28/02/2019 - Machine Translation
Machine translation (MT), in particular neural machine translation (NMT), will open the doors to good, rapid and cost-efficient language solutions in the future. But the road to this goal is bumpy since the time and financial costs should not be underestimated.
Which is why having experienced specialists with the necessary technical and linguistic know-how at your side is all the more valuable. They know what tools and measures are needed to achieve the goals that have been set.
What is neural machine translation exactly?
In the world of national and international business, multilingual communication is a must, especially in procurement. As a study from the USA shows, adapting to the local language can boost growth: eBay was able to boost its exports from the UK by 17.5% thanks to its own NMT solution.
For neural machine translation to succeed, rules and algorithms must be defined and transferred to the machine’s binary language. These create connections between bilingual texts at various levels.
Neural translation machines calculate vast volumes of data in milliseconds. This only works thanks to high computing power and deep learning: a method for optimising artificial neural networks. With deep learning, companies can slash translation costs, accelerate the process and reduce costs in the medium term. To follow this path, a strategic decision is needed on money, resources and how the company wishes to operate in the market in the future.
Artificial Intelligence (AI)
Using artificial intelligence, machines can perform work and tasks that call for human-like thinking. This technology enables computers to analyse large chunks of data and make forecasts or recommendations based on it, for instance.
Machine Learning (ML)
This area of artificial intelligence is based on self-learning algorithms. They identify patterns by using existing data and can take "decisions" independently.
Deep learning is also a sub-section of machine learning (ML). Computers learn without human help based on neural networks, similar to how the human brain works. Using huge amounts of data, machines teach themselves autonomously to learn and think.
Fundamental issues must be clarified
Do you want to invest in a professional machine translation system? What cost-benefit targets should be achieved in what time frame? Which target groups should be addressed and in which languages, what quality criteria must be met, and what guidelines such as terminology, corporate wording and reference texts are available? What are the technical requirements? Which security regulations must be considered? Are you thinking about an internal solution or connecting to an interface with an external provider?
The next step is training
The machine is fed parallel texts, i.e. source texts and the corresponding translations, as well as sector and company-specific terminology. The output is tested, analysed and corrected. Have the technical terms been translated correctly? What needs improving? The corrections are inputted and the machine is trained.
Then a new check is carried out, and another round of corrections, analysis, evaluations and inputting is performed. This training procedure is repeated until the desired level of precision is achieved and the quality of the machine translation is satisfactory.
Why should a company invest in its own solution?
Publicly accessible machine translators from the Cloud, such as Google Translate or DeepL, are becoming ever more accepted. They are free and can deliver a good result depending on the intended use. Drawing on one's own language skills, the output of a machine can be understood to a good level and is often satisfactory enough for simple, low-threshold communication. But what's the deal if content is confidential? What's the risk with sensitive information that should not be made public?
Data protect_on and confident_ality?
Cloud providers do not offer their services "for free" from purely altruistic motives. It's important to know that nothing is free (TINSTAAFL). The price for using a freely accessible translation engine from the Cloud is the data within the content. All content and rights pass into the ownership of the provider of the machine translation tool at a stroke. For example, Google's Terms of Service state that Google and those it works with may use, host, store, reproduce, change, communicate, publish, publicly perform, publicly display, distribute and create derivative works from (including those resulting from translations, adaptations or other changes) the data and content for its own purposes worldwide.
If a company deals with confidential data (e.g. personal data, information relevant to the stock exchange or patents), it will inevitably have to think about data handling for data protection reasons. In such instances, security and confidentiality are the most important drivers for investing in your own solution or for creating a secure connection with an external provider.
Do you have any questions about this topic? Get in touch with us.
Our publications on this topic [PDF-Downloads]
• „Universal Translator“ statt Profiübersetzer? (PROCURE SWISS MAGAZIN, Dez. 2018)
• Mit Trial und Error zu Effizienz und Exzellenz (SWISS EXPORT JOURNAL, 1. Quartal 2019)
Useful information on NMT
• Interesting news and articles about neural machine translation (NMT) on slator.com
• Research reports on machine learning
• Research project "Rich Contexts in Neural Machine Translation" (extended context in neural machine translation) at the University of Zurich
• OpenNMT – Open Source Initiative from MIT