A Grid-like Code for Language



Hi! This is my first time posting on the HTM Forum. I’m a big fan of Numenta’s work, so this is really exciting for me.

I recently launched a company called Emphasis AI. Our goal is the complete digitization of meaning in language. I’m posting this here because our approach has a lot in common with Numenta’s.

Our first system identifies sentence-level emphasis by analyzing language using a grid-like code. This grid-like code consists of a progression through a hexagonal lattice, which organizes information first in two dimensions, and then in three.

When representing language in two dimensions, our system makes use of a flat hexagonal lattice. When representing language in three dimensions, our system turns this hexagonal lattice into the equivalent of a face-centered cubic lattice.

Before we apply our grid-like code, we reduce language to its constituent signs. (Think of signs as simple units of meaning.) At first, the process by which we achieve this reduction is simple.

  • Nouns, pronouns, and verbs become entity signs.

  • Determiners, prepositions, and conjunctions become location signs.

  • Adjectives and adverbs become quality signs.

  • Interjections become interior signs.

Particularly within the entity category, there are several rules that complicate the process of sign identification. I’m happy to discuss these in detail below.

Once signs have been organized into categories, we apply our grid-like code. It allows us to identify sentence-level patterns of emphasis.

I was originally drawn to Numenta because of the correspondence between my location signs and Hawkin’s location signals. When I learned of Numenta’s work with grid cells, I realized the correspondence might be more than chance.

I’m posting this here because I’m eager to see if any further connections can be made between my work and Numenta’s. I’m happy to answer any questions the forum might have.