Knowledge graphs, Linguists, and the Last-mile problem of AI

Now that AI can generate fluent text at scale in multiple languages and different styles, are authors, translators, poets, copywriters, science writers, editors, and other language professionals doomed to extinction? 

The consensus seems to be that dark times are coming and many professionals despair. But I think that many more opportunities are coming for professionals with significant training, skills, and experience understanding and dealing with the nuances of text and language. I've talked about this topic several times this year, notably as a keynote at #MemoQFest in beautiful Budapest last summer.  

You can listen to a video recording of The Next "New Normal" for Language Professionals here on YouTube.

TL; DL (too long, don't listen).  Here's a quick summary of why I see reasons to think that language professionals are not doomed.

1. Huge companies in high tech have enormous megaphones to spread their views and a deeply vested interest in promoting their solutions. But they have developed translation (and lots of other) technology with only their specific use cases in mind. This means they have a sort of tunnel vision.

Speed and coverage of translation in Google's use cases, for example, are much more important than reliability and precision. But there are many other use cases and markets where precision or creativity are at a premium, so in these cases machine translation is of severely limited usefulness. Think of the expertise and skill required to write or translate multi-billion-dollar contracts -- AI is very far from being able to do that reliably. Or an ill-fated use case like translating the Harry Potter novels: you could save $30,000 or more by translating one by machine but then you would also probably lose $30 million or more in sales!

2. Technologists with deep training in software engineering are leading language automation efforts. But computer scientists have no training in language and linguistics. They even call language "unstructured" data, as if there were no grammatical, semantic, or rhetorical relations between chunks of text.

One key role of text in the enterprise is to mitigate risks:  with it, we describe products in such a way as to mitigate the risk of sluggish sales, develop agreements to mitigate the risk of legal entanglements, communicate conquests and challenges to stakeholders to mitigate the risk that they might withdraw their investments. Mechanically generating more text in more languages without a clear understanding of these functions can only serve to increase risk and eventually undermine our clients' success. 

Now the last-mile problem for new tech is raising its ugly head:  making AI work is one thing, but the expensive and most time-consuming part will be to make it reliable. Engineers simply don't have the training or skills to notice functions like risk mitigation or even to see many of the issues that are blindingly obvious to language professionals. This means it's much harder for engineers to devise the appropriate functionality that will make AIs reliable and safe. As the technology progresses, the need for sophisticated linguistic skills will skyrocket. And engineers will have to collaborate more closely with other kinds of experts. 

3. Curated knowledge sources like ontologies and rich knowledge graphs are increasingly viewed as key contributors to AI reliability and safety. Developing these resources depends on deep and often subtle understanding of semantics.

Knowledge graphs contribute improvements to all of the steps in developing, evaluating, and deploying LLMs, with or without RAGs. This increasingly obvious fact simply emphasizes that sophisticated, reliable technology only comes about by deploying teams with diverse skills – including in this case, skills and strengths that only language professionals have the training to offer. Engineers can take us to very impressive prototypes on their own, but not all the way to transformative products in the hands of satisfied clients.

The development of ontologies and knowledge graphs is perhaps the most direct way to ensure safe and reliable AI systems, to promote effective collaboration between language professionals and engineers, and to demonstrate real impact on and ROI for both business and day-to-day activities. This is the motivation behind my on-going efforts to get people to consider more carefully the increasingly important role of structured knowledge in AI.

Read more in my Knowledge Architecture newsletter on LinkedIn.

Note: For those who prefer a more informal, conversational presentation, I also discussed these topics in an interview that was released on The Translator's Turn podcast with québécoise translator Sylvie Lemieux, Ph. D.. It appeared bothin English (with an intro in French) and, thanks to interpreterLucie Battaglia, C. Tr.,in Français Canadien as well.