In this article we will discuss about the use of graph network structure to describe knowledge representation system.
Semantic nets, semantic network or associated network, is used to describe a knowledge representation system based on graph network structure. Originally they were developed for use as psychological models of human memory but now they are being used as standard methods for knowledge representation system in Artificial Intelligence and Expert Systems too. At the time of their origin they were used mainly in understanding natural language, where semantics (meaning) of associate words in a sentence was extracted by employing such nets.
A semantic net S/N consists of nodes connected by links called arcs, describing the relation between the nodes. The nodes in a semantic net stand for facts or CONCEPTS. Arcs can be defined in a variety of ways, depending on the kind of knowledge being represented.
Common arcs used for representing semantic nets Arcs represent relations or associates between concepts and included is a and has part. In natural language processing arcs are represented by AGENTS, OBJECTS, RECIPIENTS etc. (Fig. 6.18 and 6.19) A simple semantic network is shown in Fig. 6.15. As is clear both nodes and arcs are generally labeled.
As simple example consider, “NIS Arjuna is a ocean liner” and “Every ocean liner is a ship”.
These statements are represented in semantic nets in the Fig. 6.16.
We can know about the properties of the relations linking the nodes. For example, is a relation is a transitive one; we can infer a third statement from the net through this arc (relation). The INS Arjuna is a ship, even though it was not explicitly stated. The is a relation and others like the has-part, establish a property called INHERITANCE, of the semantic net, meaning that the items lower in the net can inherit properties from items higher up in the net.
This saves space since the information about similar nodes need not be repeated at each node. Instead it can be stored in a one central location entailing that an object or a concept may be a member of a class and it is assumed to have the same attributes (values) as the parent class, unless alternate values over-ride. Stated in an other way classes can also have subclasses or super classes which inherit the properties in a similar way. Instead it can be stored in a central location.
Property inheritance is over-ridden, when a class member or a sub-class has an explicit alternative value for an attributes as shown below:
Consider the parent class ‘Dog’ which has attributes, such has as tail, barks, and has four legs (Fig. 6.17). A subclass of that parent class may be a particular breed say Great Dane which consequently inherits all the attributes above, as well as having its own attributes; is tall.
A particular member (or instance) of this subclass, that is a, particular dog, Moti, may have additional attributes such as has brown colour. Property inheritance is overridden where a class member or subclass has an explicit alternative value for an attribute; specialized property should be given importance For example, Rottweiler as subclass of dog, the parent class, dog, may have the attribute “has no tail”, Deberman who has “cut tail”, French Buldog a subclass of parent class dog may have the attribute “snoars”. So property inheritance “bark” is over-ridden by “snoar” in the case of Deberman.
Alternative value may also be given at the instance level. For example, Rottweiler variety of dog has no tail, so as a subclass of the parent class dog its inherited property “has tail” is over-ridden by its value “has no tail.” Sharu is Rottweilerian dog owned by Ramu.
It is clear from the Fig. 6.17 that there are two types of nodes – Generic and Individual or instance. Generic nodes are very general nodes (canine or dog) because there can be many types of dogs or canines, whereas individual or instance nodes are members of these generic types (Moti, Basenji, Great Dane etc.).
Similarly, there are different types of arcs like is-a, has owns, instance etc. Of all these links, the link is a and has are the most important and links a generic node to generic node or a generic node to a instance (a special node). Also it may be noted there are two generic areas which support property inheritance, is a indicating class inclusion (subclass) and instance indicating class membership.
ADVERTISEMENTS:
Further, proper inheritance supports inference, in the sense that we can derive inference about an object by considering the parent classes. For example, in the ‘Dog’ network in the above figure we can derive the facts: “Great Dane has a tail and is carnivorous” from the facts that a dog has a tail and a canine is carnivorous, respectively. It may however be noted that we cannot derive the fact that a Basenji can bark since we have an alternative value associated with Basenji i.e., it does not bark.
We may also note how the network links together information from different domains (dogs and cartoon-character) by associations and hence the name associated nets.
In summary, we can remark that network representations are useful where an object or a concept is associated with many attributes and where relationship between objects are important.
Semantic nets have also been used in natural language research to represent complex sentences expressed in English.
ADVERTISEMENTS:
One example is the English sentence:
“Bobby gives Arisha a gift”
Its semantic representation is given in Fig. 6.18.
The arcs define the relationships between the predicate (GIVE) and the concepts, (such as, ARISHA and GIFT), associated with that predicate.
A semantic network of a more complicated sentence “Bobby told Krishan that he gave Arisha a gift” is given in Fig. 6.19.
The semantic representation is useful because it provides a standard way of analysing the meaning of a sentence. Also it provides the similarities in the meaning of sentences which are clearly related but have different structures. This is brought home by the fact that, although, the sentences in Fig. 6.18 and 6.19, look very different, the semantic nets representing the meaning of these two sentences look similar. In fact the semantic net of Fig. 6.18, is completely contained in net of Fig. 6.19.
Inheritance in semantic net can be explained with the help of following Semantic net which describes the properties of Snow and ICE.
Graph-based representation of network is a representation of properties of snow. This network, with the appropriate inference rules, can be used to answer a range of queries about snow ice and snowman. These inference are made by following the links to the related concepts.
Semantic net also implement inheritance, e.g., frostily inherits all properties of snow.
Semantic net encompasses a family of graph-based representations, the most common being conceptual graphs Sowa (1984). (Fig. 6.20)
Knowledge representing using semantic nets have similarities with other knowledge representing structures as well. The “isa” arc structure can be easily represented using a predicate logic, for example in Fig. 6.17, “isa” link can be represented in FOPL by-
Properties of Knowledge Representation:
1. Expressiveness:
Semantic nets allow representation of facts and relationships between facts. The levels of hierarchy provide a mechanism for representing general and specific knowledge. This representation is a model of human memory and it is therefore relatively understandable.
2. Effectiveness:
They support inference through property inheritance. They can also be easily represented by PROLOG, LISP and other AI languages making them amendable to computation, as shown by “isa” arc in LISP given by
3. Efficiency:
They reduce the size of the knowledge base, since knowledge is stored only at its highest level of abstraction rather than for every instance or example of a class. Further they help maintain the consistency in knowledge base because high level properties are inherited by subclasses and are not added for each subclass, as is clear from Fig. 6.17.
Explicitness:
Reasoning amounts to following paths through the network, so relationship and inference are explicit in the network lines.
Example:
Fig. 6.21., describes a ship semantic net. The parts of a ship, such as the engine, hull or boiler are stored once at top level, rather than repeatedly at lower . levels like ship type or particular ship. This saves huge amount of memory space. The net can then be searched, using knowledge about the meaning of the relations in the arcs, to establish fact like “the INS Arjun has a boiler”.
