Spreading-Activation Theory

UI/CM LAB

Spreading-Activation Theory

Quillian (1962, 1967) proposed a spreading-activation theory of hu-man semantic processing memory search as activation spreading from two or more con-

cept nodes in a semantic network until an intersection was found

Priming in semantic memory – spreading activation from the node of the primed concept

QUILLIAN’S THEORY OF SEMANTIC MEMORY concepts correspond to particular senses of words or phrases a concept as a node in a network, with properties of the concept rep-

resented as labeled relational links to other concept nodes

UI/CM LAB


Links are pointers, usually go in both directions between two concepts Links can have different criterialities, which are numbers indicating how

essential each link is to the meaning of the concept The full meaning of any concept is the whole network as entered from

the concept node The search in memory between concepts involves tracing out in parallel

along the links from the node of each concept specified by the input words

at each node reached, an activation tag is left that specifies the starting node and the immediate predecessor

UI/CM LAB


Common Misinterpretations of Quillian’ Theory1. The idea of cognitive economy (Anderson & Bower, 1973; Conrad,

1972) Strong theory of cognitive economy (Conrad, 1972)

All properties are stored only once in memory and must be re-trieved through a series of inferences for all words except those that they most directly define

Requires erasing info whenever it applies at a more general level Weal theory of cognitive economy

People surely store certain properties at more than one level in the hierarchy

UI/CM LAB


2. Quillian’s theory that all links are equal Links were assumed t have differential accessibilities (criteriality) Rips et al. (1973) – intermediate nodes are necessary for a network

model to explain the reaction time differences in categorizing different birds

Smith et al. (1974) – feature model can account for data that network models cannot (a concept consists of a set of values on a large number of semantic dimensions)

3. Categorization judgment (Juola and Atkinson, 1971) a categorization judgment proceeds from the instance to the category

but Quillian’s theory (1966, 1969) assumes the search proceeds from both the instance and category in parallel

UI/CM LAB


Anderson and Bower (1973) reject a Quillian-like model of a parallel search, while acknowledging that their data are compatible with a paral-lel model whose search rate is slower in proportion to the number of paths that must be searched

THE EXTENDED THEORYLocal Processing Assumptions1. When a concept is processed, activation spreads out along the paths of

the network in a decreasing gradient.2. The longer a concept is continuously processed, the longer activation is

released from the node of the concept at a fixed rate3. Activation decreases over time and/or intervening activity

UI/CM LAB


4. With the assumption that activation is a variable quantity, the notion of intersection requires a threshold for firing

Global Assumptions about Memory Structure and Processing5. The conceptual (semantic) network is organized along the lines of se-

mantic similarity (Fig 1)6. The means of concepts are stored in a lexical network (or dictionary)

that is organized along lines of phonemic (and to some degree ortho-graphic) similarity

7. Loftus’s data lead to the further assumption that a person can control whether he primes the lexical network, the semantic network, or both

UI/CM LAB


UI/CM LAB


Assumptions about Semantic Matching Process8. In order to decide whether or not a concept matches another concept,

enough evidence must be collected to exceed either a positive or a negative criterion

9. If the memory search finds that there is a superordinate (or a negative superordinate) connection from X to Y, that fact alone can push the de-cision over the positive (or negative) criterion

UI/CM LAB


10. If the memory search finds properties on which X and Y match (i.e., common properties), this is positive evidence proportional to the criteri-ality of the property for Y

11. The Wittgenstein strategy is a variant of the property comparison strat-egy

12. The mutually exclusive subordinates strategy was necessary for pro-gramming a computer to answer questions

13. Counterexamples also can be used as negative evidence

UI/CM LAB


RECENT EXPERIMENTSProduction Experiments of Loftus1. Freedom and Loftus (1971) produce an instance of a category that began with a given letter or was

characterized by a given adjective (name a fruit that begins with the let-ter A or a fruit that is red)

faster when the category was given first than when either the letter or the adjective was given first?

UI/CM LAB


2. Loftus (1973b) used four kinds of category-instance pairsa. Pairs where both the category and instance evoked the other with high

frequency (e.g., tree – oak)b. Pairs where the category evoked the instance with high frequency, but

the instance evoked the category with low frequency (seafood – shrimp)c. Pairs where the category evoked the instance with low frequency, but

the instance evoked the category with high frequency (insect – butterfly)d. Pairs where both the category and instance evoked the other with low

frequency (cloth – orlon) Category first a and b equal, faster than c and d Instance first a and c equal, faster than b and d

UI/CM LAB


3. Grober and Loftus (1974) noun-adjective (fruit – red), noun-letter (fruit –

A) a noun-adjective trial occurs in the semantic

network and requires the further step of retriev-ing the corresponding name in the lexical net-work

On the other hand, a noun-letter trials occur at the name in the lexical network. Therefore, the name does not need to be retrieved

4. Loftus and Cole (1974) – dictionary-network model noun-adjective-letter or noun-letter-adjective faster when the adjective is presented before the letter

UI/CM LAB


Juola and Atkinson’s Study with Multiple Categories category-search model (Landauer and Meyer, 1972) – subjects searches

through instances of the categories in memory seeking a match for the stimulus word

as the number of categories or words in the memory set increases, RT for the category-matching task should increase at a greater slope than RT for the word-matching task

UI/CM LAB


Effects of Typicality and Semantic Relatedness in Categoriza-tion Tasks semantic similarity speeds up positive decisions and slows down nega-

tive decisions1. the way evidence is aggregated distinguishing properties slows down the positive decision – matching

properties slows down a negative decision2. related to those cases where a superordinate connection is found accessibility will be highly correlated with typicality ratings Holyoak and Glass (1975) – two different cases where semantic relat-

edness or typicality does not produce the usual effect on RT for nega-tive judgments – mutually exclusive subordinates and counterexample

UI/CM LAB


Relation of the Theory to the Model of Smith, Shoben, Rips the meaning of a concept by semantic features of two kinds: defining fea-

tures (a member of concept) and characteristics features (not necessarily for concept membership)

in a categorization task, the two concepts are first compared in Stage 1 with respect to all their features, both characteristic and defining

second comparison in Stage 2 based on just the defining features there is no feature that is absolutely necessary for any category the model of Smith et al. has several inherent difficulties without the use

of superordinate links1. when people have superordinate info stored, they do not use it2. people seldom know the defining properties of concepts

Documents

Spreading-Activation Theory