Member-only story
Machine learning to support human language interactions has been unsuccessful when measured by its complex setup and lack of accuracy in use. While hopes were high when every major IT corporation setup a platform to exploit deep learning around 2014, the promise of conversation has been systematically replaced by technologies to cover-up the limitations in understanding, such as with click-bots.
A Gary Marcus article — click here is a good background reference. As is my scene setting for NLU in 2022 here. In that scene, publishers in the future move to a single meaning-based system, publish to that repository, and users select their choice of language to consume the content. Today’s article looks at the industry processes involved in the conversion from initial text to meaning and the reverse, from meaning to text in a selected language. We’ll discuss the use of a Super Knowledge Graph to store meaning without loss at another time.
In industry as in science, one failed experiment like deep learning chatbots opens the door to innovation. Since the dominant focus in Natural Language Understanding (NLU) moved to statistical models in the 1990s, advances have not led to systems that enable conversational AI. In fact, they cause problems with return on investment (ROI) to corporations spending millions on systems that may not even be as effective as their IT predecessors.
In 2022, note that the IT industry and computer science’s academia often calls those systems that use word vectors like in deep learning chatbot platforms — NLU . As they demonstrably do NOT understand us, they shouldn’t be called NLU!
Today’s review looks at what my company proposes as the future industry standards for reliable NLU. The idea is to follow the successful computer industry’s approach pioneered by IBM in the 1960s with their System/360 architecture.
Today I’ll explain what the new NLU industry will look like when complete. This makes the NLU industry “come of age”: when you can use a system without configuration for conversation, NLU will be the default way to work.
