Introduction to semantic primitives

by Beate Grawemeyer
School of Cognitive and Computing Sciences
The University of Sussex


Contents


Introduction



The first step in developing a Knowledge Based System is the selection of a proper representation formalism. The next important design question examines the concepts which are to be defined, such as the definition of predicates, functions and constants use and meaning, in a logical knowledge representation. For example in the application of frames, individual frames and their attributes have to be specified.

Many of the concepts to be defined, can be reduced to other concepts already defined. 'Running' for example, might be defined as a 'quick movement of a person by foot'. Semantic primitives are the most basic concepts, which can not be reduced to other concepts.

During the development of a knowledge base for a particular domain a set of such knowledge units must be stated. A default set of basic knowledge units, the semantic primitives, are used to build more complex units, which are independent of one-other.

The selection of primitive elements for a presentation of knowledge in a particular domain might be seen as a basic problem, which is independent of the representation formalism (for example, predicate logic, frames, semantic networks, production rules or procedures). The selection of primitives determines the expressiveness and the problem solving abilities of the system.
 
Some knowledge representation systems provide the specification of primitives and defined concepts. For example KL-ONE.

There exists several approaches (Wilks, Schank), in the area of the development of natural language processing systems, which have tried to use a basic set of primitives, to represent the meaning of any natural language sentence.




b a c k


Wilks System



Wilks states a set of around 80 concepts as semantic primitives as a basis for the development of a mechanical translation system. He divided these concepts into five classes. Some examples of such primitives are:

  • entities
      man (human being)
      things (substances)
      part (parts of a things) etc.

  • actions
      causes
      be
      flow (moving as liquids do) etc.

  • cases
      to (direction)
      in (containment) etc.

  • sorts
      good (being morally acceptable)
      cont (being a container) etc.

  • type indicators
      how (being a type of action) etc.


The use of such semantic primitives concerning the representation of the meaning of natural language predicates does not solve the problem of ambiguity and uncertainty in natural language.

There also does not exist a 'correct' set of primitives for a particular language or domain. The proof of the adequacy of such a primitive set can only be done operational or experimental. The adequacy of a primitive set will be shown, if it is possible to present the knowledge in terms of the primitives, in a simple and meaningful way, and to solve the range of possible problems which the task may involve.




b a c k


Schank´s Conceptual Dependency Theory



Roger Schank, who also deals with the development of natural language processing systems, tried to identify a small number of primitives, which would make it possible to represent every possible action. While Wilks work concentrated on the area of language translation, Schanks aim was to develop a domain and task independent basis for knowledge representation. Schank views this approach as an intuitive theory of human language processing.

Schanks aim was to produce a  knowledge representation formalism, which when interpreted, was free of any ambiguity and uncertainty. To achieve this aim Schank defined the following eleven primitive actions (ACTs):

  • physical actions
      PROPEL application of physical force to an object
      MOVE  movement of a body part by its owner
      INGEST ingestion of an object by an animal (e.g. eat)
      EXPEL expulsion of something from the body of an animal (e.g. cry)
      GRASP grasping of an object by an actor

  • actions, which might be characterized through the resulting state changes
      PTRANS transfer of the physical location of an object
      ATRANS transfer of an abstract relationship

  • actions, which might be seen as an instrument for other actions
      SPEAK production of sounds
      ATTEND focusing of a sense organ toward a stimulus

  • mental actions
      MTRANS transfer of mental information
      MBUILD building new information out of old (e.g. decide)

In addition to the primitive actions (ACTs) of the CD-Theory there are other concept categories, which allow other dependency structures to be built:

  • Picture Producers (PPs),  physical objects

  • Picture Aiders (PAs),  modifiers of PPs, attribute of objects

  • Action Aiders,  modifiers of ACTs,  attributes of actions

  • Tenses

  • Places

For example the sentence: ' I gave the man a book '
will be shown in CD-Theory as:

The Symbols of this example have the following meaning:



The double arrow indicates a two-way links between actor and action.
The letter p over the double arrow indicates past tense.



The arrow indicates the direction of dependency.
o indicates the object case relation.



Describes the relationship between the action (ACT), the source and the recipient of the action.
R indicates the recipient case relation.



In addition to the rules above, there are many more rules in Schank´s CD-Theory, each with their own corresponding graphical description. For more details see (Schank,1973).
In CD-Theory the process of splitting the knowledge into small sets of low-level primitives makes the problem solving process easier, because the number of inference rules needed is reduced.

A disadvantage of the CD-Theory might be seen in the fact, that the whole knowledge has to be slit into several small units. This is often very inefficient and sometimes impossible.
The CD-Theory is a theory to describe events. To use this theory as a universal model in knowledge representation, it must be able to represent all knowledge. Some attempts have been made to expand the CD-Theory to achieve this goal, and these take similar approaches to describe physical objects and social actions.


b a c k


Conclusion



When selecting knowledge representation formalism, the primitive concepts (lowest level) must be stated, as these are responsible for building higher level concepts. The selection of primitives determines the abstraction level, the detail facility and the problem solving ability. Through the selection of a primitive it will be stated how detailed knowledge will be described, is influenced by the kind of problems that has to be solved.



b a c k

Other Resources



Useful References

  • Knight, R. (1991).Artificial Intelligence (2nd ed.), New York: McGraw Hill, Inc.
  • Partridge, D. (1991). A new guide to artificial intelligence,New Jersey: Alex Publishing Corporation
  • Schank, R. (1973).Identification of conceptualizations underlying natural language. In Computer Models of Thought and Language,Schank, R. and Colby, K., editors, San Francisco:W. H. Freeman Co.
  • Wilks, Y. (1772). Grammar, meaning and the machine analysis of language,London: Routledge & Kegan Paul

People

On-Line Resources


b a c k
© 1999 b.grawemeyer