Since ChatGPT was launched, we’ve seen an explosion of generative AI offerings making content easier than ever to create. That’s great news for organizations, enabling them to rapidly create new content at a lower cost.
You may be wondering if generative AI can be used to create multilingual content, eliminating the need for translation and potentially saving significant time and money.
The short answer is…no, not in its current standalone form (for example, ChatGPT). The issue is quality and cost. However, generative AI is already revolutionizing translation and enabling us to produce human-grade translations for a fraction of the cost and time it used to. If you’re in the business of global growth, this is something you need to know about. First, a little background.
HUMAN PROFESSIONAL TRANSLATORS ARE HIGHLY ACCURATE, BUT SLOW AND EXPENSIVE
Many people consider human translation to be the gold standard for professional translations because translation is not only substituting words in one language with words of another, but also getting the correct grammar, the right tone, the accurate brand terminology and the exact style of the original text. These, and other considerations, drive the belief that humans are the best at navigating those nuances, which are especially important in areas such as marketing.
According to the Multidimensional Quality Metrics (MQM) framework, the industry standard for assessing translation quality, humans produce the best quality translations. The MQM scorecard documents error types across four error severity levels. The maximum score value is 100, and human professional translators can achieve a score of over 97.
But humans are also the slowest and most expensive option. Top translators can translate around 2,000 words a day—impressive for a human, but not for a machine. You also need another human translator to review and edit translation drafts. In sum, it’s very expensive to use humans and difficult to scale.
MACHINE TRANSLATION IS INSTANTANEOUS BUT DELIVERS A SIGNIFICANTLY LOWER QUALITY SCORE THAN HUMANS
Neural machine translation (MT) works well for general and high-volume translations for which speed and costs are big considerations, such as customer reviews, training materials, subtitles, and time-sensitive translations. And its use is growing around 200% per year. In fact, machine translation now accounts for 55% of our total translations. While MT may not capture the tone and subtlety of humans, it conveys the meaning of the original text. MT is inexpensive, instantaneous, and has a place alongside humans in your mix of translation services. To put a number on it, untrained MT delivers an MQM score in the 78 to 80 range. Trained MT engines can reach an average MQM translation score of between 86 and 90.
GENERATIVE AI GENERALLY DELIVERS LOWER QUALITY THAN NEURAL MACHINE TRANSLATION
Our testing shows that both GPT 3.5 Turbo and GPT 4.0 deliver lower quality scores than current untrained and trained commercial MT engines like Google Translate or DeepL. They score lower than neural MT engines for certain language pairs due to inaccuracies and inconsistencies in results, also called “model hallucinations,” where they deliver wrong answers that sound highly confident. Moreover, large language models (LLMs) are also markedly slower, usually more expensive than MT, and their pricing models change frequently.
HOWEVER…GENERATIVE AI IS QUITE GOOD AT IMPROVING FLUENCY
Robots, don’t despair! Combining MT, generative AI, and putting a human in the loop can generate human-level quality at a fraction of the price.
This approach combines the instantaneous speed and low cost of MT for the initial translation; LLMs to correct the grammar, modify the style, and smooth the language; and highly trained linguists to monitor, review, and perform any minor edits. This approach helps organizations boost productivity and meet accuracy requirements, while ensuring that voice, word choice, and phrasing land well with the audience.
Hybrid translations with MT engines, LLMs, and human review can produce human-quality translations with an MQM score of 98 to 99 at half the cost and up to 200% faster than human translation. This allows you to localize more products and more content globally, improving conversion rates and customer satisfaction.
WHY LLMs ARE IN THE MIX BETWEEN MT AND HUMANS
Although it’s ideal to have a human editor for machine translations, especially of sensitive or important content, the addition of automation and generative AI is improving MT technology rapidly and in many ways.
LLMs enable translation engines to capture the nuances of human language that MT engines don’t have yet. For example, LLMs can manage gender nuances, such as translating a female doctor’s title from “doctor” in English to the feminine “doctora” in Spanish. MT engines incorporating a post-processing step by LLMs and designed for business use can also be trained to maintain brand compliance and voice by translating selected words and phrases in a source language to specific words and phrases in a target language every time. LLMs also can take MT drafts and adjust them for formality and word choice based on the target audience. A human can then optionally review a draft from MT and LLMs that’s close to final.
Generative AI is a game changer for companies that harness it for global growth. Will it replace your human translator? Not quite. But if you’re looking for a competitive advantage in your global growth strategy, implementing AI can help get you there.
Bryan Murphy is the CEO of Smartling, a LanguageAI company trusted by global brands to create multilingual experiences their customers love.