation (FAHQMT) system seemed tangible, which could, essentially, render human translators redundant.
However, to this day FAHQMT has not been achieved, and it is beginning to appear unlikely that it ever will become a reality. Although output quality has greatly increased in recent years, many inherent limitations of MT systems persist. Research efforts have shifted from trying to devise the ultimate system to understanding the potentials and constraints of the available MT architectures, and a great deal of work has been dedicated to evaluating the quality of MT output. In fact, researchers have suggested that the evaluation of MT is better understood than MT itself, referring to the wealth of evaluation techniques for assessing MT quality.
MT evaluation techniques range from qualitative analyses like the FEMTI framework to automatic quality-evaluation metrics like BLEU and METEOR. These metrics aim to determine the suitability of a MT-generated translation for a certain task. As such, an output may be deemed suitable for “gisting” but not for publishing as a final translation. However, although these approaches are useful for research purposes, they have not been implemented in a unified quality estimation framework to date, and a consensus in the translation industry about the merits of MT systems for aiding translators remains amiss.
Nonetheless, MT is nowadays integrated into most commercial and free computer-aided translation (CAT) tools to provide translators with machine generated translation suggestions of fresh content that has no translation memory (TM) matches. Although MT is still met with a hefty dose of scepticism among translators (perhaps rightfully so?), several studies have shown vastly increased user productivity achieved by such MT-CAT integration. For instance, Lange and Bennett (2000) found that translation productivity could be increased by 50 – 60% when integrating the MT engine Logos with the Star Transit TM system, and Guberof (2009) noted productivity gains of 25% when post editing MT rather than TM output. Furthermore, Federico et al. (2012) noted relative productivity gains of 55.5% for the language pair English – German and 74.2% for English – Italian when using an MT-integrated CAT over the sole use of CAT.
These studies are just a few examples from a large body of knowledge surrounding the modern implementation of MT into CAT, with other studies also focusing on the potential of MT when implementing controlled language or effects on post-editing.
So, although the notion of FAHQMT was never realised, MT should not be written off as useless too quickly. Rather, more research needs to be dedicated to leveraging its potential to best aid human translators, and perhaps future MT systems will be able to overcome some of the limitations that have been lingering for over half a century.
Sources and resources:
Federico, M., Cattelan, A., & Trombetti, M. (2012). Measuring user productivity in machine
translation enhanced computer assisted translation. In Proceedings of the Tenth Conference of
the Association for Machine Translation in the Americas (AMTA)(pp. 44-56).
Guerberof, A. (2009). Productivity and quality in MT post-editing. In MT Summit XII
Workshop: Beyond Translation Memories: New Tools for Translators MT
Lange, C. A., & Bennett, W. S. (2000). Combining machine translation with translation memory at Baan. Translating into Success: Cutting-edge Strategies for Going Multilingual in a Global Age, Amsterdam: John Benjamins, 203-218