This article throws light upon the top two models of semantic memory. The models are: 1. Hierarchical Network Model 2. Active Structural Network – Model 3. Feature-Comparison Model.
1. Hierarchical Network Model of Semantic Memory:
This model of semantic memory was postulated by Allan Collins and Ross Quillian. They suggested that items stored in semantic memory are connected by links in a huge network. All human knowledge, knowledge of objects, events, persons, concepts, etc. are organised into a hierarchy arranged into two sets. The two sets are superordinate and subordinate sets with their properties or attributes stored.
These properties are logically related and hierarchically organised. The following illustration explains the relationship between the sets – super ordinate for dog is an animal, but it is a mammal too; belongs to a group of domesticated animals, a quadruped; belongs to a category of Alsatian, hound, etc. Let us look at Collins and Quillian study as an example for a better understanding of this model.
In this hierarchically organised structure one can see that the superordinate of canary is bird, of shark is fish and the superordinate of fish is animal. One can notice further that a property characterizing a particular class of things is assumed to be stored only at the place in the hierarchy that corresponds to that class. This assumption forms the basis of the cognitive economy.
For example, a property that characterizes all types of fish (the fact that they have gills and can swim) is stored only at the level of fish. It should be noted that gills and other such features are not stored again with the different types of fish (salmon, shark, etc.) even though they have gills. Similarly, a bird which is the superordinate of canary is an animal. Specific properties are stored only at appropriate levels in the hierarchy.
Given this hypothesized network structure, Collins and the Quillian’s next task was to determine how information is retrieved from the network. To answer this question an experiment was carried out in which subjects were asked to answer ‘yes’ or’ no’ to simple questions.
Consider, for example, the following questions about canaries:
1. Does a canary eat?
2. Does a canary fly?
3. Is a canary yellow?
The three questions mentioned above may be challenged by the semantic level at which the information needed to answer them is stored. Consider the first question, “Does a canary eat?” The information “eats” is stored at the level of animal, two levels away from canary. Likewise, the information has “wings” and is “yellow” (needed to answer the second and third questions) are stored at one and zero levels away from canary, respectively.
The major point of interest in this model of Collins and Quillian was the reaction-time or time taken to respond to the questions. Results of the experiment revealed that with the increasing level of information it takes increasing amounts of time to retrieve the information.
Their explanation about this is as follows- in order to answer the third question, the subject must first enter the level in memory that corresponds to ‘canary’ and here find the information that canaries are yellow. The question is, therefore, answered relatively fast. To answer the second question the subject still enters the memory level that corresponds to ‘canary’ but does not find any information at that level concerning whether or not canaries fly.
However, the subject moves up the hierarchy to the level where information about birds is stored and there finds that birds fly. This is done by combining the information that canaries are birds and that birds fly and then the question can be answered. Due to the extra step of moving up the hierarchy, question two takes somewhat longer to answer than question three.
The first question takes even longer for the same sort of reason. To answer question one, the subject cannot use any of the information that is stored at either the level of ‘canary’ or ‘bird’ but must move up to an additional level in the hierarchy to ‘animal’. Thus, it was concluded that, because a canary is a bird and a bird is an animal and animals eat, the canary must eat too. Therefore, the reason why some questions take longer to answer than others is that some questions require more travelling in our memory from level to level in the semantic hierarchy.
Using a similar rationale Collins and Quillian predicted that it takes less time to answer “Is a canary a bird?” than to answer “Is a canary an animal?” We see in the figure that to answer the latter question, a subject must move up two levels from canary to animal, whereas to answer the former question, the subject must move up only one level.
It was revealed that on an average, people take about 75 milliseconds longer to answer the question, “Does a canary eat?” than to answer, “Does a canary fly?” and about 75 milliseconds longer to answer the question about flying than to answer, “Is a canary yellow?” 
2. Active Structural Network – Model of Semantic Memory:
The active structural network model postulated by Norman & Lindsy can be understood by their analysis of two simple sentences. Let us now see how they go about explaining it. Peter put the package on the table. Because it wasn’t level, it slid off.
These sentences refer to objects, person and events. Figure 10.9 shows the diagrammatic sketch representing information in a semantic network. This network consists of information expanded in terms of events, instances of the movements involved or modes of their relations, the direction of the relationship, etc. This elaborate network representation is said to form the basis of human memory.
Let us consider the figure for a moment. The basic conceptual information shows that Peter caused the package to move from its earlier location to the top of the table, and that gravity was the causal agent that then acted upon the package causing it to move from the table top to the floor.
The first movement is represented by a node, the oval numbered. The oval (or words in the figure) are called relations. The relations show how the different node structures in the figure are related to one another. Thus, looking at the node we see that it represents an instance of the act of ‘move’. This particular instance of ‘move’ has its cause – Peter (shown diagrammatically) and the object being moved is package (again shown diagrammatically). The location to which the moved object is placed is the table.
The second node, the oval labelled 2, is another instance of ‘move’. Here the cause is gravity, the object is the same, i.e. the package, and the movement takes place from a ‘From’ location, (the table-top) to a ‘To’ location (the floor).
The drawings of the package and Peter are instances of the nodes that are named “package” and “Peter”. The representation shown and described can further be elaborated. Peter put a package on the table, an event of which Peter was the agent, caused the result that causes the package to change its location from place unspecified to a new place, on top of the table. It changed its place because the first position was higher than the second position.
Moreover, the movement was caused by the force of gravity. In a similar fashion detailed analysis can be carried on and on. But the conceptual network presented here is assumed to be sufficient enough to give us an idea about how words and events create relationships, concepts, etc. and form a complex network. Thus, one can see that this model of semantic memory conceives of human memory as a giant network of interconnected nodes, and these nodes are assumed to correspond to individual concepts, ideas, or events in the system.
3. Feature-Comparison Model of Semantic Memory:
E.E. Smith, E.J. Shoben and L.J. Rips postulated a theory in which emphasis was laid on semantic features. Their assumption was that there are two distinct types of features.
First, there are those features which are essential aspects of the item’s meaning. These are known as defining features.
The second type of features do not form any part of the item’s definition but are nonetheless descriptive of the item and are referred to as characteristic features. For instance, if we take the word Robin, there are some features true to Robins, such as that they are ‘living’, have ‘feathers’, have ‘wings’ and have ‘red-breasts’. All these are defining features.
Other features, however, may be associated with robins, but they are not necessary to define a robin. These include features such as ‘like to perch on trees’, ‘undomesticated’, ‘harmless’ and ‘smallish’. In situations where a subject must decide whether an instance belongs to a specific category (for example, deciding whether a robin is a bird), it is assumed that the set of features corresponding to the instance and category are partitioned into the two sub-sets corresponding to defining and characteristic features. Figure 10.10 illustrates the above features.
Now this process of verifying whether an instance belongs to a category, i.e. in this case ‘is a robin a bird?’ is assumed to be accomplished in two major stages as given in the figure. The first stage involves a comparison of both the defining and the characteristic features of the instance and the category to determine the degree to which the two sets of features are similar. If there is a high degree of correspondence between the instance features and the category features, the subject says “yes” immediately.
If the two sets of features have very little correspondence (low similarity), the subject can say ‘no’ immediately. However, if there is an intermediate level of similarity between the features of the instance and the features of the category, then a second stage is needed before the subject can reach a decision. In the second stage, the subject compares only the defining features of instance and then a ‘yes’ response is made, otherwise the subject says ‘no’.
Smith et al. extended their model further by including the concept called typicality effect. When a subject is asked to verify whether an instance belongs to a category, say birds, one is consistently faster in verifying some instances, for example, robin, canary, than chicken.
The faster instances are those that are judged by other independent subjects to be more typical of the category. If the instance to be verified is highly typical of the category, the two share a large number of features, both defining and characteristic.
When it is discovered during stage one that the instance and category have largely overlapping features, the subject can make an immediate response without executing stage two. For atypical instances in contrast there is not much overlap in terms of the characteristic features. Stage two must, therefore, be executed and response-time is accordingly longer.
Though these models have been built on highly scientific lines with detailed analysis, they are not free from certain limiting factors. Rips Shoben and Smith criticising Collins and Quillian pointed out that most of the college students know what a mammal is and if we add this concept to a hypothetical network that contains collie (a dog of specific breed), dog and animal, it is placed between dog and animal. In a semantic hierarchy, mammal is closer than animal to either dog or to some particular type or breed of dog (for example, collie).
According to the Collins and Quillian model a person should answer the question “Is a collie a mammal?” faster than the question:
“Is a collie an animal?” They found that people do not react as predicted by Collins and Quillian. Similarly, people take longer to answer the question “Is a potato a root?” even though vegetable is logically closer to potato in a semantic hierarchy.
The concept of cognitive economy was criticised by Conrad. She simply asked subjects to describe a canary as a bird, an animal and so on. She then tabulated the frequency with which various properties were mentioned.
It turned out that the properties frequently associated with canary (such as the fact that they are yellow) were the properties presumed by Collins and Quillian to be stored directly at the canary node whereas the properties that Conrad found to be less frequent were presumed by Collins and Quillian to be stored with bird or with animal.
She concluded that property frequency rather than the hierarchical distance determines the retrieval-time. The active structural network model has been criticised on the grounds that it expresses semantic memory through a gigantic network which is so expansive that the underlying conceptual framework cannot be presented in a representational system.
Collins’ criticism against the feature comparison model is that the distinction between defining and characteristic features poses an inherent difficulty – there is no feature that is absolutely necessary to define something.
For example, if a person removes the wings of a bird, it does not cease to be a bird. If the feathers are plucked from a robin, it does not stop being a robin. Furthermore, people do not appear to be able to make consistent decisions as to whether a feature is defining or characteristic. Is “having four legs” a defining feature of tables? What if you see a table-like object with only three legs?
Do you still call it a table? Smith and his co-workers realised the meaning underlying the questions but continued to maintain this artificial distinction between defining and characteristic features. With all these loopholes, we still see the contribution of these models to various fields of human and material world as something incredible. There are a few other models like the Human Associative Model propounded by Anderson and Bruner.