Build Intuit

A Resource Project for Natural Language Processing

“To understand is to experience harmony between what we aim at and what is given, between the intention and the performance – and the body is our anchorage in the world. ”  ―  Maurice Merleau-Ponty


Reference Frames & Artificial General Intelligence

Site Summary: Machines can’t acquire commonsense naturally as humans do. To assist the machine acquisition of commonsense concepts, the idea is to analyze the words of a text utilizing reference frames. We call it Word-Sensibility. The goal is a text analysis to assess word sense by simulating human responsiveness. That is, the ability to act quickly and positively to situations in the world. Reference frames move quickly between general and relevant viewpoints. In this way, machines can learn to make better sense of the things that people typically know. A conceptual model and propositions are proposed that could form the basis for further research.

The Quadranym Model of Word-Sensibility (Q): An Ecological Psychology Perspective on Word-Level Concepts and Situational ContextsNon-Mental-Representation Representation — Action Based Theory of Context.

Q System: Dynamic word sense relations represented in units, scripts & layers.

Visit the About page for the project’s, motivation, theory, scope and goals. 


Q System Outline

Context isn’t just the surrounding circumstances, because it includes and interacts with the subject that is surrounded, and the agent that tries to comprehend it all. ― Andrew Hinton 🔗

  1. Introduction
  2. Reference Frames
  3. Systems of Orientation
  4. Services to Assist Orientation
  5. Reference Frame Models
  6. The Theory Summary
  7. Q System Database
  8. Scripts and Layers
  9. Summing-up
1. Introduction

Word-Sensibility is About Agents Orienting to Communications:

In this article, we will introduce, reference frames, dynamical contexts and quadranyms. These features aim to represent the responsiveness of agents. 

  • Goal: Improve commonsense prediction with units of responsiveness.

Word-Sensibility is a theoretical approach that can be interpreted for a wide span of applications: from  API services for NLP to metaphor-analysis.

The theoretical ideas are most concerned with the non-linguistic aspects of using language, aspects that ground the responsiveness of communication.

(Note: Ontological concerns involve: Mereology and Part-Whole Reasoning.)

2. Reference Frames

Normative Responses for Artificial Intelligence:

Word-Sensibility (Q) is proposed as a model for text analysis. Data training involves a database with its features tagged to identify apt reference frames

Consider the sentence: Let’s move the couch over there.

General reference frames aptly apply, such as: agent, energy, time and space.

The words of the sentence are matched to reference frames:

  • Agent(let’s), Energy(move), Time(the_couch), Space(over_there).

A reference frame is represented as a predicate for each word (i.e., subject):

  • From Agent To subject: let’s + allow, we = goal of agent
  • From Energy To subject: movechange = matter of energy
  • From Time To subject: the_couchsequence = event of time
  • From Space To subject: over_thereposition = between of space

(Note: Reference frames are general semantic orientations to topic fields. They are like semantic perspectives and deal with what’s not visible in an expression.)

Latent Topic Variables:

Reference frames anchor on topic-states that target topic variables.

  • If agent: then self targets goal to find ‘agent’ topic variables.
  • If time: then present targets event to find ‘time’ topic variables.
  • If space: then void targets between to find ‘space’ topic variables.
  • If energy: then motion targets matter to find ‘energy’ topic variables

(Note: Reference frames are populated by quadranyms. See: nymology.org. A quadranym is the basic unit of a faceted classification scheme (Q). Any topic is analyzed into its component parts beginning with the prime quadranym classes.)

Source and Target Conditions: 

The reader intuits the general conditions when given the relevant context.  

  • Relevant Context: Move(the_couch, we) 

General Conditions: 

The source conditions (i.e., anchor states) are self, present, void and motion. The target conditions (i.e., target states) are goal, event, between and matter.

For example:

  • Basic Condition States: source ➝ target
  • Space Frame topic-states: voidbetween 

(note: Above, condition variables provide topic-states for the ‘space’ topic.)

In this spaceframe example, the anchor void targets between to find apt spatial measures such as, objects, regions, solid separation & spatial separation.

  • Void is a factor that is a constant in relation to any variable of between. Void is in essence the background in a figure-ground relationship with between

(Note: Source and target conditions borrow and abstract from James Gibson: Perception is the process of selecting, organizing, and interpreting information from our senses. Selection: Focusing attention on certain sights, sounds, tastes, touches, or smells in your environment. Something that seems especially noticeable and significant is considered salient. Organization: Taking the information selected organizing it into a coherent pattern in your mind. Structuring the information you have selected into a chronological sequence that matches how you experienced the order of events is known as punctuation. Interpretation: Assigning meaning to the information you selected by calling to mind relevant, familiar information to make sense of what you are hearing/seeing. Source and target conditions help model this perception process.)

Source and target conditions are virtual mental states. Situations occur in the world. A source relies on a target to find what’s relevant in a situation.  

  • A source condition is unmeasured. So, a target is required to find measure. 

In theory, a source is needed for frames to elicit potential target conditions. The idea is straightforward, there can be no foreground without background.

  • The invariable source is the standpoint taken. Topic variables are the target. 

So, for all the examples (motion targets matter for Energy | present targets event for Time | self targets goal for Agent) it is all  subjective objective.

  • Subjective: general source condition ➝ Objective: relevant target condition
  • Source conditions refer to the intersubjective construction of communication.
  • Target conditions refer to the objective relations that provide understanding. 

As one more example, say there is Energy. In the Energy frame, the source condition is motion. What makes it relevant to a situation is the matter of Energy. For instance, if something is falling on Pat, falling being motion, then relevant is what’s falling as matter  e.g., raindrops vs. a piano is relevant.

(Note: Source and target conditions are semantic orienters for relevant contexts. A quadranym has a subjective state or source condition and an objective state or target condition.  We introduce quadranyms in section 8. Q System Database.)   

Condition Classes & Domain Classes:

Source and target conditions are subsets of source and target domains.

  • Conditions ⊂ Domains 

Theoretically, conditions and domains develop on the same continuum and become distinct over time. Conditions become domains and prove the logic.

  • In the system, conditions are transient while domains are steady. 

(Note: Mental conditions are constructs held between mental domains. Domains such as social and spatial can share conditions e.g. In the club = self ➝ between. Progression: conditions form references which form realms which form domains. Realms are random reference classes and domains are ordered reference classes. • You might say that the realm is discriminant while the domain is determinant.)

3. Systems Of Orientation

Relevant Responses Require General Responses to Anchor On:

Consider our example: Let’s move the couch over there.

The relevant response is about why people move furniture around. The general response is about intuitive grounding and orientation. The aim is to create Systems of Orientation by layering frames ― scripts form on layers.

  • Systems of orientation combine, modify and layer reference frames together.

The system works to engage states e.g., agents in a temporal sequence, in a space, using energy. The sentential content is parsed to anchors and targets

  • The source conditions virtually anchor the target (topic) variables.

The words of a text are represented as subjects predicated on topic-states:

  1. From: Agent: Anchor Self(let’s) Target Goal(move).
  2. Orient: Time: Anchor Present(the_couch) Target Event(move)
  3. Orient: Space: Anchor Void(move) Target Between(over_there)
  4. To: Energy: Anchor Motion(move) Target Matter(the_couch).

Above represents a nested hierarchy with layered viewpoints. Notice how the base verb move is represented as both, a target and a source condition.

  • The base verb (move) is assigned a subject condition on each layer.

(note: See naive semantics for some background on the representation above.)

Overview:

  • Constituents: {<let’s, move, the_couch, over_there>}.
  • The Relevant Viewpoint: Move(the_couch, we)
  • The General Viewpoint: Agents <use> Energy
  • Hierarchical Range: From Agent To Energy
4. Services to Assist Orientation

Network Systems Situate the Words of a Text to Reference Frames:

Services to help link situational categories (networks) are available such as, WordNet semantic relations and ConceptNet commonsense knowledge graphs.

All frames have related semantic states e.g., void > vacant > unobstructed. This and knowledge graphs align states to a situation e.g., move <isA> goal

Relevant Predicate: the base verb of the sentence example is move.

  1. move = goal < Agent 
  2. move = event < Time 
  3. move = unobstructed < Space 
  4. move = motion < Energy. 

(note: Above, each reference frame gives ‘move’ a unique dynamical context. We introduce the dynamical context in the next section 5. Reference Frame Models.)

Grammar Links for Source and Target Conditions:

Word-sensibility uses syntactic analysis when parsing words to reference frames.  Syntax assists finding, grouping or conjoining words for placement.

Syntactic Analysis see: Link Grammar (API)

    +-----------------------Xp-----------------------+
    |                     +-------MVp-------+        |
    |       +------I------+----Os----+      |        |
    +---Wi--+--Op--+      |    +--Ds-+      +--J-+   |
    |       |      |      |    |     |      |    |   |
LEFT-WALL let.v s[!].n move.v the couch.n over there . 

Constituent tree:

(S (VP Let [Contraction let's]
       (NP s )[Pronoun Phrase us]
       (VP move
           (NP the couch)
           (PP over
               (NP there))))

D connects determiners to nouns.
J connects prepositions to their objects.
I connects certain verbs with infinitives
O connects transitive verbs to direct or indirect objects.

Relations between frames and syntax add a potential level of predictivity.

Each layer has an anchor, target and grammar links

  1. Agent: let’s ➝  move: From imperative verb To infinitive (base) verb (I). 
  2. Time: the_couch ➝  move: From direct object To transitive verb (O). 
  3. Space: move ➝  over_there: From verb To prepositional phrase (MV). 
  4. Energy: move ➝  the_couch:  From transitive verb To direct object (O).

(note: Agent = let’s : subject (we) of the sentence; first person plural imperative.)

Words of Text Become New Anchors for New Sets of Target Variables:

Neural network deep learning methods cluster words in reference frames.

The system identifies new anchors. For example: 

  • Anchor: Let’s ➝ target variables  (We(allow) = self < Agent(s))

In the example, the relation relevant to the situation is;  If let’s Then move. Let’s is the anchor and is a constant. Move is the target variable. It could be, If let’s Then try, sit or eat, or whatever becomes relevant in a given situation.

  • The aim is to amass clusters of target variables for each anchored frame.
  • The target variables available to a frame are based on the given situation.

Agents: We see the agent as a system in an ecological relationship with its environment. For a helpful view of agents in general see: Multi-Agent Systems.

(note: There is no actual subjective and objective distinction in the hierarchical structure. Mental attributes and environmental attributes are nested layers in a single global system. The distinction is in quadranyms and is temporal and local.)

5. Reference Frame Models 

Word-Sensibility Proposes Two Complimentary Systems of Context:

The Situational Context is predicated on Move(the_couch, we). We introduce the Dynamical Context*.   It’s about responsive alignment to the situation.

  • Words (text data) cluster in reference frames becoming situational variables.
  • Content collected in each reference frame has a specific Dynamical Context.

*(note: The dynamic context communicates the changes in a situation while the dynamical context communicates a system’s response to changes in a situation.)

The Dynamical Context Couples with The Situational Context.

Source Condition: motion ➝ Target Condition: matter

The images above are reference frame models for Energy. The images show different sets of mode-variables. Modes align target states to the situation:

  • left: general, active-passive, calibrate to situation [move ➝ matter ].
  • right: relevant, size-weight, calibrate to matter [move ➝ the_couch].

The more-general left frame cues a mean for how much weight people can move. The more-relevant right frame cues a mean for the weight of couches.

  • General: How much the average person can lift, slide or budge.
  • Relevant: Average weight of something e.g., couch, vase or mountain.

Situational Alignments: (Virtually aligning the agent to the world.)

Below, the content is parsed into the more-general reference frame.

:[Active(motion:(y=Let’s, move, over_there)) ⊇ Passive(matter:(x=the_couch))]

  • Modes and Causal Relations: dependent y and independent x 

Alignment depends on situational variables e.g., xcouch, xvase or xmountain.

  • Modes are used in regression analyses to compare and select target variables.

To anchor the energy topic to any situation, active is dependent on passive. A more relevant set of mode-variables (size & weight) target the X(?…couch).

:[Size(motion:(y=move )) ⊇ Weight(matter:(x=the_couch))]

In terms of energy, moving something of any size depends on its weight.

  • Space and time frames use scripts to orient to any target variable.

This is one example. The calibration can be utilized in different ways. Also, other reference frames could be nested in,  such as, distance or direction.

6. The Theory Summary

Word-Sensibility: 

The agent accumulates actual dynamic sense through experience and will act upon the dynamic sense necessary to orient to a given communication.

  • Situational sense is the potential sense of one’s actual dynamic sense.
  • Dynamic sense can be literally or figuratively matched to a situation.

Two distinct systems of context adds stability and versatility for analysis.

  • The situational context is about truth conditions (less fungible).
  • The dynamical context is about responsiveness (more fungible).

Dynamical contexts have no value unless coupled to situational contexts. Situational contexts need dynamical contexts for communication orientation.

  • There is a reciprocal kind of relationship between the two systems of context.

A reference frame is like a bias that allows assumptions about topics. This includes generalizations where conditions between domains are compared.

  • The orienting above can be used for other situations e.g., reaching or passing.
  • The system orients for prediction conditions and counterfactual conditions. 
  • Systems of orientation can be repurposed to orient metaphoric conditions.

Metaphoric Condition: “Moving that little couch was like moving a mountain.”

See the slideshow for more info: The Dynamical Context (bottom of page)

7. Q System Database

1) Quadranym Database:

Quadranyms populate reference frames.

  • The quadranym is the smallest unit of context in the Q system.

 2) Quadranyms:

Quadranym example: energy: (Semantic Set = Four Dimensions)

  • Active, Passive, Matter, Motion

3) Quadranym Representation: (Semantic Set Resource Variable)

  • (x)  energy(x)  [Active(motion Passive(matter)(x)] 

An Alternate Semantic Set Resource Variable: (it’s like a word sense)

  • Alt e.g., (x)  energy(x) [Available(power Deplete(fuel)(x)]

4) Phrase Template: (Orient Semantic Unit with Mode-Variables)

Anchor (general): For the motion of Energy;

Target (relevant):

  1. Active is Dependent on Passive to Find matter
  2. Work is Dependent on Power to Find matter
  3. Size is Dependent on Weight to Find matter

5) Generic Quadranym Dimensions or Primes (Faceted Classification)

A quadranym is the basic unit of a faceted clarification scheme. Any topic is analyzed into its component parts beginning with the prime quadranym unit

Prime Unit: Expansiveness, Reductiveness, Objectiveness, Subjectiveness

  • Mode Primes: Expansive adj, Expand v, (-,+) ⊇ Reductive adj, Reduce v, (+,-)
  • State Primes: Subjective adj, Subject n,  (-,-) ⊇ Objective adj, Object n,  (+,+)

Quadranyms can be structured in a free form design or what what we call a conjugal design,  where a state is always a noun and mode is always a verb.

  • Topic: Transfer:[E = move v, R = put v, O =  location n, = S Item n]
  • (x)  transfer(x) [Move(item Put(location)(x)] < energy realm

Basic Quadranym Phrase Template: 

The generic phrase template form: FOR the subjectiveness of any-topic; expansiveness is DEPENDENT on reductiveness to FIND objectiveness.

See: nymology.org  (Database • Acquisition • Wiki)

8. Scripts and Layers

Scripts are on the Horizontal and Layers are on the Vertical in the Matrix:

Basically, layers are more logarithmic and scripts are more linear. Or,  you might say that layers are more heuristic and scripts are more deliberative.

  1. Layers: agent:[Y(self) ⊇ X(goal)]<over> energy:[Y(motion) ⊇ X(matter)]
  2. Script: energy:[Y(motion) ⊇ X(matter)]<find>[Y(matter) ⊇ X(?… couch)]

Notice the quick links; self is layered over motion and goal over matter

(note: quick links like goal <isRelatedTo> matter is due to the quadranym matrix.)

The y variable is about potential active relations between the agent and the world. The x variable is about a thing being acted upon (x thingifies anything).

In scripts, targets become the anchors of the next frame.

  • Content iterates through a script until the cycle is complete. 
  • New reference frames are constructed through the process.

Scripts are meant to run on different layers and at different cycles. In theory,  the more general a script is the less cycles it will need to run.

9. Summing-Up

Reference frames nest into various hierarchical layers (i.e., from general to relevant). These nested layers form virtual systems of responsiveness. The approach aims to assist with hard Artificial General Intelligence (AGI) tasks.

Such as:

Ultimately, the goal is to enable machines the ability to share viewpoints.

  • Topic interoperability is the ability to adopt or reject the topic orientations of other AI systems. Competitive performance comparisons advance schemes.

    Q system reference frames can be used with deep learning models, that utilize word embedding methods,  to orient the word associations learned. 

    • Developing a Q system requires a Neuro-Symbolic (hybrid AI) approach.
    • Data components: database, knowledge graphs, ontology and open text.

    Visit the About page for the project’s motivation, scope and goals.  We are in the exploratory stage and look forward to any feedback. Please Contact.


    Introducing The Dynamical Context 
    Above From: About Page.

    Eros & Psyche
    • Movement Toward the Whole of Relatedness with All the Processes of Mind (sec. C Jung) •

    About

    Model

    Contact

            Drafts - Work In Progress - Do Not Quote Or Distribute

    © 2018 buildintuit.com All Rights Reserved