Frames and Conceptual Dependency

Early artificial intelligence faced a problem that still exists:

How can a machine represent ordinary meaning?

A sentence is not merely a string.

It describes:

• objects,
• events,
• intentions,
• roles.

Two influential approaches were:

• frames,
• conceptual dependency.

Both tried to move from surface language to structured knowledge.

Frames Revisited

A frame represents a stereotyped object or situation.

It contains:

• slots,
• default values,
• relations.

The idea is associated especially with Marvin Minsky.

Restaurant Frame

A restaurant frame may contain:

customer:
server:
menu:
food:
bill:

The frame organizes expectations.

Defaults

A slot may contain a default.

Example:

Bird: can-fly = usually true.

Defaults support commonsense reasoning.

Exceptions

Penguin:

can-fly = false.

The specific frame overrides the inherited default.

This produces nonmonotonic behavior.

Why Frames Matter

Frames reduce the need to encode every fact explicitly.

Once a frame is active, many unstated assumptions become available.

This makes understanding efficient.

Frame Activation

The word:

restaurant

can activate knowledge about:

• ordering,
• tables,
• payment.

Comprehension becomes prediction.

Slots and Fillers

A slot is a role.

A filler is a value.

For:

John bought a car,

a commercial-event frame might include:

buyer = John.

goods = car.

Inheritance

Frames can form hierarchies.

A:

sports car

inherits properties from:

car.

This supports structured reuse.

Procedural Attachments

Minsky’s frame ideas allowed some slots to trigger procedures when information was needed.

Knowledge was not purely static.

A frame could contain instructions about how to obtain missing values.

Scripts Revisited

Roger Schank and Robert Abelson developed scripts for stereotyped event sequences.

The classic example was a restaurant script.

Scripts represent:

what usually happens next.

Script Structure

A restaurant script might include:

1. enter,
2. sit,
3. order,
4. eat,
5. pay.

This helps a system infer missing events in stories.

Story Understanding

Suppose a story says:

“John went to a restaurant. He ordered pasta. He left a tip.”

A script-based system can infer unstated events:

• he ate,
• he received a bill.

This is commonsense inference.

Conceptual Dependency

Roger Schank developed Conceptual Dependency, often abbreviated CD.

The goal was to represent sentence meaning independently of specific wording.

Different sentences with the same meaning should map to similar internal structures.

Surface Form vs Deep Meaning

Consider:

John gave Mary a book.

Mary received a book from John.

The surface syntax differs.

The underlying transfer event is similar.

CD tries to represent the common meaning.

Primitive Acts

Conceptual Dependency used a small set of primitive actions.

Examples traditionally included:

• ATRANS,
• PTRANS,
• MTRANS,
• INGEST,
• EXPEL.

These represented basic conceptual event types.

ATRANS

ATRANS represents transfer of an abstract relationship such as:

• possession,
• ownership.

Example:

John gave Mary a book.

Ownership transfers.

PTRANS

PTRANS represents physical movement from one location to another.

Example:

John traveled to Paris.

The actor changes location.

MTRANS

MTRANS represents transfer of mental information.

Example:

John told Mary the news.

Information moves from one mind to another.

INGEST

INGEST represents taking something into the body.

Examples:

• eat,
• drink.

Different verbs can map to the same conceptual primitive.

Why Primitives?

The goal was semantic normalization.

Many words could reduce to a smaller set of conceptual acts.

This might make reasoning easier.

Language Independence

If different languages describe the same event, a conceptual representation could in principle be shared.

Meaning would be separated from surface grammar.

This is an attractive goal.

Causal Structure

Conceptual Dependency also represented relations such as:

• cause,
• intention,
• state change.

The aim was not just labeling events.

It was modeling their internal structure.

Example

Sentence:

“John opened the door.”

A deeper representation may include:

• John performed an action,
• door changed state,
• closed → open.

This supports inference.

Inference from Meaning

If someone ingests food, we can infer:

the food enters their body.

If ownership transfers, we can update:

who possesses the object.

Structured semantics enables reasoning.

Conceptual Dependency and Scripts

CD represented event meaning.

Scripts represented stereotyped sequences.

Together they supported story understanding.

SAM

Schank’s Script Applier Mechanism, or SAM, used scripts to understand narratives.

It could fill in missing expected events.

This was an ambitious early natural-language system.

Memory Organization

Schank later developed ideas such as:

• memory organization packets,
• case-based reasoning.

Knowledge was organized around episodes and expectations.

Strength of Frames

Frames are good at representing:

• objects,
• situations,
• defaults.

They model structured context.

Strength of Conceptual Dependency

CD is good at representing:

• events,
• actions,
• semantic relations.

It aims for language-independent meaning.

Limitation: Hand Engineering

Both approaches required designers to specify:

• slots,
• primitives,
• rules.

This does not scale easily to the complexity of everyday knowledge.

Limitation: Choosing Primitives

Why should one set of conceptual primitives be fundamental?

Different domains may require different distinctions.

The choice can become arbitrary.

Limitation: Exceptions

Everyday situations contain endless exceptions.

A restaurant can be:

• self-service,
• prepaid,
• takeout.

Rigid scripts become brittle.

The Brittleness Problem

Symbolic systems often perform impressively inside expected structures.

Outside them, they can fail abruptly.

Commonsense is more open-ended than hand-coded frames suggest.

Frame Problem Again

Representing a situation does not solve how to update everything after action.

What changes?

What remains?

This remains a deep reasoning problem.

Statistical Learning Arrives

Modern machine learning often learns semantic patterns from data instead of encoding them manually.

Embeddings capture similarity without explicit frames.

The representation is less interpretable.

Neural Language Models

Modern language models can infer:

• event roles,
• likely sequences,
• semantic similarity

from statistical training.

They may implicitly learn frame-like structure.

Frame Semantics in NLP

Modern natural-language processing still uses ideas related to frame semantics.

Resources such as FrameNet represent events and their roles explicitly.

The old idea survives in new forms.

Semantic Role Labeling

A system may identify:

who did what to whom.

Roles include:

• agent,
• patient,
• instrument.

This resembles structured event representation.

Knowledge Graphs and Frames

Knowledge graphs encode explicit relations.

Frames encode contextual structure.

These approaches can be combined.

Modern AI increasingly mixes representation styles.

Neural-Symbolic Combination

A neural model may extract an event from text.

A symbolic layer may reason over it.

This combines learned perception with explicit structure.

Conceptual Dependency and Modern Events

Modern event representations no longer use Schank’s exact primitive system widely.

But the core problem remains:

map diverse language into structured event meaning.

This is still central to NLP.

Meaning Is More Than Words

Frames and CD both embody a crucial insight:

two sentences can differ linguistically yet mean nearly the same thing.

A machine must operate beyond surface form.

Context Matters

The same word can activate different frames.

“bank” in:

river bank

and:

financial bank

belongs to different structures.

Meaning depends on selected schema.

Defaults Are Powerful

Defaults let systems reason efficiently.

Most birds fly.

Most restaurants require payment.

But intelligent systems must know when defaults fail.

Nonmonotonicity Again

Learning:

Tweety is a penguin

forces revision of:

Tweety flies.

Commonsense reasoning is defeasible.

It can retract conclusions.

World Knowledge

Language understanding requires background knowledge.

The sentence:

“John dropped the glass. It shattered.”

is easy because we know:

• glass is fragile,
• falling can cause impact.

Syntax alone is not enough.

Narrative Understanding

Stories require:

• event order,
• causality,
• goals,
• unstated knowledge.

Frames and scripts were early attempts to represent this hidden structure.

From Explicit to Learned Structure

The history of AI moves from:

hand-coded frames

toward:

learned representations.

But the underlying requirement did not disappear.

Intelligence still needs structured expectations.

The Philosophical Lesson

Frames and Conceptual Dependency were attempts to make meaning computational.

Frames represented:

what normally belongs in a situation.

Conceptual Dependency represented:

what an event fundamentally means.

Their specific machinery is historically dated.

Their central problem remains alive.

The Next Question

Once knowledge is represented, a machine can try to derive new conclusions.

Can reasoning itself be automated?

Can a machine prove mathematical theorems by formal rules?

That leads to:

Automated Reasoning and Theorem Proving.