Over the past year, anyone following our series on language AI is likely familiar with the GTP-3 language model. As far as natural language generation (NLG) goes, the GTP-3 represents a new era of superior language AI. But now, it is not alone. When given abstract prompts, the GTP-3 showcased its ability to reflect human qualities, generate intelligent interaction, and even produce poetry. Simply put, this model embodies what we should both admire and fear when it comes to artificial intelligence.
Now, developers from Microsoft and Nvidia have upped the ante with the most sophisticated NLG model to date: the MT-NLG. From an outside perspective, this technology is not too dissimilar from the previous model. Yet, in the context of the language services industry, this model is far more impressive and on a completely new scale. For one, this AI is said to have unmatched linguistic accuracy and contain a parameter count of 355 billion more than the GTP-3. Similarly, MT-NLG can outperform any previous model in functions including reasoning, language inference, and reading comprehension. So, as it stands, we now have the most capable model to date, and one that in general has been cheaper and easier to train.
With advances in AI’s command of language, how are communications being impacted, and what should we make of it in language services? One important distinction to note is that for all its advances, NLG models like MT-NLG are actually a different class of language AI than, for instance, Google Translate, which relies on natural language processing (NLP) to determine and find the equivalent for what is being said to it, rather than to come up with something of its own design. Nevertheless, the challenges of relying on machines are the same in either context. With the previous GTP-3 model, a recurring issue lies in the datasets, as artificial intelligence systems of this caliber require vast amounts of datasets in order to establish parameters. During this training however, unintentional sources of bias or general toxicity from around the web can find their way into the system. In the case of the MT-NLG, this issue persists and is a further challenge to the industries reliant on language generation technology, even as it becomes more autonomous at making sense in language.
Regardless of such obstacles, this development in the AI sector should be seen as more hopeful than not. Throughout the translation industry, AI remains most effective working alongside human translators rather than as a replacement, and that model gains validation when even more astonishing technologies demonstrate the need. While the opposite reality may not be far off, a time defined by an increased reliance on artificial intelligence should also be associated with even more sophisticated NLG and NLP technology. At the end of the day, just as using technology in tandem with highly skilled linguists is needed to bridge the cultural barriers that AI cannot perceive or account for, human expertise is a persistent moderating factor in more experimental advances.
From providing technology-driven translation solutions to localizing innovative products, CSOFT’s global network of linguists and subject matter experts can help deliver some of the most cutting-edge solutions to market. To find out more about our translation and communication solutions, visit us at csoftintl.com!