Knowledge Representation | Research & Encyclopedia Articles

Knowledge Representation

In order to use in a computer system any information that is available in the real world from empirical observation, it is necessary to encode the information in some computer-usable format. For the most basic raw information such as numbers, strings of alphabets, and the like, this is a problem easily solved. However, as it gets increasingly desirable to use computer systems in sophisticated ways, it is likewise increasingly important to find ways to encode knowledge (as distinct from information). Knowledge may have many aspects, and is generally defined by a much higher degree of sophistication than bare information. Knowledge also is typically context-dependent by nature and importance.

The field of knowledge representation (abbreviated KR) is thus a broad-based one lying at the intersection of a number of distinct and important disciplines--philosophy, linguistics, psychology, computer science, cognitive science, and mathematical logic, and others. Usually, KR is considered to be a problem or area within the domain of artificial intelligence, one of the important branches of computer science. Since interest in KR is largely, if not totally, based on the desire to solve practical problems, research into it generally ignores purely theoretical problems or concerns arising from philosophy of language, philosophy of mathematics, etc., and instead focuses on engineering and performance issues.

The engineering approach to KR, sometimes called knowledge engineering, takes on either the formalisms and specific techniques used to code knowledge, or the specific type of information that is so encoded. Many AI systems in use today apply ad hoc or stopgap means of knowledge representation that are tailored for a particular application (such as a medical expert system that can offer possible diagnoses given patients' symptoms). Many AI researchers have noted that humans, on the other hand, have a generalized system of understanding and are able to draw appropriate conclusions from a wide variety of data, without being parochial about representations. For instance, a human can understand the social graces of everyday interaction, the rules to be followed in driving, the professionalism and strategies to be used in a work environment, and the personals choices available and reasonable for leisure. Each has a distinct niche, but a human can understand them well, and can rapidly draw appropriate conclusions in any one of these settings. To emulate such a multifarious capability in an computer, researchers have suggested that KR work should focus on the creation of formal systems that allow for generalized, rather than specialized, strategies for representation. In 1985, J.R. Hobbs, a linguist, urged that this concept, which he called ontological promiscuity, be achieved.

To a certain extent, work on knowledge representation is tied to and inspired by the computational theory of the mind. This theory considers the human mind to be a digital computer--a discrete, finite-state device that stores symbolic information in some unknown language, and represents them according to syntactic rules. Thoughts and other mental processes are nothing but symbolic representations in a language of thought. Mental processes are causal sequences of symbolic processing, and are driven by the syntactic rather than semantic properties of the symbol streams. This theory was first proposed by the philosopher Hilary Putnam in the 1970s, but has since been articulated by others. Although this theory has serious flaws--the most obvious one being that it is all hypothetical and nowhere backed by hard empirical data (since there is no scientific experiment that can validate it, and no way to know what the symbolic language of thought, etc., are)--it has some vociferous backers, though these are mostly outside the mainstream of computer science. The proponents of the theory hold that theirs is the only game in town, as it were--that they have the only plausible explanation for cognition that does not involve an inner homunculus or other non-physical entity. (However, many programmers and logicians remain unconvinced of the theory, noting that its glib explanation ignores insoluble computing and logical problems.)

Based on the computational theory of the mind, knowledge representation work aims to achieve a formalism for knowledge that is similar to the one actually found in the human mind. After all, if the human mind is just a computer, then it should be possible to program real (non-human) computers similarly to get the same results as well. In this direction, the many formalisms used in knowledge representation try to create a precisely defined syntax and semantics, as well as an inference procedure that is computationally tractable. Quite often, a notation for describing concept hierarchies and mechanisms to achieve property inheritance are included.

The logics used in knowledge representation may be the sentential (propositional) logic or the predicate (first-order) logic, or some variation on these. Some measure of ideas from fuzzy logic, Bayesian or probabilistic inference, and the like are also attempted in some cases. Concepts from object-oriented programming languages (such as inheritance) are not uncommon.

One of the central concerns in knowledge representation, which indeed is a primary research problem in the larger domain of artificial intelligence itself, is the Frame Problem. This problem emerged as an annoyance in the 1950s, and has since been recognized as a major deficiency in the simpler syntactic models of knowledge representation. In order to represent the status of a changing world, it is important to know just which of the many variables does in fact change. A human has no problem with this--for instance, it is obvious that turning on the oven does not change the color of the walls in a room--but a computer does not know this by default. Unless there is a specific rule, called a "frame axiom," to clarify in each instance that something does not change, the computer is unable to deduce, based on syntactic reasoning alone, that there is any persistence in the world of its reasoning. As the number of variables in a system grows, the number of frame axioms grows exponentially, until even for systems of moderate complexity, an enormous number of frame axioms must be specified. Even when specified, processing is rendered slow because for each possible change, the computer has to go through the long list of frame axioms before deciding the next state.

A common solution to the frame problem is to assume that persistence is the default--in other words, to assume that no change occurs in any variable unless there is a specific axiom that describes the change. However, if the system then has to learn, based on new evidence, that a change in one variable does affect another (for example, through research that shows that a drug used to fight obesity also causes mood swings), there is a problem. To handle this situation, much work has been done in developing specialized non-monotonic logics that allow for assumptions of no change to be withdrawn if this is proven necessary.