AI Failure: Machine Translation for Localization
via Shutterstock
Neural machine translation is a natural AI application that would allow users to localize systems by automating content translation into as many as 72 languages.
Despite the AI-powered technique’s promise for adapting international brands to local markets, a survey by an AI vendor specializing in machine translation finds that most of those AI projects fall short. A white paper released Thursday (June 20) notes that enterprise AI adoption has slowed and that as many as 85 percent of projects founder after failing to impress CIOs.
Pactera Technologies and language services consultant Nimdzi Insights found international tech companies are uniformly interested in machine translation. However, survey results released in a white paper exploring AI for localization found that only 23 percent actually used the technology while 77 percent of adopters have been stymied by lack of investment.
Poor data quality is a key reason AI projects fall short, largely because only a small fraction of available data has value for automation projects like machine translation. “Where humans are allowed to fail, AI isn’t,” the study concluded.
“AI systems [will] begin to see widespread adoption only after their quality exceeds near-human accuracy threshold,” the authors continued. “Machine translation compared to human translation is a classic example. The average edit rate for human translation is 11 percent, but machine translation is expected to have 100-percent accuracy.”
Neural machine translation, or NMT, is touted by vendors as an AI technology that aligns with business goals such as localizing systems and content. Localization specialists such as Beijing-based Pactera note the bias in most AI tools toward the English language. “In reality, customers and their data speak many languages, and there is tremendous opportunity in leveraging local content at an early stage in AI development,” the white paper noted.
Among the possible solutions are multilingual sentiment analysis via social media tracking of non-English customer reviews along with chatbots capable of interacting with customers in local languages.
The survey argues that neural machine translation for localization is a “classic AI application.” While the quality of neural machine translation has shown step-function improvement over statistical methods, adoption of the technology has been stymied by cost, operational complexity and lack of good training data.
“Not only is NMT data-hungry, the underlying infrastructures and knowledge required are also very taxing and challenging,” the white paper noted. Hence, the learning curve for production systems remains steep, hindering the reliability of emerging machine translation systems as well as related natural language processing frameworks.
The keys to advancing faltering AI projects to production are investing in data scientists and human interpreters who can, for example, educate machine translation engines through better collection, tagging, labeling and integration of training data.
“The future of AI in localization is outsourcing human tasks to the language services ecosystem,” the white paper concluded.
