Semantic Memory
- Goals of Semantic Memory Research
- General Background
- Demonstration
- A Computational Analysis of Sentence Verification
- Collins & Quillian (1969)
- Smith, Shoben & Rips (1974)
- More Recent Models
- Important Points Illustrated
- Semantic Memory and the Brain
Goals of Semantic Memory Research
- To Specify How World Knowledge is Represented in Memory
- To Specify How that Knowledge is Retrieved
General Background
- Semantic memory research has its origins in A. I. And machine translation.
- Psychological interest in semantic memory began in the late 1960s.
- The sentence verification task ("An S is a P.") was used almost exclusively to
study semantic memory until the mid 1980s.
- Error rates are low.
- Decision times are the usual response measure.
- Decision times are assumed to reflect the complexity of the underlying psychological
processes.
- Two Most Influential Models
- Collins, Quillian’s (1969) Hierarchical Network Model
- Smith, Shoben, & Rips (1974) Feature Comparison Model
Demonstration
- Read and understand each sentence.
- Say "true" or "false" out loud as quickly as you can.
Question:
- What makes some sentences more difficult than others?
A Computational Analysis of Sentence Verification
- What information is available?
- What is the goal of the computation?
- What strategy is used to achieve the goal with the available information?
- Configural Representations of Meaning/Storage
- Featural Representations of Meaning/Computation
Collins & Quillian (1969)
- Background
- Representational Assumptions
- Processing Assumptions
- Experimental Predictions
- Problems
Background
- Origin in A. I. (Quillian, 1968)
- This model was not originally intended as a model of human performance.
- Stressed Efficiency of Storage
- Collins & Quillian (1969) investigated the possibility that human memory might be
similarly organized.
- This model is a typical semantic network.
Representational Assumptions
- Memory is viewed as a network of concepts (not words).
- Concepts are connected by labeled relations that indicate:
- category membership ("isa")
- properties ("is", "has", "can")
- The meaning of a concept is its category membership and properties (configural
representation of meaning).
- Hierarchical Organization and Cognitive Economy
Processing Assumptions
- Step 1: Enter network at node corresponding to S.
- Step 2: Examine that node (or its properties).
- Step 3: If found stop, otherwise move up one step in the hierarchy.
- Step 4: Goto Step 2.
Experimental Predictions
- C0 < C1 < C2
- "A robin is a robin/bird/animal."
- P0 < P1 < P2
- "A shark can kill/swim/breath."
- Results: slope = 75 msec per comparison
Problems
- This model can’t explain why people make errors!
- Or how they determine that a sentence is false!
- Conrad (1972)
- Levels of Separation is Confounded with Associatve Strength
- When associate strength is controlled for, the difference between levels is eliminated.
- "A shark/fish/animal can move."
- Smith, Shoben & Rips (1973) demonstrated reversals in the levels finding.
- "A dog is a mammal." > "A dog is an animal."
Smith, Shoben & Rips (1974)
- Major Differences
- Representational Assumptions
- Processing Assumptions
- Experimental Predictions
- Advantages
- Problems
Major Differences
- No Hierarchical Organization
- Category membership is computed rather than stored.
- The primary determinant of RT is the relatedness of the subject (S) and predicate (P)
terms.
Representational Assumptions
- Concepts are stored as sets of features.
- Defining Features
- Must be present for an S to be a P.
- "Lays Eggs" is a defining feature of birds.
- Characteristic Features
- Features that are usually present, but don’t have to be.
- "Can Fly" is a characteristic feature of birds.
- The number of defining features decreases as we move up Collins & Quillian’s
hierarchy.
Processing Assumptions
- Stage 1: In parallel match of all the features in S and P.
- Yields a similarity value X.
- If X exceeds an upper criterion, respond YES.
- If X is less than a lower criterion, respond NO.
- Stage 2: Sequentially check all the defining features of P to determine is S has them.
- The upper and lower criteria are adjustable.
Experimental Predictions
- The Semantic Distance (Typicality) Effect
- Typical members of a category should be responded to more quickly because Stage 1 is
more likely to produce a fast YES response.
- "A robin is a bird." < "A chicken is a bird."
- People will make errors when Stage 2 is bypassed.
- "A whale is a fish." (high similarity but false)
- "A penguin is a bird." (low similarity but true)
- The Category Size Effect
- If an intermediate level of similarity ensures that you go to Stage 2 . . .
- then comparisons to superordinate categories should be faster . . .
- because fewer defining features have to be checked.
- "An ostrich is a bird." > "An ostrich is an animal."
Advantages
- This model does a better job of predicting verification times for true sentences!
- Semantic Distance Effect
- Category Size Effect
- It is able to predict verification times for false sentences!
- It is able to account for errors!
Problems
- What is a feature?
- Some concepts appear to have no defining features (e.g. "game" according to
Wittgenstein).
- The definition of a characteristic feature is circular.
- The presence of highly related negative items should attenuate the semantic distance
effect but it does not (McCloskey & Glucksberg, 1979)
- "A bat is a bird."
- "A dolphin is a fish."
More Recent Models
- Collins & Loftus’ (1975) Semantic Network Model
- McCloskey & Glucksberg’s (1979) Feature Comparison Model
- McClelland & Rumelhart’s (1985) PDP Model
Collins & Loftus Semantic Network Model (1975)
- This model is an intellectual descendent of Collins & Quillian’s model.
- Major Differences:
- Many Connections (No Cognitive Economy)
- Spreading Activation Replaces Search
McCloskey & Glucksberg’s (1979) Feature Comparison Model
- This model is an intellectual descendent of Smith, Shoben & Rips’ model.
- Major Differences:
- No Defining/Characteristic Distinction
- Single Comparison Stage
- Bayesian Decision Rule
- Have we lost track of the goal?
McClelland & Rumelhart’s (1985) PDP Model
- Self-Organizing Artificial Neural Network
- Shares Network Structure with Collins & Quillian
- Shares Features with Smith, Shoben & Rips
- But what are the features?
- These models do many interesting things:
- Classification of Novel Exemplars
- Prototypes
- Inferring Missing Features
- Semantic Distance Effect
Important Points Illustrated
- Storage versus Computation
- Tradeoff between Structure and Process
- Configural versus Featural Representations of Meaning
- Distributed Representations have Properties of Both
- Importance of Converging Operations
Semantic Memory and the Brain
- Warrington & Shallace (1984) showed a double dissociation between biological
categories (e.g., "animals", "foods") and human artifacts (e.g.,
"tools", "clothing").
- Does this suggest that semantic memory is hierarchically organized with animate versus
inanimate as a major distinction?
- Farah & McClelland (1991)
Farah & McClelland (1991)
- Model borrows heavily from Seidenberg & McClalland (1989) and McClelland &
Rumelhart (1985).
- Major Distinction:
- Visual Semantic Features
- Functional Semantic Features
- Reproduces Double Dissociation
The End!
|