Paper Structural representations of unstructured knowledge Wiesław Traczyk Abstract—Knowledge should be represented in a formal, Knowledge management consists in organization and facili- structured manner if we want to process and manage it. Un- tation of knowledge generation and utilization, with a main fortunately a source knowledge presented in many documents task – increase of institution profit. has informal, unstructured shape. The goal of these consider- In the last two domains a form of primary knowledge can ations is to present the methods of translation from the textual, be very diversified. The low and company regulations, in- unstructured knowledge to the structured knowledge, preserv- stitution rules, medical procedures, web pages, fragments ing textual form. of books and other documents written in a natural language Keywords— textual knowledge, knowledge representation lan- are frequently the main sources of knowledge. But, if the guages, ontology. knowledge is to be processed by computers, its form has to be converted to a formal shape, with precisely defined syn- tax and semantics. Then, the proper knowledge representa- tion (KR) is needed, easy to obtain from natural language 1. Introduction descriptions and easy to understand by computers. General demands for such a representation are usually summarized Knowledge is used in all areas of human activities but there by the ontology: a set of formally specified concepts (such are domains, relevant to computer applications, in which as things or events) and their properties and relations that the word knowledge is located in the name of a branch. describe a domain of interest, in order to create an agreed- Three more popular such fields, with their links to data, upon vocabulary for exchanging information. knowledge and human agents, are presented in Fig. 1. Since the goal of KR is to express knowledge in computer- tractable form, the two questions should be answered before discussing the possible solutions: – what does it mean “knowledge”? – what are the demands from languages used for knowl- edge representation? There is no universally accepted definition of knowledge. Most likely the cause of it lies in very big diversity of the notion. Knowledge (in its intuitive meaning) can be descriptive or procedural, explicit or tacit, qualitative or quantitative, individual or collective, and so on (even east- ern or western). It is essential to determine the need that the knowledge must fulfill. It can be used for problem Fig. 1. Knowledge generation, processing and utilization. solving and decision support, for inference, as an interlin- gua for cooperating agents or software modules, to define Knowledge discovery looks for specific, previously un- semantics for natural language interpretation, etc. known but important, regularities in large data bases. If the range of knowledge application is limited to knowl- These regularities (or patterns) determine a new knowl- edge engineering and management, the following definition edge gained from data. Different forms of knowledge ob- seems to be appropriate: knowledge is a special kind of in- tained from data mining are adapted to the goals and dic- formation able to transform the source information to the tated by experts. Usually patterns are described by associ- information in demand. ation rules, decision trees, regression functions and neural In another words – it is a mapping: networks. KNOWLEDGE: information → information. Knowledge engineering constructs knowledge-based sys- tems used for reasoning, intelligent search and decision The definition of information can be taken from C. Shan- support. Knowledge, related to particular domain, is built non, the father of information theory: information is every- into a system and utilized for data transformation (e.g., from thing that decreases uncertainty. conditions to conclusions). It is assumed that the meaning of uncertainty is well known. 81 Wiesław Traczyk Knowledge as a means of information transformation is well useful in many applications. In some cases properties suited to different domains: of a class are passed on to its subclasses, simplifying representation. – knowledge of problem solving: problem → solution, • Relational dependencies between concepts represent – knowledge of decision support: connections linking together either concrete or ab- problem criteria decision, + → stract elements of the domain. This can help in – knowledge of inference: conditions → conclusion, searching for complex dependencies between given concepts. – knowledge of search: demands → goal, etc. Source and resulting information can be interpreted as input There are numerous languages describing some or all of and output data, and knowledge – as a program used for these dependencies. Short presentation of selected, repre- data processing. Such a program should be described by sentative examples will help in farther considerations. specialized languages of knowledge representation. Demands for representation language depends on the layer of processing [1, 2]. 2. Languages for knowledge • The domain layer concerns declarative knowledge representation about the domain of application. Such knowledge describes the objects of discourse in a particular do- A common problem during the development of ontologies main, facts that hold about these objects, and rela- and languages is a range of their applications. It is more tionships among them. This type of knowledge is costly to create representations that will be reusable across often represented by concepts, relations, hierarchies, multiple domains than it is to create a language that is suit- properties, rules, etc. A crucial property of this cat- able for just one application. Languages described below egory of knowledge is that it is represented as much attempt to be more or less universal. The original termi- as possible independently from how it will be used. nology from the source papers is preserved. The domain layer describes the formal model of a do- (ML)2. A formal language (ML)2 [2] has been developed main. for the representation of KADS [1] models of expertise. • The inference layer describes the roles of domain It uses an extended first order predicate calculus. Con- expressions and specifies how these expressions are cepts are represented by constants, i.e., nameable and dis- to be used in the inference steps. One can say that the tinguishable entities. All constants and variables have asso- inference layer is a meta-layer of the domain layer. ciated sorts (types, classes), organized in a subsort hierar- chy. Predicates and functions are also typed. The relation • The task layer enforces control over the inference between constants and sorts is the equivalent of the IS-A steps specified at the inference layer. Here it is de- relation. A knowledge base is divided into theories, consist- cided in which order the inferences should be ex- ing of import relations, signature (sorts, constants, models ecuted. The procedural knowledge is used on this of functions and predicates), variables and axioms (instan- layer and is concerned with sequences, actions, iter- tiations of predicates, production rules and functions). ations, etc. The main ideas can be illustrated on the following example: In this hierarchy of layers only the domain layer formally theory carFailure describes a particular domain of application and should be signature adapted to textual form of knowledge sources. That is why sorts number ( vehicle (bus car (station-car only languages of the first layer will be considered below. limousine))); the subsorts tree Knowledge has many forms that have to be represented constants by appropriate languages but the most important and fre- myCar, yourCar : car; quently used are three groups of demanded dependencies bus-412 : bus; between elements of representation. functions • Logical dependencies between statements with val- noOfWheels : vehicle → number; ues TRUE or FALSE determine principles for rule predicates based reasoning, the most popular method of infer- sameBrand : car × car; ence. The long history of logics and its established greater : number × number; position as a tool for formal analysis help to solve variables a,b,c : number; numerous problems. axioms • Hierarchical dependencies organize objects of a do- same Brand(myCar, yourCar); main into class taxonomy. The ability to represent the greater(a, b) ∧ greater(b, c) → greater(a, c); relationships between an object and its class or be- noOfWheels(yourCar) = 4; tween a class and its superclass has proven to be very endtheory 82 Structural representations of unstructured knowledge Telos. A hybrid language Telos [3] supports three different der uncertainty (using Dempster-Shafer theory) and XML representation formats: a logical, a graphical (semantic net- as a serial syntax definition language. work with relations) and a frame representation. A Telos F-Logic. In a deductive, object oriented language F-Logic knowledge base is a finite set of interrelated propositions [6, 7] objects are organized in classes and methods repre- presented as vectors: hoid,x,l,y,tti. Their meaning is as sent relationships between objects. Facts are collections of follows: objects with classes and appropriate methods, e.g., The object x has a relationships called l to the peter:student[father−>john:man]. object y. This relationship has identifier oid and is believed by the system for the time tt. Rules use similar forms, often with quantifiers and vari- ables: A standard way to describe objects together with