Localization Management Series Part 3: Machine Translation
It must be tough tackling localization management these days. On the one hand, you’re facing pressure from senior management to both reduce localization expenditure but also ensure that the linguistic quality of translations does not suffer, whilst on the other, you’re receiving requests from customers and partners to provide *even* more localized content to them quicker than ever before. And on top of all that, you have your ongoing localization efforts to plan and coordinate. It must be a real headache sometimes.
You’ve been keeping half an eye on the developments with machine translation; certainly MT seems to be a far more mature product these days, and may well be the solution to the cost savings/time to market conundrum described above. Now, how to incorporate MT into my existing localization management workflow?
Please be realistic about MT
Machine translation has certainly made giant strides in recent years, but it is paramount that you do not see it as the one stop solution to all your localization needs. If the vast majority of your localization efforts are focused on producing marketing collateral, then I would not recommend investigating machine translation as a solution here. MT is certainly capable of producing good literal translations, but it is foolhardy to expect any MT system to possess the ability for content which requires a more creative linguistic flair.
If, however, you undertake a lot of technical documentation localization, whose content is far more descriptive than creative and where consistent application of terminology is key, then configuring machine translation into your localization workflow would certainly be a fruitful endeavour. But only if this integration is managed sensibly, for it is not the most straightforward of tasks.
You will of course need to build MT engines which have been customized using your companies language assets e.g. translation memories, glossaries, dictionaries etc. Again it is important to be realistic about what you can achieve internally within your company. Optimizing your training data, selecting appropriate test and tuning sets, fine tuning and re-training your engines, and running post-editing tests to determine productivity metrics requires a full engagement of time and technical resources which may be unfeasible to manage in house with all your “live” project work to contend with as well.
You may find that you have to outsource the MT system customization and integration to LSPs or MT technology providers in order to get the process moving quicker than you will be able to manage yourselves. Indeed outsourcing these tasks is advisable in most cases; an LSP or MT tech provider will more than likely have significantly wider experience with customizable MT systems and therefore the technical expertise required, whilst they may also be able to supplement your existing training corpus with additional relevant data which would give your customized MT systems a wider linguistic coverage. Even after a six to eight week training period, you should already be able to start making worthwhile time and cost savings, and with continual post-editing and retraining of your engines, these savings should become even more worthwhile over the long term. Which should make the members of your senior management happy, as well as your customers who are getting localized material quicker than ever before!