Most technology reflects the social norms, biases, and creativity of its creators and subsequent developers. This remains so with artificial intelligence. AI knows what it knows based on what it learns.
Importantly, when scientists train an algorithm to perform a task, they feed the AI data that reflects what is going on in society at the moment and in the past. The algorithm processes, accumulates and learns from these data inputs.
Weaknesses and issues are often apparent in terms of language and AI has often erred towards chauvinism. For example, when translating from English to Spanish, even the most advanced machine translation approaches tended to follow “masculine” word agreement and translate the word “Senator” as “Senador” or “Doctor” as “Doctor.”
In addition, AI often unintentionally misgenders female “Senadoras” and “Doctoras.” This is because of how gender is reflected in model training datasets.
One company has a new approach to produce gender-accurate translations. The technology not only translates “Senadora” and “Doctora” based on the context, but it also ensures that the entire translated sentence is grammatically fluent as well as being factually accurate.
The generative AI technology comes from Smartling, Inc., an enterprise translation company. The objective is to bring human-quality machine translation closer to reality, according to Bryan Murphy, CEO of Smartling.
The advances in translation quality include implementing style guidelines, brand voice, locale-specific conventions, grammatically accurate terminology handling and the proper use of linguistic gender preferences in translations.
The main challenge was with designing a repeatable and predictable prompt engineering process in order to obtain the most suitable prompt templates for each specific use case. This has been achieved by enabling the use of large language models, such as GPT-4 and ChatGPT.
These are neural networks composed of many parameters (billions of weights), each trained on large quantities of unlabelled text using self-supervised learning or semi-supervised learning
These technologies can be deployed to ensure that machine translation follows the purchasing company’s style guidelines. In addition, nuances like maintaining brand voice and ensuring further perfects the grammar of the automated translation output are working.
Trials have demonstrated that the technology can improve a customer’s preferred terminology handling and eliminate gender bias in translations.
