Types of knowledge Semantic memory: organization
Lecture 9 Representation of knowledge -
Types of knowledge Semantic memory: organization Models of semantic memory
Types of knowledge Types of knowledge
Episodic vs. Procedural -
Cohen and Corkin • Subjects: Normals vs. amnesics • Task: “tower of Hanoi task” 3 pegs and 4 disks (small, medium, large, extra large) • Results: Normals: Learns set of procedures Transfer Have episodic memory Amnesics: Can learn task (procedures) Transfer of training No episodic memory! • Suggests: Dissociation between episodic and procedural memory
Semantic memory: organization Semantic memory: organization -
Semantic knowledge is highly organized •
E.g.: psychoanalysis would not work if memory was not organized
•
We wish to examine and formalize
Tulving -
List of unrelated words
-
People do better and better after each trial
-
Interesting part people will subjectively organize
Models of semantic memory Models of semantic memory -
Feature model: •
-
Semantic feature-comparison model
Two network models: •
Hierarchical network model
•
Spreading activation model
I. Semantic Feature-Comparison (Rips, Shoben and Smith) Structure -
Knowledge consists of sets of features •
Bird = wings, feathers, beaks, flies
-
Features weighed on n-dimensional space •
Small ------------------------------------------------ large
•
Defining: important Ex: lots of emphasis. Bird has feathers
•
Characteristic: less important
Process/decision
-
How people would decide that something is something else
-
For question is a robin a bird?
-
Encoding •
-
Features for bird retrieved and for robin
Stage 1 •
First decision stage
•
Comparison of all features associated with robin and bird
•
Generate index of similarity/overlap (x)
X has a high value. So high that the index would be greater than some upper criterion (Cupper) If there is a really large amount of overlap, you say YES a robin is a bird! X has a low value. If it doesn’t even meet the lowest criterion, then just say no X is lower than upper criterion, but higher than lowest criterion. Ex: maybe an ostrich is a bird Can’t make a quick yes or no, so go to 2nd stage -
Stage 2 •
Compare defining features
•
A little bit of overlap, but only with defining features Plane has wings, but doesn’t lay eggs
Evidence -
Typicality ratings
• -
All birds have most defining features, so assume that typicality reflects characteristic features
Sentence verification (Rosch) •
Is a _______ a bird?
•
Model predicts: Typical items have defining and many characteristic features
Fast RT based on stage 1 Less typical have fewer characteristic features Slow RT based on stage 1 + stage 2
Problems for semantic feature comparison model -
Disconfirming sentences •
Responses should be faster to
RT (collies are birds)
STAGE 1
Would say no •
Than to Collies are not poodles
-
Results are opposite!!
II. Hierarchical Network Model
Collins and Quillian
RT (collies are poodles) STAGE 1 + 2
-
RT (robin is a bird) < RT (robin is an animal)
Structure -
Semantic memory vast collection of associated nodes (concepts)
-
Links:
-
•
ISA } subset
•
HAS, CAN } attribute relations
Hierarchy: •
-
Higher up is more inclusive
Cognitive economy: •
Common information stored at only one level
Processing -
Information retrieval = search of pathways •
“A canary can sing”: Find canary Retrieve properties
•
“A bird can breathe”: Bird node Move up level
-
Assume:
-
•
Takes a set amount of time to move across levels
•
Fanning (parallel)
•
Self-terminating search
Decisions: •
“yes” if retrieval successful
•
“no” if not
Evidence
-
Going from one level to the next takes more time
Problem: typicality effects -
Why is Robin rated as more typical than chicken?
-
Why faster to respond to typical than atypical?
III. Spreading Activation Model (Collins and Loftus) Structure
-
Not hierarchical
-
Related nodes are connected •
Boy-girl, branch-tree
-
Length of line = degree of relatedness
-
Links: ISA and ISNOTA (bi-directional)
Process -
Activate one node: •
-
Is a spread of activation to related nodes
Amount of spread: •
Strength of activation
•
Semantic distance
Decisions -
Evaluate evidence at intersections •
ISA or ISNOTA (gives super ordinate)
•
Strength of activation at intersections
Evidence -
Verification RT
• -
Robin is more closely connected, therefore the nodes reach each other faster
Semantic Priming •
Meyer and Schvaneveldt
•
Task: lexical decision:
-
Results:
•
If words are related (doctor-nurse), people are quick to respond
Summary -
Different types of representations in LTS •
Procedural, declarative (episodic and semantic)
-
LTS is highly organized
-
Several approaches to modeling semantic memory •
Feature lists
•
Hierarchical network
•
Spreading activation
Types of knowledge Semantic memory: organization Models of semantic memory
Types of knowledge Types of knowledge
Episodic vs. Procedural -
Cohen and Corkin • Subjects: Normals vs. amnesics • Task: “tower of Hanoi task” 3 pegs and 4 disks (small, medium, large, extra large) • Results: Normals: Learns set of procedures Transfer Have episodic memory Amnesics: Can learn task (procedures) Transfer of training No episodic memory! • Suggests: Dissociation between episodic and procedural memory
Semantic memory: organization Semantic memory: organization -
Semantic knowledge is highly organized •
E.g.: psychoanalysis would not work if memory was not organized
•
We wish to examine and formalize
Tulving -
List of unrelated words
-
People do better and better after each trial
-
Interesting part people will subjectively organize
Models of semantic memory Models of semantic memory -
Feature model: •
-
Semantic feature-comparison model
Two network models: •
Hierarchical network model
•
Spreading activation model
I. Semantic Feature-Comparison (Rips, Shoben and Smith) Structure -
Knowledge consists of sets of features •
Bird = wings, feathers, beaks, flies
-
Features weighed on n-dimensional space •
Small ------------------------------------------------ large
•
Defining: important Ex: lots of emphasis. Bird has feathers
•
Characteristic: less important
Process/decision
-
How people would decide that something is something else
-
For question is a robin a bird?
-
Encoding •
-
Features for bird retrieved and for robin
Stage 1 •
First decision stage
•
Comparison of all features associated with robin and bird
•
Generate index of similarity/overlap (x)
X has a high value. So high that the index would be greater than some upper criterion (Cupper) If there is a really large amount of overlap, you say YES a robin is a bird! X has a low value. If it doesn’t even meet the lowest criterion, then just say no X is lower than upper criterion, but higher than lowest criterion. Ex: maybe an ostrich is a bird Can’t make a quick yes or no, so go to 2nd stage -
Stage 2 •
Compare defining features
•
A little bit of overlap, but only with defining features Plane has wings, but doesn’t lay eggs
Evidence -
Typicality ratings
• -
All birds have most defining features, so assume that typicality reflects characteristic features
Sentence verification (Rosch) •
Is a _______ a bird?
•
Model predicts: Typical items have defining and many characteristic features
Fast RT based on stage 1 Less typical have fewer characteristic features Slow RT based on stage 1 + stage 2
Problems for semantic feature comparison model -
Disconfirming sentences •
Responses should be faster to
RT (collies are birds)
STAGE 1
Would say no •
Than to Collies are not poodles
-
Results are opposite!!
II. Hierarchical Network Model
Collins and Quillian
RT (collies are poodles) STAGE 1 + 2
-
RT (robin is a bird) < RT (robin is an animal)
Structure -
Semantic memory vast collection of associated nodes (concepts)
-
Links:
-
•
ISA } subset
•
HAS, CAN } attribute relations
Hierarchy: •
-
Higher up is more inclusive
Cognitive economy: •
Common information stored at only one level
Processing -
Information retrieval = search of pathways •
“A canary can sing”: Find canary Retrieve properties
•
“A bird can breathe”: Bird node Move up level
-
Assume:
-
•
Takes a set amount of time to move across levels
•
Fanning (parallel)
•
Self-terminating search
Decisions: •
“yes” if retrieval successful
•
“no” if not
Evidence
-
Going from one level to the next takes more time
Problem: typicality effects -
Why is Robin rated as more typical than chicken?
-
Why faster to respond to typical than atypical?
III. Spreading Activation Model (Collins and Loftus) Structure
-
Not hierarchical
-
Related nodes are connected •
Boy-girl, branch-tree
-
Length of line = degree of relatedness
-
Links: ISA and ISNOTA (bi-directional)
Process -
Activate one node: •
-
Is a spread of activation to related nodes
Amount of spread: •
Strength of activation
•
Semantic distance
Decisions -
Evaluate evidence at intersections •
ISA or ISNOTA (gives super ordinate)
•
Strength of activation at intersections
Evidence -
Verification RT
• -
Robin is more closely connected, therefore the nodes reach each other faster
Semantic Priming •
Meyer and Schvaneveldt
•
Task: lexical decision:
-
Results:
•
If words are related (doctor-nurse), people are quick to respond
Summary -
Different types of representations in LTS •
Procedural, declarative (episodic and semantic)
-
LTS is highly organized
-
Several approaches to modeling semantic memory •
Feature lists
•
Hierarchical network
•
Spreading activation
