Back to Cognitive Psych Home page Back to Topics

• OVERVIEW
• METHODOLOGY
• SIMPLE MODELS OF SEMANTIC ORGANIZATION
• NETWORK MODELS OF SEMANTIC ORGANIZATION
• FRAMES AND SCRIPTS AS SCHEMA

OVERVIEW

RETURN TO TOP

METHODOLOGY

1. The time required to retrieve one piece of information relative to another can provide important information about how this information is organized.
2. In the SENTENCE VERIFICATION TASK, Ss are required to consult semantic memory to determine if the relationship between two elements is true or false. For example, for each of the items below, answer as quickly as possible either true or false:
1. A poodle is a dog
2. A squirrel is an animal
3. A flower is a rock
4. A carrot is a vegetable
5. A mango is a fruit
6. A petunia is a tree
7. A robin is a bird
8. A rutabaga is a vegetable

RESULTS OF SVT STUDIES

• THE TYPICALITY EFFECT -- People reach a decision faster when an item is a typical member of a category, rather than an unusual member.
• THE CATEGORY SIZE EFFECT -- People reach decisions faster when an item is a member of a small category rather than a large category.
• CONTEXT EFFECT-- People reach decisions faster when an item was preceded by a similar item. Sometimes referred to semantic priming.
• THE TRUE-FALSE EFFECT-- People respond to true items faster than they respond to false items.

SIMPLE MODELS OF SEMANTIC ORGANIZATION

1. The Feature Comparison Model
2. The Exemplar Approach
3. The Prototype Model

THE FEATURE COMPARISON MODEL

• suggests that concepts are stored in memory according to a list of features or attributes.
• A two-stage decision proces is necessary to make judgments about these concepts.

(click here for graphic)(from Matlin, Cognition (1994))

THE EXEMPLAR APPROACH

• This model suggests that information is classified in terms of how well it relates to stored examples. Note that no abstraction is assumed -- simply, what example is it like.
• Not to be taken seriously. Lacks necessary cognitive economy and fails to recognize the fact that we do abstract.

THE PROTOTYPE MODEL

• DESCRIPTION
• Prototype theory suggests that knowledge is organized in natural abstract categories built around prototypes -- abstract, idealized representations or members.
• Unlike formal categories, natural categories may have
• Fuzzy borders
• --membership may overlap
• --partial membership is possible
• Centrality of category members
• --some members are more representative than others
• -- a graded structure.
• Categorization is made on the basis of the number of shared features with the prototype (which may not really exist).
• Ratings of prototypicality support this view.

(click for demo) (from Matlin, Cognition (1994))
• CHARACTERISTICS OF PROTOTYPES
1. Prototypes are supplied as examples of category. Evidenced in typicality effect.
2. Prototypes serve as reference points. Less prototypical members are identified as the more prototypical (11 is essentially 10).
3. Prototypes are judged more quickly after priming. Priming helps prototypes more than it helps nonprototypes.
4. Prototypes can substitute for a category name in a sentence.

(click for example) (from Matlin, Cognition (1994))
5. Prototypes share common attributes in a family resemblence.

NETWORK MODELS OF SEMANTIC ORGANIZATION

(click here for graphic) (copyright Fidura, 1995)

• Quillian's (1968) Teachable Language Comprehender (TLC)
• Collins & Loftus (1975) Spreading Activation Model
• Anderson's (1983) ACT* Model

QUILLIAN'S (1968) TEACHABLE LANGUAGE COMPREHENDER (TLC)

• OVERVIEW
• In TLC, concepts are stored at the nodes and the nodes are linked in a conceptual or classification hierarchy.
• Each node demonstrates two kinds of relationships, a superordinate link to a higher node called an "isa" relationship, and property characteristic called "has" relationships.

(click here for graphic)(copyright Fidura, 1995)
• Notice that the hierarchical arrangement creates the kind of cognitive economy that characterizes human memory.
• All subordinat nodes have the properties characterizing all superordinate nodes.

(click here for graphic) (from Best, Cognitive Psychology (1995)
• The properties of one node apply to all nodes below it.
• SEARCH STRATEGY
• Search involves scanning node by node and is constrained by the nature of the links and is also self-terminating.
• The search is terminated when an intersection of activation occurs and thus is sometimes called an intersection search.
• The process in response to a sentence verification task is essially as follows:
• The concepts in a SVT sentence activate their corresponding nodes in the network.
• Activation fans out in parallel from these two entry nodes in all directions along the links of the network. The TLC model assumes unlimited enrgy.
• As each node is scanned, flagging occurs, i,e, a pointer to the origin of activation is left behind pointing to the origin of activation.
• The search terminates when the two spreading patterns of activation intersect on a single node. When this happens, the flags indicate the pathway linking the nodes of origin.
• An inference process uses these pathways to determine the truth of the statement.

(click here for previous graphic)
• RESEARCH SUPPORTING TLC
• Collins & Quillian (1969), predicted reaction times in a sentence verification task using the model. They asked questions at three different superset (isa) levels and three different property (has) levels.
• Property levels
• P0: A canary is yellow
• P1: A canary can fly
• P3: A canary has skin
• Superset levels
• S0: A canary is a canary
• S1: A canary is a bird
• S2: A canary is an animal
• As predicted, the greater the semantic distance, the longer the reaction time.

(click here for graph) (from Best, Cognitive Psychology (1995)
• There is a significant amount of similar supporting research including much of the earlier research using the sentence verification technique.
• CONTRARY FINDINGS
• Rips, Shoben, & Smith (1973), using the same procedure but different materials reported contrary findings. In the superset sentences: "A dog is a mammal" "A dog is an animal," the model predicts that the reaction time for the first sentence should be shorter.
• Rips et al. found that the second sentence was actually verified faster.
• In the same study, sentences at the same level, which should have resulted in similar reaction times, such as: "A peach is a fruit" "A watermelon is a fruit." did not.
• Reaction times were shorter for the first sentence.
• While these findings were considered damaging, notice that the model could have been salvaged if, rather than being organized around formal concepts, it was seen as being organized around prototypes.
• The findings of Rips et al. may simply reflect that the conceptual hierarchy is neither arbitrary nor necessarily logical, i.e., it doesn't follow the rules of formal classification.
• Beyond this, however, there were other problems:
• The model predicts that verifying remote concepts ("A flower is a bus") should require a long time. In fact, it does not.
• In addition, the model does not account for findings in studies of semantic priming.

COLLINS & LOFTUS (1975) SPREADING ACTIVATION MODEL

• OVERVIEW
• Because of the shortcomings of the TLC model, Collins and Loftus (1975) suggested that the network is not, in fact, organized as a formal conceptual hierarchy.
• Based on the literature on semantic priming, they suggested that the links in the network were based on semantic distance or relatedness (semantic association).
• SEMANTIC PRIMING
• Semantic priming is identified in a classic experiment by Meyer & Schvaneveldt (1971).
• In this experiment, Ss were presented with pairs of elements made up of letters. Their task was to judge as quickly as possible, whether BOTH elements were words.
• There were five kinds of pairs: 1. Both elements were words but unrelated 2. Both elements were words but related 3. First element a word, second a nonword 4. First element a nonword, second a word 5. Both elements nonwords
• The trials on which both elements were words were referred to as positive trials. Notice there were two kinds of these.
• The trials on which at least one of the pairs was a nonword were referred to as negative trials and there were three kinds of these.
• The RTs on the negative trials are as predicted and suggest the decision can be made on the basis of the first element.
• The important results are seen in the positive trials on which both elements were words. Notice that the RTs are significantly shorter on those trials in which the words were semantically related.
• These results, which are consistently confirmed, suggest that activation of a conceptual node facilitates retrieval of semantically-associated concepts or words, as if the activation of a node spreads to those semantically nearby.

(click here for results)(from Best, Cognitive Psychology (1995)
• This could be the case only if the nodes of networks are linked by semantic associations.
• THE SPREADING ACTIVATION MODEL
• Like TLC, this model assumes that concepts are stored at the nodes. Unlike TLC, the links are made up of semantic associations rather than a conceptual hierarchy.
• The lines in the graph indicate semantic associations. The length of the lines indicate the strength of the association, specifically: the shorter the line, the stronger the association.

(click here for graphic) (click here for graphic)(from Matlin, Cognition (1994))
• The superordinate "isa" relationship is retained. In addition, this model assumes that certain relations are stored directly.
• These include an "isnota" relation. This direct storage of this relation (Example: "A school is a bus") prevents an extensive search for highly remote conceptual relations.
• THE SEARCH PROCESS
• The basic search precedure is the same as in TLC, namely, intersection searches. One difference, however, is that as the nodes are searched, they change status -- they become activated.
• Activation is understood as knowledge being brought into an increased state of accessibility.
• This mechanism would predict semantic priming.
• Notice that in the sentence "A McKintosh is a fruit," the node 'apple' is activated because of its associative proximity (strong assocition represented by a short link) with the other nodes.

ANDERSON'S ACT* THEORY

• OVERVIEW
• ACT* (Adaptive Control of Thought) is the latest incarnation of a series of increasingly complex models of John Anderson to account for all of cognition.
• It bears some similarities to both TLC and the Spreading Activation models. It is different in the kind of information assumed to be stored in the nodes.
• Rather than concepts, ACT* assumes that propositions are stored at the nodes. Declarative knowledge is organized around an interconnected set of propositions.
• A proposition is the smallest unit of knowledge that has truth value.
• Propositions are inherently relational so the links of this propositional network consist of these relationships.
• ELEMENTS OF THE MODEL
• While the propositions of Anderson's model are really abstract cognitive elements, propositional networks are most easily understood in their concrete expression in sentences.
• The meaning of any sentence can be represented in a propositional network: as a pattern of interconnected propositions.
• Consider the sentence "Susan gave a white cat to Maria, who is president of the club." It consists of three propositions:
1. Susan gave a cat to Maria.
2. The cat was white.
3. Maria is the president of the club.
• The network can be represented graphically by establishing a node for each proposition and letting the relationships among them constitute the links.

(click here for graphic) (from Matlin, Cognition (1994))
• Note that the exact form of the sentence would make no difference in how it is stored (e.g., if the sentence was in the passive voice). This is confirmed in the literature on memory for prose.
• Also note that each of the concepts in the sentence can be represented by a network as well.

(click here for graphic)(from Matlin, Cognition (1994))
• ASSUMPTIONS OF ACT* THEORY
1. STRENGTH ASSUMPTION: Each link has a specified strength, and the strength of a newly formed link is low but is incremented each time the link is used.
2. ACTIVATION ASSUMPTION: At any instant, a small portion of the nodes in logn-term memory are in an active state; all other nodes are not.
3. SPREAD OF ACTIVATION: Activation spreads out from an active node to the passive nodes to which it is linked.
• The stronger the link between the two nodes, the more likely it is that the activation will spread along that link
• The spread of activation has limited capacity in that the more links that are being activated at once, the less activation that will spread to any one.
4. DAMPENING ASSUMPTION: Periodically, activation is dampened throughout the network (i.e., all nodes and links are deactivated).
5. ACTIVE LIST ASSUMPTION: A maximum of ten nodes can be kept on the active list (ACT*'s equivalent of working memory). Nodes on the active list are not dampened, so they remain active as long as they are kept on the list.
• EVALUATION
• POSITIVE ASPECTS
• Accounts for semantic priming.
• Suggests retrieval from memory is abstractive. What is stored is not exactly what is seen or heard, rather it is meaning that is stored.
• Retrieval is constructive.
• Accurately portrays certain memory difficulties like TOT failures (with many links, activation weak but you know you know it).
• NEGATIVE ASPECTS
• Assumes sequential (serial) processing
• No simulation of memory has come close to duplicating human memory. Consider:
• --metaphor
• --sarcasm
• --exceptions
• The assumptions seem ad hoc.

FRAMES AND SCRIPTS AS SCHEMA

Frames and scripts represent two kinds of schemas which originated in Artificial Intelligence -- as ways of embodying in a machine, behavior which if observed in a human would be characterized as intelligent. In frames and scripts, one might think of knowledge as being stored in little packets or action sequences.

THE FRAME APPROACH

• Frames may be thought of as packets of organized and related information about a thing or event. Frames contain slots which are filled in from experience.
• Two typical examples of frames are the TRIANGLE FRAME and the CHILD'S BIRTHDAY PARTY FRAME.
• (click on each icon for examples) ******

SCRIPTS AND SHANK'S CONCEPTUAL DEPENDENCY THEORY (SAM)

The concept of scripts originated in AI studies of Roger Shank who was trying to write a computer program that would allow a computer to process printed stories. It was clear that the ability of humans to process stories is dependent on a context of implicit concepts about how sequences of events occur in the natural world -- conceptual dependency. Shank tried to incorporate this ability into his program. He began by trying to understand how people did it. In looking at human ability, he suggested that people use a set of expectations about sequences of events as the context which is why they don't need to be explicit. Since the notion of sequences of events in stories suggested drama, Shank called these scripts.

Scripts are simple, well-structured, event sequences that we use to interpret situations and as a basis for action. As a set of expectations, they allow us to infer important elements in a situation. Consider the following demonstration suggested by Matlin (1994):

• Read the following paragraph:
John was feeling very hungry as he entered the restaurant. He settled himself at a table and noticed the waiter nearby. Suddenly, however, he realized he'd forgotten his glasses.
• How are the last two sentences related?

CONCEPTUAL DEPENDENCY THEORY

• Scripts are built up as series of primitives which are basic actions that are incapable of being broken into smaller actions.
• This conceptualization is defnied as an actor doing something to some particular object.
• Scripts can be represented in machines by formalizing a theory of world knowledge.

AXIOMS OF SAM

1. For any two sentences that are identical in meaning, regardless of language, there should be only one representation.
2. Any information in a sentence that is implicit must be made explicit in the representation of the meaning of that sentence.
3. The meaning propositions underlying language are called conceptualizations. A conceptualization can be active or stative.
4. An active conceptualization has the form: Actor, Action, Object, Direction (instrument).
5. A stative conceptualization has the form: Object (in) state (with value).

PRIMITIVE ACTS OF CDT

• ATRANS
• Transfer of an abstract relationship such as possession, ownership, control (give, take, buy consists of two atrans)
• PTRANS
• Transfer of the physical location of an object (put, go is prtrans ones self).
• PROPEL
• The application of physical force (push, pull, throw, kick -- may cause ptrans)
• MOVE
• Movement of a body part of an animal by that animal -- nearly always an act (usually used instrumentally but not always -- kiss)
• GRASP
• Grasping of an object by an actor (grab, hold, throw, let go)
• INGEST
• Taking in of an object by an animal to inside that animal (eat, drink, smoke, breathe).
• EXPEL
• Expulsion of an object from the body of an animal (sweat, spit, cry).
• MTRANS
• The transfer of mental information between or within an animal. Memory partitioned into CP (conscious processor) and LTM. (see, tell, remember, learn).
• MBUILD
• Construction by an animal of new information from old (conclude, decide, imagine).
• SPEAK
• The actions of producing sounds. Many objects can speak, but humans usaually speak to mtrans (say, play music, purr, scream, bark).
• ATTEND
• The action of attending or focusing a sense organ towards a stimulus (listen, look at, see).

MEMORY ORGANIZATION PACKET (MOP)

• A recent construct that includes generalized groups of events called scenes.
• The MOP specifies how scenes are organized. The scenes, however, are relatively abstract and may serve in a number of scripts (the ordering scene for instance).
• Allows for flexibility. Scenes can be reorganized to explain adaptability when an inappropriate script is used initially.

SUMMARY

We have examined in this section, a number of ways in which the organization of semantic memory is conceptualized. From the rather flat structure suggested by simple models to sohpisticated organization put forth by network models and finally models suggested very closely tied to computer science.

Back to Cognitive Psych Home page Back to Topics