Language versus knowledge in AI
Refining what AI needs to do with language
My recent trip to Europe and Canada to refine my theory of knowledge is drawing to a close. We have visited a number of experts at universities in Europe and Canada to refine the challenge of knowledge representation for generalization, common sense and lossless accuracy.
The key to science, in my view, is to understand what the experts have discovered while maintaining a critical view and refining my own model. In the case of knowledge representation, I can see how easily a human brain deals with the vast number of semantic (meaning) distinctions we make with language, while limiting our understanding to what makes sense.
This is the distinction in AI known as the frame problem (described in Wiki and Stanford’s Plato very well). Paraphrased: how experts only seem to consider things that are relevant to the problem at hand. It’s common sense to only consider relevant possibilities that we also need to use in human language by excluding possible sentence meanings that aren’t relevant to the current context.
Upcoming Articles
In my recorded talks on the trip, they can be seen in YouTube. The recordings aren’t studio quality, but the questions are excellent given the audience’s expertise on the topics. Each talk was different and being refined to cover AI’s current limitations, the way forward with Patom brain theory, and how RRG has completed the heavy lifting to show the common features of the world’s diverse languages.
Soon, I will be writing about each of the talks in sequence from the NLU to NLE focus in the Netherlands (natural language understanding to natural language execution) where NLE takes the output of meaning to perform the tasks requested before returning responses to the NLG (language generator).
I often talk about Professor Thomas Dietterich from Oregon State University, partly because he is an expert in machine learning and partly because his solution is exactly the model that I advocate for - a split of language from knowledge and the integration of context to create a general knowledge base.
Of course Patom theory proposes a single representation in a network, with the use of intersection to divide knowledge into elements that align with the context set (aligns the current context with all available knowledge).
My flight to Vancouver takes off shortly, so I will hold of on more details of the upcoming articles.
Until then, think about what it takes to represent human language without loss to build up a repository of all the world’s knowledge in a way that also deals with the creation of a dictionary. Humans are able to both explain the meanings of words in their language as well as explain what has happened in situations, like an encyclopedia.
More on those topics shortly.