There are two major classes of models of semantic memory:

(1) One assumes that people compare the features of two categories to determine their relationship.

(2) The other assumes the relation between two categories is stored directly in memory in a semantic network, which consists of concepts joined to other concepts by links that specify the relation between them. This model makes use of the concept of spreading activation: activation of one concept leads to activation of related concepts as activation spreads along the paths of the network.

Hierarchical organization can facilitate the recall of semantic information.

Recall of Hierarchical Information

Bower, et al (1969) presented subjects 112 words: 4 groups of 28 words related by category membership. Half saw the words grouped into categories, the other half saw the words randomly arranged.

They were then to recall the words in any order.

Across four trials, subjects in the organized condition significantly out-performed subjects in the randomized condition.

Thus, subjects could organize the input into hierarchies. Such organization allows one to structure memory in such a way that it can later be searched more efficiently.

Category Size

Each category in a hierarchy is divided into smaller categories. This raises the question of what makes for maximal category size.

The advantage of grouping is reduced if the categories are too small; it is also reduced if there is too much information to remember.

Several experimenters, using different experimental paradigms showed the ideal category size is 2 to 5 items.

One paradigm for testing semantic know-ledge is to have subjects answer true or false, as quickly as possible, to simple statements like, 'a bird is an animal'.

Response times to different types of statements give an insight into the organization of semantic knowledge.

This is an example of using reaction time (RT) as a dependent variable in cognitive research.

Two findings from such studies are:

people can verify that an instance is a member of a basic-level category faster than they can verify that it is a member of a superordinate category: a canary is a bird is faster than a canary is an animal.

people can verify more typical instances faster than they can verify less typical instances.

An important related finding is that of category size effects--the fact that people can classify a member into a smaller category faster than into a larger category.

That is, the smaller category is reached sooner because it appears lower in the hierarchy.

Thus, 'canary is a bird' is responded to more quickly than 'canary is an animal'.

Two models have been proposed to account for these findings.

Features true of all category members are stored at the highest level. Features that apply to basic-level categories are stored at an intermediate level. Properties stored at the lowest level are true for that particular member but not for all members of the category.

Information is not repeated at each level, but only appears once, making this an economical way to store information. However, retrieval may be more complicated because we may need to access more than one level to access the necessary features to decide category membership.

The model has two assumptions:

it takes time to move from one level of the hierarchy to another; and

additional time is required to retrieve the features stored at any one of the levels.

These assumptions are demonstrated with sentence verification--it does take longer to verify a sentence in which the two components each come from a different level & it takes longer to verify sentences when features need to be retrieved.

This model makes predictions of underlying physiological mechanisms because retrieval is facilitated when a previous question requires information from the same category level.

If two consecutive verification sentences have information about a category member stored at the same level, verification of the second is faster--putatively because verification of the first already activated that level.

If the second sentence requires information stored at another level you don't get this facilitation.

There are two findings which do not fit the model:

Sometimes verification does not follow naturally occurring levels of a hierarchy--it takes longer to verify that a monkey is a primate than to verify that a monkey is an animal; and

(2) the model does not account for the typicality effect--it takes longer to verify that an ostrich is a bird than it does to verify that a canary is a bird.

Thus, the meaning of words can be represented in memory by a list of features which are used to define categories; but they vary in the extent to which they are associated with a category.

The most essential features are called defining features--those an entity must have in order to be a member of a category (ex: BIRD: wings, feathers, lays eggs). The rest are called characteristic features--those that are usually possessed by category members but are not necessary (ex: BIRD: flies).

According to the feature comparison model decisions are made in two stages:

all features of the two concepts are compared to determine how similar they are to one another. If they are very similar or very dissimilar a true or false decision can be made.

If features are intermediate then only the defining features are examined to determine whether the example has the necessary features of the category--grouping hinges on similarity and not on category size.

Because the second stage is not necessary if the concepts are similar, the model predicts that the more typical members of a category should be classified more rapidly than the less typical members. This ability to account for typicality is an advantage of the feature comparison model over the network model.

This model also accounts for reversals of the category size effect--the comparison can be made more quickly because its predictions are based on similarity rather than on category size.

Limitations of the Feature Comparison Model

One of the problems with this model is that it relies on ratings to make most of its predictions. Some of these are weak, i.e., people rate how similar an item is to the global concept of the category.

A second problem is that all classifications require computations. This seems cumbersome and seems better explained by the network model which assumes association come to play a role.

A third problem regards defining features. There is little direct support that people can in fact identify the more defining features of a category.

Spreading Activation Theory

This first model was developed by Collins and Loftus (1975). The emphasis was on concepts joined together by links that show relationships.

The length of each link represents the degree of semantic relatedness between two concepts.

When a particular concept is 'activated' nearby concepts become activated as well.

This model assumes that when a concept is processed, activation spreads out along the paths of a network, but its effectiveness is decreased as it travels outward. It predicts typicality because more typical members will activate the superordinate category sooner than less typical members, i.e., have a shorter link.

Experimentally this model can be assessed with RT studies by the assumption that "spreading" of activation takes time and so less associated concepts take longer to get to and more associated ones take less time. This shows how we use RT as a DV.

This model also alludes to underlying physiological mechanisms with the concept of semantic priming, which is much like the concept of facilitation of recall we saw in the Collins & Quillian model.

Priming occurs when a decision about one concept makes it easier to decide about another concept.

Using a lexical decision task (decide if a string of letter is a real word) Meyer & Schvaneveldt (1976) showed that the spreading activation model suggests that the presentation of a word activates related words (BUTTER activated by BREAD but not by NURSE).

General criticisms of the model are that it makes too many assumptions while making too few predictions, and that it fails to account for spontaneous recall after an initial failure. These criticism are both fairly weak.

Value of Semantic Network Models

While on the one hand they are very flexible and explain many findings, it is at the expense of being so flexible that they explain almost any findings, and so cannot be falsified and lose their predictive power.

The test of a 'good' model is that it predicts what should not happen.

So, when hierarchical network models were revised into spreading activation models, the added assumptions corrected limitations at the expense of sacrificing precision.

A.)     Schemata
          Bartlett (1932): schema (plural = schemata)--active organization of past reactions or events.
          Schemata organize our past knowledge or a particular set of material--not very popular at a time
          when introspection had just been abandoned and behaviorism was rising.

         Schemata--basis for reconstruction in memory--also help at encoding, allowing inferences about
         information as it is being encoded.
         Scene schemas --tell us what to expect in certain places--may have some negative effects on
         retrieval, e.g.: Brewer & Treyens (1981) asked subjects to wait in an office which lacked typical
         office furnishings but had atypical items.

Then asked to describe the office.  They were good at recalling schema-consistent items, but poor at recalled schema inconsistent items. They also falsely recalled items that were absent but were schema-consistent.

B.) Value of Schemata

Several theories have different interpretations of the value of schemata, but concur that having a schema has several effects, including:

• 1. Selection - top down selection of relevant  stimuli from the environment.
• 2. Abstraction - coding of central ideas (gist)  without need to encode details.
• 3. Framework Effects - framework for encoding  new information
• 4. Interpretation - interpret incoming informa- tion from inferences made during encoding.
• 5. Integration - integrate the new information,  which can make changes to existing schema
• 6. Guides Retrieval - helps us locate    information in memory.
• 7. Output Editing - allows your output to be  appropriate to the contextual situation.

An event schema--refers to stereotyped knowledge about routine activities as they relate to particular situations.  Most sequences of events for frequent activities are pretty standard--no need to recall all events that take place on each occasion, only those that are not standard.
Scripts help retrieval of unusual aspects of a situation, and free us from retrieving usual aspects.

Schank & Abelson (1977): subjects who read script-based stories which contained some unexpected events (called obstacles or distractions) would better remember these interrup tions than the routine events.  This was confirmed:  Subjects recalled 53% of the interruptions; 38% of the script actions.

D.) Scripts:  Internal Organization

A key question is whether scripts are temporally or spatially organized.

If asked to list the important activities related to a particular event most subjects will make their list in a temporal or spatial fashion.

There is also evidence that scripts are goal-organized.  Galambos & Rips (1982) discriminated between the premises that scripts can be organized according to the centrality of an activity in achieving a goal, or according to the temporal order of an activity.

The former can be represented as a semantic network with sequences being nodes and degree of association between a script and an activity being the link.  The latter are represented as sequential nodes.

To test which internal structure better represents scripts, Galambos & Rips had one group of participants rank-order the component activities of different scripts.  Next they showed another group of participants a pair of items, the name of the script and an activity,

These participants were then asked to respond if the activity was part of that script.

More central events were verified more quickly than less central ones; earlier ones were not verified more quickly than later ones.  So the centrality hypothesis was supported by the data.

They found that goal-directed activities were recalled 70% of the time, but the same actions not embedded in a plan were only recalled 30% of the time.

So now, the task is to relate this back to the Brewer & Treyens (1981) study above! On the one hand schema consistent activities were recalled better, but on the other, schema inconsistent elements in an environment were recalled less well.

E.) Themes and Story Structure -  When the Central Idea is NOT Part of "Prior Knowledge"

Stories can be broken down into a setting, a theme, a plot and a resolution.  Thorndyke (1977) had subjects read stories where the theme was in its usual place, early in the story--in which case recall was best; moved to the end of the story--in which case subjects recalled less information; or left out, in which case they recalled even less.