For NLP semantics – some books from my grad school days that provide good case examples (when you want more than the Yale Shooting Proble):

Sowa’s Concpetual Structures (there is a short paper & a longer book)
Harman’s Change In View (talks about revising previously held false beliefs)

One concept of “Where to apply background processing work?” Which boxes are more in focus view? How well is is the current model integrated there & how well elaborated are the views of interest wrt anticipated needs? Add Sampler work where we currently have the greatest anticipated shortfall. Long term & short term storage concepts apply there too.

Each box should be relatively clear about the kind of individual “objects” or persistences that it picks out, notices as things have a beginning & an end, and could describe using NP. It’s a function of NL and/or persistences in sensor measurements & active investigations/engagements which what seem to be instances of sensor persistence…”Turned out it was a copy!”

Having picked out objects, then other bits of knowledge are often organized around those objects & their relationships.

Views can be thought of as including a way of describing the level of detail & precision in a situational or compound situation (e.g. a lifetime of scenes) that includes an a particular object. There is some background process for compressing and/or discarding data that is not in view.

Non-perceptual relations must be available to describe, in a consistent way (wrt to available data & noticing) how unconnected situations are related to one another in a larger box. Our psycho-scientific theory describes the nature/condition of transitions from one to the other (plane travel, heating, smashing, corporate legal registration or breakup,…)

Based on the discussion above, not available linguistic reflection, I think it is important for the architecture to be able to synthesize unique names for persistences in boxes/box locations & ways of transitioning between objects that are connected in the same scene & unconnected scenes. Those are primitive KR elements prior to words.

The focus here is basic research rather than applications. However, as noted the other day, the planned approach to CompLang is to use applications as a staircase of test exercises that provide, a way to simultaneously improve, deal with issues, & progress toward valueable goals.

Here I think:

Level 1: In a video game, generate & understand texts that describe the tactics used by a human or computer player to successfully achieve goals.

Level 2: Model & understand text descriptions that fit with YouTube howto videos & similar.

Level 3: Computer Generation of HowTo videos from text description.

Level 4: Generate text & video of mechanical plans to repair a given real world situation.

A lot of different directions are possible. The thought is to pick paths where success can be achieved initially & then inspection can focus on what is harder & new about the next level.