This article was originally published in July 2019 and has been updated.
Artificial Intelligent Machine Translation (AI MT), like ULG's MT technology Octave MT, can unlock tremendous opportunity for global organizations—from virtually instant on-demand translation of internal reports or general web content to enabling real-time translation of chat and live interactions. It can also be integrated into virtually all translation tasks to create a level of efficiency when combined with human linguistic resources in tackling high-sensitivity customer engagement or sales activities.
Like all technologies, the key to unlocking this value is in understanding the opportunity and limitations of the technology based on the desired outcome you want to achieve.
What is the intended goal of the translated communication you need?
Knowing the intended purpose of your content is the first place to start when looking to increase efficiency and drive cost control with AI MT solutions. Materials all organizations need often have very different purposes and actual use or access patterns. Knowing where your content fits has a dramatic impact on the range and application of translation technologies such as AI MT.
About 85% of web content does not achieve heavy, active use in any language. Users, members, or customers will often opt for user-generated content, phone support, videos, or user guides rather than structured materials and formal guides.
If you have the data to know you are dealing with this type of usage model, AI MT offers an excellent opportunity to drive an efficient and cost-saving solution. In many situations, and with relatively little additional support, a trained AI MT can produce baseline translated content, or content where the translation is good enough to post and fulfill a translation requirement for this type of low-access content. In some cases, AI MT can reduce costs and turnaround time for this type of language need by around 80% or more.
If your desired outcome requires more engagement or confirmation that the material contains no errors (perhaps it is critical engagement content, or there are components linked to your overall brand visibility, brand value, or regulatory controls), then AI MT alone probably won't get you a working solution. However, it can still be a meaningful part of the production process.
Even with highly engaging critical path regulated content for J&J in the medical device space, we have been able to deploy MT within a human ISO compliant linguistic process to produce a 10 to 15% efficiency improvement with a guaranteed full human quality output.
Once you have a GOAL for the outcome of the content, then you can train the MT towards that goal.
Understanding the intended purpose of the content allows a professional language service company to set up the AI MT solution in the right way and at the right cost to match that goal; it is the critical first step to achieving increased efficiency and saving money. Once that goal and intention is set, the next step is to focus on the training.
Like all intelligent systems, AI MT needs to be trained to gain its maximum value. Much like we go to schools or universities to drive to our goals and maximize efficiency, the quality of the training you do is critical. If we have set a goal of baseline translation for low-access content, then the training can often be lighter, focused on key words that are critical to the outcome. If the desire is a high-quality human linguistic integration, then the training needs to cover the voice, tone, and style of your organization.
Training an AI MT solution at the right level helps manage costs and ensures output is aligned to your intended purpose.
Great testing is the key to great training!
Don’t forget to test...or, more accurately, don't forget to access the testing step of the vendor who is implementing the solution for you. Training is the key to making an AI MT perform in the desired way. If you can’t test a system accurately, it is impossible to know if the training you are doing is actually working.
Testing AI MT output is a science in and of itself; it's important to triangulate both human and automated testing to really know if you are building an efficient system. Automated scores like Bleu score or edit distance are very fast but also often don’t relate to the actual translation experience of working with the output. Similarly, the human tests often are biased against the technology. A great testing solution should triangulate multiple points of data to get an accurate view of how the quality of output is performing to the intended use and goal of the content.
Set a goal for the AI MT; then train for the goal and keep testing along the way to ensure you are driving to it.
In summary, these three steps will ensure you can drive translation efficiency and save money with AI MT solutions:
- Identify the goal of your content.
- Train the MT to align with that goal.
- Frequently test the MT to ensure that the AI system is achieving the intended results.
To learn more about AI MT, please feel free to contact one of our experts.