General Knowledge Dr. Claudia J. Stanny EXP 4507 Memory & Cognition Spring 2009
Overview How is knowledge represented in semantic memory? Models of the structure of semantic memory Feature comparison model Prototypes and family resemblances Exemplars Network models Schemas and scripts Influence on memory storage and retrieval Claudia J. Stanny2
Semantic & Episodic Memory Semantic Memory General knowledge Facts, ideas, concepts, categories Generic information Conceptually organized No temporal coding Subjective experience of retrieval: “KNOW” Episodic Memory Specific event knowledge Events, episodes May include specific information about self (Autobiographical Memory) Temporally organized Subjective experience of retrieval: “REMEMBER” 3Claudia J. Stanny
Types of concepts represented in semantic memory Logical categories and concepts Clear definitions Clear category membership Natural categories and concepts Things that occur naturally in the environment Tend to be thought about in terms of essential elements or features, but specific examples do not always have these features 4Claudia J. Stanny
How are concepts represented in semantic memory? Defining sets of features (Feature comparison model) Prototypes Family resemblances Exemplars Network models 5Claudia J. Stanny
How to study these models? Sentence verification tasks Present a sentence Measure RT to respond that the sentence is true or false Use patterns in the RT to make decisions about organization and retrieval of semantic information Example of trials in a sentence verification task Claudia J. Stanny6
Sentence Verification Task Findings Some sentences take longer to verify than others (semantic difference effect) Typicality Effect: RT is faster for sentences about typical examples of a concept or category A canary is a bird. A penguin is a bird. Claudia J. Stanny7
Feature Comparison Model Concepts are defined in terms of features Defining features (necessary / required features) Features must be present for meaning of a concept or category membership Characteristic features (features that are descriptive but not required) Anything with all of the necessary features is automatically included in the concept 8Claudia J. Stanny
How well does this model work? Logical categories Easily described with a list of defining features Membership is clear and unambiguous All members of the concept are equally good as examples of the concept Natural categories Not all members have all the “defining” features Features are correlated with one another Members vary in how well they fit the concept (typicality effects, graded category membership) 9Claudia J. Stanny
Prototypes Abstract, idealized representations of a concept The prototype stored need not correspond to any specific example Features of the prototype are highly typical of the concept What might the prototype be for dog ? What might the prototype be for animal ? 10Claudia J. Stanny
Prototypes Evidence for prototypes Typicality effects (graded structure of categories) Ease of access as an example of a category: Name a type of fruit Prototypes benefit more from semantic priming than non-prototypes Problem: prototypes do not address how we represent our knowledge of the variability of members of a category 11Claudia J. Stanny
Family Resemblance Category membership is not determined by a common set of defining features Games Instead, category members share an overlapping set of common traits that create a family resemblance for the category 12Claudia J. Stanny
13Claudia J. Stanny catch tennis bridge
Levels of Categorization (Rosch) Superordinate level categories General categories furniture, food, animals Basic level categories Specific enough to identify objects clearly chair, tomato, cat Subordinate level categories More specific, more detail than needed for some purposes Chippendale arm chair, beefsteak tomato, Siamese cat Claudia J. Stanny14
Evidence related to category levels Basic level categories are our “default” category levels We use basic level category names to identify and talk about objects We access basic category names faster than other levels of category names Memory for category information migrates toward basic level names (errors in recall will be basic level substitutions) Bigger priming effects for basic level names Experts develop more category levels in their domain of expertise Claudia J. Stanny15
Exemplars Concept is represented by the set of specific representations for members of the category we have previously encountered and classified Variability of category members is represented directly (in a set of examples) Typical members and prototypes are created from existing representations No economy in storage: all examples are stored 16Claudia J. Stanny
Network Models Characteristics are derived from the pattern of associations or linkages among concepts stored in semantic memory Collins & Loftus Model (knowledge) Anderson’s ACT-R Theory (knowledge and procedures) Claudia J. Stanny17
Collins & Loftus Knowledge is stored in a network of connected nodes and links Retrieval and sentence verification task entail activation of relevant information in the network Spreading activation moved from node to node though links, RT depends on number of links Claudia J. Stanny18
Based on Collins & Quillian (1969) Semantic Network Model
Testing Semantic Network Models Assume that activation of a node takes time Questions that require activating nodes at greater distances in the network will require more time than questions that activate nodes close together in the network Property QuestionsCategory Questions A canary can singA canary is a canary A canary can flyA canary is a bird A canary has skinA canary is an animal
Spreading Activation Model (Collins & Loftus)
ACT-R Model (Anderson) Adaptive Control of Thought Declarative memory Information represented in networks of interconnected nodes Procedural memory Knowledge represented as production rules Goal → Required Conditions → Actions Model for acquisition of skilled behavior motor programs as production rules Application to skilled cognition problem solving algorithms as production rules
24Claudia J. Stanny
Neural Network Models Parallel Distributed Processing (PDP) approach Connectionistic, neural network model Networks of neuron-like units or nodes Highly interconnected – multiple connections between units System learns by adjusting connection weights Claudia J. Stanny25
Knowledge represented as a pattern of connections Claudia J. Stanny26 Stimulus input Response output
Knowledge represented as a pattern of connections Claudia J. Stanny27 apple
Knowledge represented as a pattern of connections Claudia J. Stanny28 pear
Knowledge represented as a pattern of connections Claudia J. Stanny29 cat
Characteristics of distributed network models Network knowledge is built up by encoding specific experiences (exemplars) Spontaneous generation of categories emerges from patterns of connectivity & shared units Fill in missing information in new examples (default assignment) Protection from damage to part of the network Graceful degradation (partial retrievals) Claudia J. Stanny30
Schemas & Scripts Heuristics or organizational structures Categorical information (schemas) Event information (scripts) Facilitates comprehension Organizes information in memory Provides retrieval cues to facilitate recollection Potential explanation for errors in recollection Use of schematic information to “fill in blanks” during memory reconstruction 31Claudia J. Stanny
How schemas and scripts are used Direct attention to relevant details during encoding Fill in partial recollections with details from relevant schema or script Schemas represent the gist or general meaning of an experience or event Use schemas to make inferences about ambiguous information presented in a story Claudia J. Stanny32