Semantic networks

Fritz Lehmann
1992 Computers and Mathematics with Applications  
A semantic network is a graph of the structure of meaning. This article introduces senmntic network systems and their importance in Artificial Intelligence, followed by I. the early background; II. a summary of the basic ideas and issues including link types, frame systems, case relations, llnk valence, abstraction, inheritance hierarchies and logic extensions; and III. a survey of 'world-structuring' systems including ontologies, causal link models, continuous models, relevance, formal
more » ... ries, semantic primitives and intersecting inference hierarddes. Speed and practical impleme~atation are briefly discussed. The conclusion argues for a synthesis of relational graph theory, graph-grmmnar theory and order theory based on semantic primitives and multiple intersecting inference hierarchies. • .. when controversies arise, it will not be a work of learned disputation between two philosophers, but between two computists. It will be enough for them to take pen in hand, sit at the abacus, and say to each other, as friends: 'Let us cMculatet. ' --Leibniz All thought is diagrammatic. --Charles S. Peirce 2 F. LEHMANN units, and directed links (drawn as arrows between the nodes) representing the relations between the units. The network in Figure 1 has a relational graph describing two individuals (Toby and the unnamed tigress) with their asserted qualities and relations, on top of which is an abstraction hierarchy of more general concepts and relations. Prom this combined structure it is possible to deduce things about the composite concept as a whole and its relations to other concepts. An abstract (graph-theoretic) network can be diagrammed, defined mathematically, programmed in a computer, or hard-wired electronically. It becomes semantic when you assign a meaning to each node and link. Unlike specialized networks and diagrams, semantic networks aim to represent any kind of knowledge which can be described in natural language. A semantic network system includes not only the explicitly stored net structure but also methods for automatically deriving from that a much larger structure or body of implied knowledge. For example, the assertion in Figure 1 that Toby is hungry implies that he is a conscious animal, and everything true of conscious animals is automatically true of Toby. Almost all systems have structured concept-hierarchies or taxonomies used for this kind of derivation (described in Section 6), and these hierarchies themselves are also 'semantic networks.' In the 1970's, semantic network research emphasizing this 'structure of knowledge' approach became predominant in AI, later contending with Rule-based Expert Systems for center stage. Since then it has waxed and waned periodically and many of its ideas reappear in other guises such as Object-Oriented Systems. Many Expert Systems now include Object-Oriented extensions which allow easy implementation of semantic networks (see Section 15). The latest vogue in AI, neural nets, often seems to be an opposing, anti-analytic approach using no identifiable symbols or concepts in the computer at all, but there are systems (Section 14) which have enough internal compositional structure to be used 'semantically.' Semantic networks are seldom claimed to exist physically in the brain; rather, they are viewed as idealized reasoning structures with practical 1 Throughout this volume, "graph" means an interconnected vertex-and-arc (dot-and-line) structure M studied in Graph Theory rather than a graph plotted in Cartesia~ X-Y coordinates. Semantic networks 3 computer application. A common goal of this kind of AI is to impart obvious 'common sense' to computers. 2 The specialized semantic network inference methods discussed below are often combined with other standard AI reasoning methods such as rule-based search, automatic theorem-proving, constraint satisfaction, machine learning algorithms, and others. Semantic networks are used in almost every application area of AI, including natural language understanding, deductive databases, library document retrieval, business planning, medical diagnosis, legal case analysis, analogical reasoning, expert systems, robot control, intelligent Computer Aided Design, visual pattern recognition, simulated aircraft control, and many more. Overview There are now eight major research families of semantic network systems plus countless independent projects around the world. The eight are: CONCEPTUAL DEPENDENCY • CONCEPTUAL GRAPHS • ECO • KL-ONE PATH-BASED INHERITANCE • PREFERENCE SEMANTICS • PSN • SNEPS In some of these there are all sorts of variants and it is quite confusing to the newcomer. The eight invited survey articles following this one introduce the basic ideas and give some guidance to the research directions within each family. These are followed by 25 articles on diverse subjects in the field, a The surveys and articles cover most contemporary work. In this article, following this introduction, Part I describes the historic origins of semantic networks; Part II covers the basics such as frame systems, relational graphs, deep case relations, link valence, inheritance hierarchies, IS-A links, relational inheritance and logic extensions; and Part III discusses 'world-structures': ontologies, continuous models, relevance, dictionaries, semantic primitives and intersecting inference hierarchies. I then briefly treat some speed and implementation issues. (If you're very familiar with semantic networks you can skip Part II.) My emphasis will be on knowledge structures rather than connections with natural language or particular notations, and I'll mention how the other articles relate to the subject. 1.P. Flat or Deepf Most commercial AI systems are Rule-based Expert Systems with large sets of IF-THEN rules supplied by an expert in some field to embody his or her expertise. There is near-universal disappointment at the 'flatness,' or lack of structured knowledge, in Expert Systems based on rules whose symbols may represent anything at all. It is hard just to keep track of numerous arbitrarily interrelated rules, and, during automatic inference, the computational burden of exhanstive search through a large space of unconstrained rules is often overwhelming. Serious users are soon frustrated by the indiscriminate stupidity of pure rule-based systems. A knowledgeable person will spot obvious errors in a description whereas a dolt will just accept it. A typical Expert System often seems clever due to the suggestive names of internally meaningless symbols, but the system cannot distinguish between telling it "IF a gas pipe bursts THEN gas leaks" and telling it "IF a gas pipe bursts THEN Michael Jackson is a parallelogram." It just responds "OK." You quickly get a yearning for some real understanding, some 'conceptual structure' forming a deeper model of the subject area. The most promising solution is thought to be highly structured Knowledge Bases built using Knowledge Representation Systems (semantic networks and symbolic logic are two possible representation systems--in the conclusion I'll discuss the difference). In a practical application you need to build an ontology of the concepts and principles of the particular subject area in question (Section 8); some thorny problem areas even require a general metaphysics. Some people like practical things and dislike airy metaphysical discussions. Other people, the philosophers among us, have the opposite taste. This difference in temperament is now obsolete: to get reliable practical results in AI you have to be a kind of philosopher, and you have to make 2For this reason AI research examples may often seem trivial to outsiders--most of what is perfectly obvious common sense to a three-year-old now eludes the most sophisticated computers. 3One subject which hardly appears at all in this volume is automatic macldne leaning; the articles deal mainly with representation and automatic use of knowledge irrespective of how it is obtained. F. LEHMANN every effort to think deeply and get the philosophy right. Semantic networks and kindred methods in AI are mechanized philosophy, and that is what researchers around the world are now doing. On the one hand the hard-headed businessman or military planner using AI who s~ys "let's get practical and skip the philosophical stuff" is headed for wrong turns, mistakes and rniAfortune due to faulty analysis; on the other hand the philosopher who cooks up vast and complicated AI theory without testing it on practical examples is likely to drift far from soundness and relevance.
doi:10.1016/0898-1221(92)90135-5 fatcat:4nli67lrsjdzdiu56i5ixwhm2a