Caution: Will this tech change how we read and write?

There’s been a flurry of fundraising around companies seeking to help computers understand natural language. It could presage a change in the way we use and experience language and the written word. [I think it’s not the funding that presages but the tech. Maybe … “Development of such technology could presage a change in the way we use and experience language and the written word”?]

As we covered this summer, speech recognition company Speechmatics raised over $60M toward its ambition of global language comprehension. Text-to-speech company Verbit has raised over a quarter million to date. 

Into that fray steps AI21 Labs, which just raised $64M on a $664M valuation to continue expanding its suite of natural language processing (NLP) offerings.

All of this leads to a dawning reality where computers can understand natural speech and writing. That is an ability that so far has proven to be an elusive target due to the variability and extreme dynamism of natural language as used by real people. The immediate practical applications of that development seem almost prosaic: automated transcription or voice command functionality applied to a thousand goofy gadgets. But the real power of the technology could fundamentally impact our relationship with language and how it’s generated and consumed.

Consider Wordtune, a browser extension with millions of active users that was chosen by Google as one of its favorite extensions for 2021, and the related Wordtune Read. The latter analyzes and summarizes documents, the idea being that the summaries enable users to quickly and efficiently absorb the meaning and message of long and complex text. The benefit of the extension is obvious in a world where time perpetually is stretched thin. Effectively, this application condenses knowledge into something akin to its most efficient package. [I don’t understand this. What is knowledge’s most efficient package? Maybe “Effectively, this application condenses knowledge to a more efficient package.”?] 

Where’s the rub? Well, in truth we don’t know yet. But consider the consequences of stripping a nuanced speech down to its barest elements. In that context we’re sacrificing rhetorical power, certainly, but also, potentially, some degree of nuance and meaning that exists primarily in natural language’s ability to come at a subject from multiple angles, often by use of oblique inference and layered metaphor. One oft-hurled criticism of Stoic philosopher Seneca is that his writing consists of sound bites that repeat over and over. However, that approach is strategic; language does not create a perfect overlay of our conceptual world, so a language’s use must, at some level, exist imperfectly and inefficiently, a tool to be wielded by practitioners both skilled and unskilled.

The question, then, is not how well a computer can understand the individual words and sentences we speak and write (which is where most natural language processing goes awry) but rather how skilled a computer can hope to be at analyzing human concepts and abstractions and divining a speaker’s or a writer’s intentions in conveying those concepts. Given that humans still outperform even advanced deep learning models in problems requiring common sense and abstraction, it’s a safe bet we’re not quite there yet.

That’s no knock on the companies developing these technologies, which are providing very useful technologies with widespread applications. In fact, AI21’s tagline on its homepage is both appropriate and worth bearing in mind: “AI has a long way to go before it can match human intelligence. We aim to get it a little bit closer.”

So concomitantly, it also warrants that recognizing words and patterns is not the same as understanding natural language, and that should inform how we use these emerging technologies. Condensing a nuanced technical paper into a Cliff’s Notes version, even if done with remarkable fidelity, likely sacrifices some meaning, and there are circumstances where that’s unacceptable (issues of policy making or criminal prosecution, say).

For now, the market is certainly bullish on the future of natural language processing, and it is genuinely making impressive leaps in very short order.

“We completed this round during a period of market uncertainty, which highlights the confidence our investors have in AI21’s vision to change the way people consume and produce information,” said Ori Goshen, co-founder and co-CEO of AI21. “The funding will allow us to accelerate the company’s global growth, while continuing to develop advanced technology in the field of natural language processing.”

Most recently, the company’s AI21 Studio added a feature called Jurassic-X, an advanced natural language processing system that can handle tasks that lie beyond the reach of neural models alone.