The Hierarchical Network Model | Semantic Knowledge

INTRODUCTION TO SEMANTIC KNOWLEDGE

We have discussed already in the previous lesson on knowledge of rules, facts, and meanings that are not tied to particular contexts. For example the fact that 'there are nine planets' is semantic knowledge. In this lesson we will examine the different models of semantic memory, which represents semantic knowledge.

Analyzing semantic knowledge is an enormously complex task. Some investigations ! have discovered basic representation and process, which will set the stage for analyzing more complex aspects of semantic knowledge.

Semantic organization refers to the way concepts are organized and structured in memory. In this section, we discuss the semantic organization in the context of three models; the hierarchical network model, the spreading activation model, and feature comparison model.


THE HIERARCHICAL NETWORK MODEL


This model of semantic memory was proposed by Collins and Quillion (1969, 1972). They studied the manner in which people comprehend simple sentences and make simple inferences. Collins and Quill ion proposed that semantic knowledge can be represented as a network of interconnected concepts. The following figure illustrates this model.


Fig. 21.a A hierarchical network representation for semantic knowledge pertaining to animals (From Collins and Quillons, 1869

In the above figure darkened circles are nodes that stand for concepts, for example, the concepts-animal, canary and shark. The concepts are arranged in a hierarchy. The -more inclusive a concept is, the higher it is located within the hierarchy. Thus, animal is · located above bird, for bird is a subset of animal. Similarly, shark and salmon are subsets· of a broader category of fish and fish is super ordinate to shark and salmon, so fish is place  higher in the hierarchy. 

The pointers emanating from the concept nodes designate two types of relations. 
They are:

a. The pointers that move toward superordinate concepts designate sub set relations. For example, the upward pointer from bird to animal indicates that bird is a subset of animal. This link connecting bird and animal is called "is a". In the same way the links connecting canary and bird, and ostrich and bird are "is a" links i.e. canary is a bird etc.

b. The smaller pointers designate attribute relations. For example, birds have wings, so bird and wings are connected by the link labeled has. Birds can also fly, and this is indicated by the link labeled can that connects bird and fly. 


Characteristics of the Hierarchical Network Model 

l. In this model the properties are ·stored in a very economical way.

2. Each property is represented only once-at the highest level.

For example, even though canaries, ostriches and salmon all breathe the 'breathes' attribute is stored at the most inclusive level possible, namely, the level animal. This approach eliminates that need to represent the attribute breaths many times, one for every animal.

3. Network representation depicts our knowledge as highly organized and interrelated.

4. The interconnections in the network provide pathways through which memory may be searched.


Collins and Quillian proposed that information is retrieved from semantic memory by searching the pathways in the knowledge-structure. For example, to determine the truth of 'A canary can sing', one needs to simply find the canary node and to retrieve the properties stored at that node. But comprehending 'A canary has skin' requires finding the canary node, moving up two levels to the animal node, and retrieving the properties stored at the animal node. The greater the number of levels one need to search through, the longer it takes to verify the statement.

Problems with the Hierarchical Network Model 

1. The model predicts that it will take longer to verify 'A bear is an animal' than 'A bear is a mammal' since animal is more levels above bear than mammal is. But in fact just the opposite occurs i.e. it takes longer to verify 'A bear is a mammal'.

2. The hierarchical network model fails to accommodate the effects of typicality some instances of a category can be verified faster than others, even though the instances are equidistant from the super ordinate category in the network. For example, robin and chicken both belong to the category of bird, but people verify 'A robin is a bird' faster than 'A chicken is a bird' because of the concept of typicality that was discussed in the preceding lessons.

3. Some of the studies conducted by Conrad ( 1972) contradicted the hierarchical
network models cognitive economy assumption. According to Conrad when the noun and property of a statement are strongly associated, one can verify the statements quickly irrespective of the levels in hierarchy. For example, the concept canary is more strongly associated with the property can sing than with has skin.

In spite of the above limitations the model may provide a valid account of some of our semantic knowledge.


THE SPREADING ACTIVATION MODEL 


The spreading activation model is the revised model of the hierarchical network  model This model was proposed by Collins and Loftus (1975). According to this model the memory network is no longer structured hierarchically. Instead, it is structured around the principle of semantic relatedness or semantic distance.



Fig. 2l.b A portion of the semantic memory network proposed by the Spreading Activation Model.

A small fragment of the memory network is shown in the above figure. Many of the concepts are interconnected as indicated by the lines running between the ovals. The shorter the line· connecting two concepts ie, the more closely related the two concepts are. Thus, bus is more closely related to school and to student than to accident or fire engine. The clusters of related concepts are found together.

The extent to which various concepts are related is determined through the measures such as ratings of typicality, subjective experience, norms for production frequency, and strength of association. Unfortunately there is no consensus about how to measure semantic relatedness. Like its predecessor, the spreading activation model includes labeled 'is a' links that indicate which categories are superordinate to others. For example daffodil flower are connected by an 'is a' link. The model also includes 'is not a' links, which indicate for example, daffodil is not a fruit i.e. daffodil and fruit are connected by an 'is not a' link. With 'is not a' links the memory search becomes quick. Because the knowledge that daffodil is not a fruit is stored directly in the network, this bit of knowledge is said to be pre-stored knowledge.

Assumptions Of The Spreading Activation Model 

1. Links differ in accessibility or strength; it takes less time to traverse a strong path than a weak one. How accessible or strong a link is depends on variables such as how frequently that link has been used.

2. When a concept is processed, activation spreads from that concept to neighboring ones. The spread of activation in the memory network depends on the strength of the initial activation, the proximity to the pint of activation, and the amount of time that has passed since the activation began.


Decision Process And Spreading Activation Model 

The model also includes a complicated set of decision process that will only be outlined here. If the sentence 'A car is a vehicle' were read, the car and vehicle nodes would be activated and spread. Various intersection points and links provide the evidence. The link between car and vehicle (an 'is a' link) would indicate that vehicle is superordinate to car, and this would provide strong evidence in favor of verifying the sentence. Similarly, in evaluating the sentence 'A bat is a bird' the decision process would discover an 'is not a' link between bat and bird. This information would allow the decision process to determine rapidly that 'A bat is a bird' is false.

In one of the experiment done by Meyer and Schvaneveldt ( 1971) subjects classified butter as a word faster if they had just classified bread than if they had just classified nurse. This outcome is called the semantic priming effect because the processing of one word prepares or primes the system for processing a semantically related word.

Conclusion 

Overall, the spreading activation model is an advance over the hierarchical network model. Network models have provided useful frameworks for conceptualizing human knowledge and are useful because of their potential scope and power.


THE FEATURE COMPARISON MODEL

The network model assumed that much of our knowledge is prestored in memory. In contrast, the feature -comparison model assumes that much of our knowledge is computed from the information stored in memory. For example the know ledge that a robin is a bird could be computed or inferred from other knowledge. Assume, for instance, that memory contains these two lists of features: 

Robin: living, animate~ feathered, rea-breasted ....... . 
Bird~ living, animate, feathered, two-legged ..... · 

Perhaps we verify 'A robin is a bird' by comparing these two list of features and deciding upon the degree of overlap. 

The feature-comparison model, proposed by Smith, Shoben, and Rips (1974), ··represents words as sets of features or attributes. The semantic features are assumed to vary along a continuum of definingness. At the one of the dimension are the features, called definite features, that are essential for defining a word. At the other end of the dimension are characteristic features that are characteristic of the concept but not essential to it. For example, the word bird has the defining features animate and has feathers. These feathers are essential for defining the concept. Bird also has characteristic features such as can sing or can be eaten by dogs. The later features are characteristic of many birds, but they are not essential for defining the word

Stages Of The Model: (or) Making Feature Comparisons 

The model asserts that verifying statements such as a robin is a bird occurs by comparing the features of robin and bird. The model includes two stages. In the first ,; stage, the overall featural similarity of the two nouns is compared. This comparison includes both the defining and the characteristic features. If the featural similarity is very ' high, as it is in the comparison of robin and bird, then the subject rapidly responds true. Similarly _if the featural  similarity is very low as in the compares off of robin and car, then the subject rapidly responds false. 

In many instances two nouns will be neither very high nor very low in featural similarity. If the first stage had determined that there is an intermediate level of similarity between the two nouns, then the second is begun. In this stage, only the defining features of the two nouns are compared. If all of the defining features match exactly, then the subject responds true. But if any of the defining features of one noun do not match with those of the other noun, the subject responds false. Characteristic features are essential for the occurrence of typicality effect. 

Problems With the Model 

1. The model does not include rules that stipulate which features are defining and which are characteristic. 
2. People have a variety of strategies for disconfirming sentences, and comparing features may be only one useful strategy. 


CONCLUSION OF THE ABOVE THREE MODELS 


Research concerning our knowledge of word meanings is of recent origin, and it is too early to decide the extent to which our knowledge  is prestored or computed. Most likely, a particular piece of knowledge can be both presorted and computed. Human knowledge is too diverse and flexible to be categorized as strictly prestored or computed. Future research, then should take this flexibility into account. In the long run, theories of semantic knowledge must clarify how people understand complex sentences and stories, remember prose passages, and so on. 


SUMMARY 


Three models have been put forth to explain how people make simple inferences and organize their semantic memory. In the hierarchical network model Collins and Quillion proposed that semantic knowledge can be represented as a network of interconnected concepts which are arranged in a hierarchy. Super ordinate concepts are arranged at a higher level and the concepts, which are at a lower level, are called ·as subsets. Network representation depicts our knowledge as highly organized and interrelated. 

In the spreading activation model knowledge is represented in a network containing concepts and links that interconnect different concepts. The links are labeled and indicate the relationship between two connected concepts. The more closely related two concepts are the closer together they are in the semantic network. Verification of sentences such as a bear is an animal occurs through process of spreading activation. 

A rival model, the feature -comparison model, proposes that sentences can be verified in either one or two stages. In the first stage subject compares both characteristic defining features. In the second stage only the defining features of the two nouns are I compared. The defining features are necessary for the defining of the word, whereas characteristic features are properties that occur often but are nonessential. People use these models as complementary to each other in various contexts. 

Scroll To Top