Knowledge Representation Calculi – Current View

[Josh SternModerator]

Implicit in what I say above: Scientific knowledge/theory is very important to language understanding, but it fits in several different places:

I. What sorts of boxes exist & how are the linked
II. What sorts of different levels of physical abstraction/detail exist & which are compatible in theory
III. How dynamics actually work/play out inside of a box
IV. The distribution of what is found where in particular natural/real/true environments.
V. How sensing & measurement instruments interact, in detail, with environments & their contents to yield the data that underly our multi-sensory visions of meaning.

[Josh SternModerator]

Q: How should we think about the distributed representations of primitive situation VR parts in relation to OUR BEST INTROSPECTIONS ABOUT THOSE PRIMITIVE PARTS INCLUDING, FOR EXAMPLE, SIMPLE ENGLISH SENTENCES ?

A: At a given time, there is a computational process that can quickly come up with a representative example or can more slowly try to come up with a probability distribution over sample regions/types or a vaguess of boundaries around ‘most pleasant archetype’. Say the probability distribution is a component of a sort of generative/monte carlo sampling using imagination & theory, re-weighted by other relevant factors/theories that we notice might or should apply to a specific context. In both of these cases, the “answer” is not a precise data structure stored somewhere – it’s a result of mental processes that are hard to introspect about. But we can say is that a good/competent process for coming up with a ready example and for judging distributional properties/weights/chances is one that satisfies a set of performance conditions/evals & one that is amenable to teaching/correction/adaptation to changing environment or new theoretical understanding (the throw hammer goes in a straight line rather than a circle after we release it).

[Josh SternModerator]

Try to make tough language gyms from micro world semantics can be a good way to test compositional adequacy of KR proposals. Of course the tests that get used are biased by what researchers think/hope an adequate theory would look like. When predicate/argument forms are represented by distributed mappings, it would also be helpful to try and shape the grammatical requirements of the mapping shapes using a micro world. Nobody thinks they would have a different net for every task variant, but the shape of the general net becomes part of a KR theory, relating language to something else with sensory and memory inputs. Microworld shaping of reusable distributed reps would be interesting to see.

[Josh SternModerator]

By micro world I mean that we artificially create adqequate coverage of variety of sensory inputs, events, ways of talking, performance criteria, syntactic gymnastics, hypotheticals, etc. so that the formal adequacy of a way of processing can escape from failure due to incompleteness of inputs & focus on adequacy for other rigorous demands.

[Josh SternModerator]

One possible platform format for computing good exemplars & prob distributions over situation models that fit arguments + predicate in discourse context & events involving them:

Define a generative model/syntax for situations and a way of sampling with constraints. Constraints can be provided by environment boxes & discourse context & scientific theories & other mechanisms.

Probabilities of picking certain paths can be adaptively adjusted over time/experience.

Best exemplar – Do 1 or more runs & then perhaps move the result closer to most similar match in stored memory.

Probability – Do a bunch of monte carlo samples & derive a factor based likelihood model about the Q of interest based on the samples. This also can be adjusted/tuned by learning experience.

[Josh SternModerator]

Cognitively, an object or individual is a clump/grouping of stuff that is localized or localizable in space & tracks together over time. It is thought to have a beginning time & an ending time in various possible worlds – e.g. reality, fiction, hypothetical plan, someone else’s understanding of a name, etc.

A word/syntax predicate is often ambiguous & resolved to a specific meaning in context. Normally, the meaning of the resolved predicate stays the same within a given box while the set of things it includes will vary over time. Changing the meaning of a predicate within a box is kind of a special operation that could be said to produce a mildly different environment, only if the changes reflect changes in our scientific theory of what is possible. In that sense, scientifically grounded & social/verbal predicates are somewhat different. Radical changes to environment boxes can result in world situations where the presupposition/reqs of a predicate no longer apply in any sensible way. Again, it makes a difference whether or not that is due to a change in scientific paradigm in the other environment. How do I “look solid” at quantum scale at 20 degrees C but not at 200 degrees C? In the first case “I” am said to still exist while the 2nd case presumes that I perished. But our connections between looking/sensing/understanding at radically different scales, for example, are like links between different environments. The conditions for “looking solid to human vision” are no longer met at the quantum scale.

Intuitively it seems that there is a theoretical/virtual/algorithmic process of combining resolved predicates & objects/individuals in diffrent environments/scenes that works as a kind of cognitive test for correctness & viability. Consider an algorithmic version of “Can I make a compatible set of predicate/argument situations in that given environment? If not, what is the explanation for failure? The contents of these explanations are part of the hidden meaning of predicates & individuals.
‘

[Josh SternModerator]

Try to unify the remarks about objects, predicates, & environmental boxes in this way:

Any given environmental box + tech level admits a finite set of ways that our cognitive minds can use to sense/perceive/measure

Objects/Individuals in a given box are persistent localizations/clusters of perceived measurements/perceptions. They are framed in that way.

Predicates/relations are the rest of what we can notice/perceive about those objects/individuals & their relations & events those boxes.

Canonically, objects are visual. But our cognition can be very abstract. What is a set of algorithms? We can notice processes that proceed in steps, describe the steps as information processing changes in some state space, & describe algorithms as homomorphisms over those sequences of state changes, and we can talk about common properties of algorithms. Very abstract, but there is a reason that we are language requires grounding algorithm in some sort of representation we can point at.

[Josh SternModerator]

Meta-level common sense about language & knowledge is important to human intelligence.

Consider claims that are meant to be true or false? What supports the assertion? Generally, there are some sort of leaf/edge observational experiments, formal or informal, in situations that record facts as “true”, and then various linguistic & information processing activities are involved in summarizing/collating, recording, expressing, & forwarding those experimental results, faithfully or insincerely. Commonsense reasoners evaluate claims in a way that involves thinking about the provenance of these experiment & relay chains. How credible is the total package, as far as we can see?

Language interpretation often involves filling in the hidden picture of observational experience & relay – seeing what is implied by the most plausible model.

[Josh SternModerator]

For computing with that type of abstract structure – details to be possibly forthcoming in dribs & drabs on an as available & add needed basis, the relationa with expanding variables of successors to Prolog feels like it would be parsimonious & convenient.

X saw IT, X told Y, ….

[Author]

Posts