A few months ago, we were mesmerised by DALL·E 2’s ability to generate realistic images and art based on short descriptions. Now the people behind DALL·E 2 have gone viral again with the launch of ChatGPT.
Both DALLE and ChatGPT highlight the advancements in generative AI. Compared to predictive AI, generative models have impressively fluent results, however, their quality is harder to assess. How do we know if a blog generated with ChatGPT is correct or if it’s accurate? Most of the output is fluent, and tangentially relevant, however these systems are inherently unreliable, making errors of both reasoning and fact.
Let’s take an example.
We asked ChatGPT how its development will impact the language and translation industry. Here’s what it said:
“It is likely that chat GPT (Generative Pretrained Transformer) will have a significant impact on the translation industry. Chat GPT is a type of natural language processing (NLP) technology that uses machine learning algorithms to generate human-like text based on input from a user. In the context of translation, this technology could potentially be used to create real-time, high-quality translations of spoken or written text.
One potential benefit of chat GPT for the translation industry is that it could help to reduce the time and cost of translation services. With the ability to generate accurate translations on-the-fly, chat GPT could potentially allow translators to work more efficiently and deliver translations faster, without sacrificing quality. This could make translation services more affordable and accessible, particularly for small businesses and individuals.
Another potential benefit of chat GPT for the translation industry is that it could improve the accuracy and nuance of translations. By leveraging large datasets and advanced machine learning algorithms, chat GPT systems can generate translations that are more accurate and nuanced than those produced by traditional translation methods.”
This could be particularly useful in industries where accuracy and precision are critical, such as in legal or medical translation.”
Devil is in the detail
It can’t be denied that the content reads well and sounds believable. But if you break down the response, there are some holes that need further investigation.
ChatGPT says it “will have a significant impact on the translation industry.” It sounds convincing, but as we will see later, this might not be true.
It also states that it “could improve the accuracy and nuance of translations.” This is very unlikely in the short term, given that accuracy requires specific domain training data, and terminology, which is not available on the world wide web data used to train ChatGPT.
The statement claims that ChatGPT could be “particularly useful in industries where accuracy and precision are critical, such as in legal or medical translation." In our experience, these highly regulated industries require a lot of adaptation to hit the accuracy levels required. And, realistically, this is likely to be the last type of content where a generic multi-purpose model – trained with web data – will be applied. Translators and localization professionals will continue to use state-of-the-art machine translation, like Language Weaver, where the output text needs to be fluent, accurate and utilise specific terminology and style-guides.
While sounding confident and fluent, the paragraph above suffers in terms of accuracy – and detail. At this stage, ChatGPT is a bit like having a junior analyst with surface knowledge about a particular domain, running high-level research, and then stitching together the top five Google results into something fluid.
Having gone through the above illustrative example, let’s think about the potential implications of using ChatGPT from the perspective of global organizations that need to manage an ever-increasing volume (and velocity) of content in today’s environment. In trying to meet consumer appetite for instant, relevant, and accurate information, organizations across private and public sectors today rely on Neural Machine Translation (NMT) solutions. Therefore, the crucial question we need to ask is – for global organizations, how does state-of-the-art NMT compare against ChatGPT translations in meeting their business requirements?
Well, state-of-the-art NMT models are better than general large language models (LLM) like ChatGPT due to three factors: quality, data privacy and deployment options.
1. NMT models have been trained as predictive models and optimized for accuracy. This can be further enhanced through adaptation, to generate additional an extra quality boost within a given domain. It is unclear how a shared generic AI model can be adapted yet.
2. Even if you could use ChatGPT as a translation tool, it is unlikely that you should, or be allowed to. Such technologies cannot be used by enterprises where data security and privacy are of paramount importance. This is a hotly debated topic, mainly around the data used to train large LLMs.
3. Models can’t be deployed securely and segregated from other users. There is just one single ChatGPT model owned by OpenAI, shared with all customers.
While machine translation models are inherently predictive, where both accuracy and precision are expected, generative models like ChatGPT can open up new avenues for translators and the localization industry.
The cost of creating content will drop, which means that even more content will be created. This creates a need for linguistic services for revising, adapting, and certifying the AI output. The concept of Machine Translation Post Editing will expand into linguistic validation, cultural adaptation, tone adjustment, fact checking and bias removal.
Another application is generating shorter or longer examples of the translation to help with formatting in-context. It is clear that NMT models will be complemented by LLMs for improving fluency or offering value added services like content insight.
While ChatGPT isn’t quite there yet – it’s certainly raised interest in conversational AI and the possibilities around the technology. But, is it going to transform the translation industry? No - it still has a long way to go. And until then it’s important to remember that only a robust, secure, and adaptable NMT solution can be trusted if you’re looking to engage with global audiences.