3203. Introduction to Artificial Intelligence
Knowledge Representation
1. Knowledge
In general, AI systems increasingly depend on domain knowledge, and this tendency is being strengthened by the availability of more knowledge.
There are several dimensions by which knowledge can be classified into types:
- declarative, episodic, procedural
- commonsense, scientific, mathematical
All these types of knowledge need to be handled in AI, so they must be first represented in AI systems, so as to be processed. Requirements for knowledge representation includes expressiveness, modularity, naturalness, flexibility, and efficiency. Usually, tradeoff among the factors needs to be considered.
There is always a relationship between the form in which knowledge is represented and the way in which the knowledge is used. The study in knowledge representation has been influenced by several disciplines:
- In AI history, the mainstream representation approach is based in first-order predicate logic, though various attempts have been made to modify it to satisfy the needs of AI. According to this approach, knowledge is represented by a set of propositions and rules.
- In some mathematical models, knowledge is represented as points and trajectories in multi-dimensional spaces.
- In some computational models, knowledge is represented as procedures and data structures.
The relationships among different representations are important. For example, here are some tips
on translating from English into first-order predicate logic.
2. Structure of knowledge
Though in principle is is possible to see all the knowledge of a system as a set of propositions, it is not a good idea for implementation. For efficiency considerations, it is necessary to organize knowledge into larger structures. Here are some attempts:
- Semantic network:
a graph structure in which nodes (or vertices) represent concepts, while the
arcs between these nodes represent relations among concepts.
- Frame:
similar to object-oriented representation, a frame corresponds to a concept,
containing attribute-value pairs (sometimes with types and default values).
- Script: a script represent generalized episodes, or typical sequence of events corresponding to a concept that describes a type of event.
Here are some examples where these structures are implemented using Prolog:
- example 1 (search "Prolog code" in the page)
- example 2 (search "knowledge expressed in the semantic network" in the page)
One conclusion reached by many researchers in this study is that certain conceptual relations play important roles in knowledge, and therefore deserve special treatment. In particular, the relation between a concept and its subconcept forms the concept hierarchy that is crucial in problem solving. This relation is often called the "is-a" relation, and is used with a special inference rule that realize property inheritance.
Description Logic is an attempt to combine some of the above ideas with first-order predicate logic by balancing expressing power and inference efficiency. It is more expressive than propositional logic but has more efficient decision problems than first-order predicate logic. It can also describe concept hierarchies.
3. Knowledge base
Usually knowledge in a certain domain or for a certain purpose is collected together in a knowledge base. Beside providing retrieval capacity, the knowledge base is often equipped with an inference engine to generate derived knowledge. Also, the consistency and integrity of knowledge need to be maintained.
Representative knowledge bases in AI:
- CYC: constructing a foundation of basic "common sense" knowledge -- a semantic substratum of terms, rules,
and relations -- that will enable a variety of knowledge-intensive products and services. The language is based on first-order predicate logic.
- Open Mind:
an attempt of building a huge commonsense knowledge-base in English, participated by volunteers over the world.
- WordNet:
an online lexical reference system whose design is inspired by current psycholinguistic theories of human lexical memory.
- Mizar:
the project has built a database of mathematical knowledge, which includes more than 2 thousands of definitions of mathematical concepts and more than 30 thousands of theorems.
With the coming and growing of the Internet, structured online data are also widely used in AI systems:
- Semantic Web: the idea of having data on the web defined and linked in a way, that it can be used by machines - not just for display purposes, but for using it in various applications.
- Ontology: a catalog of the types of things that are assumed to exist in a domain of interest from the perspective of a person.