ARCHIVED Some Aspects of the Terminology of Artificial Intelligence (Part One)

 

Consult the Pavel in ...

Silvia Pavel*

1. A Multidisciplinary Field: Artificial Intelligence and the "Cognitive Hexagon"

   In a state-of-the-art report on research in cognitive science commissioned in 1978 by the Sloan Foundation in the U.S.A., the interrelationships among the constituent subfields of cognitive science were represented in a graph called the cognitive hexagon (See diagram.) In this graph, the unbroken lines represent strong disciplinary ties connecting philosophy, psychology, linguistics, artificial intelligence and neuroscience, while the broken lines suggest weaker ties between philosophy and neuroscience, as well as between philosophy, artificial intelligence and anthropology.

   

   Fig. 1 Connections among the Cognitive Sciences. The cognitive hexagon. (Gardner, 1985)

   The connections among the six cognitive sciences are represented as the six points of a ‘cognitive hexagon’ (clockwise: Philosophy, Linguistics, Anthropology, Neuroscience, Artificial Intelligence and Psychology). In it, strong connections are rendered by continuous lines (such as between Philosophy, Psychology and Linguistics or Psychology, Artificial Intelligence and Neuroscience), and weaker ones are rendered by discontinuous lines (such as between Philosophy and Neuroscience, or Artificial Intelligence and Anthropology).

   The discipline of cognitive science had been unofficially founded in the mid-1950’s in the course of an MIT symposium in information theory, and defined as "a contemporary, empirically based effort to answer long-standing epistemological questions, particularly those concerned with the nature of knowledge, its components, its sources, and its development. The term ‘cognitive science’ applies chiefly to efforts to explain human knowledge, but its use is sometimes extended to include all forms of knowledge - animate as well as inanimate, human as well as nonhuman" (Gardner 1985 : 6).

   More specifically, cognitive scientists claim that mental representations are necessary constructs in analyzing human cognitive activities, that the electronic computer is central to any understanding of the human mind, and that affective, historical or cultural factors should be deemphasized in studies of human cognition.

   The term artificial intelligence itself was coined by the mathematician John McCarthy, one of the organizers of the 1956 Summer Institute at Dartmouth College in Hanover, New Hampshire. Basing themselves on the assumption that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" (McCorduck 1979: 93), the members of the Institute considered the possibility of producing computer programs that permit intelligent "thinking" or behaviour.

   During the following three decades, however, the new field of artificial intelligence also became "renowned for its lack of consensus on fundamental issues" (Wilensky 1983: XII), including how the field itself should be defined. Among the main conflicts of opinion were those between the generalists, "who believe in overreaching programs or families of programs that can be applied to most problems" and experts, "who place their faith in programs that contain much detailed knowledge about a specific domain but are relatively restricted in their applicability" (Gardner 1985: 141).

   These conflicts originate from two different approaches that have arisen in Al research: the product-directed approach, which is "concerned mainly with the technology of getting computers to do a range of things that are quite fantastic in human terms", and the theory-directed approach, which is dedicated to "the representation of knowledge, learning, and human thought processes" (Schank 1984: 31).

   While a number of researchers see AI as a theoretical scientific discipline meant to replace epistemology, there are others who do not consider it a discipline at all, but merely a form of applied engineering. However, most AI specialists accept the definition provided by Feigenbaum and McCorduck (1984: 312):

   "A subfield of computer science concerned with the concepts and methods of symbolic inference by a computer and the symbolic representation of the knowledge to be used in making inferences. A computer can be made to behave in ways that humans recognize as intelligent behaviour in each other."

   At present, most AI laboratories do research in some or ail of the following fields: reasoning strategies, inferencing and knowledge representations, programming, computer architecture, design and analysis systems, natural language processing, machine translation, voice recognition, machine-learning, vision, sensing, and manipulation (Winston and Prendergast, eds., 1984: 2; Parent 1984: 25). These research areas allow for the development of computer programs that can read, speak or understand language; visual and tactile programs for sensate robots; and symbolic knowledge programs that simulate the behaviour of human experts or build other expert programs. Such "expert systems" allow for countless applications of machine intelligence in every conceivable field: business, engineering, manufacturing, farming, mining, education, health care, and household concerns, and include work in hostile or hazardous environments.

   Harmon (1985, in English) and Parent (1984, in French) provide a detailed analysis of the main differences between conventional programming and expert systems development (i.e., knowledge engineering), as well as between theoretical and applied artificial intelligence.

   Notwithstanding the differences in view points described above, AI today is "a truly integrated field of study with its theories, experiments and applications coming from computing, engineering, mathematics, logic, physics, psychology, human engineering, neurology and other fields" (Science Council of Canada SS-24-19E 1983: 10).

2. Aspects of AI terminology
   1. Conceptual Borrowings as Reflected by AI Keywords

      The association of the academic side of AI with the fields of philosophy, cognitive psychology, logic and linguistics, and of its technical side with computer programming, computer architecture and information retrieval is strongly confirmed by a whole series of AI concepts and keywords borrowed from these disciplines. All the terms in the following lists represent common concepts and issues in AI:

      From philosophy: act, action, actuation, analogy, belief, causality, concepts vs. objects, cognition vs. knowledge, completeness vs. incompleteness, human communication and information theory, consistency vs. inconsistency, credibility, fact / factual, generalization, heuristics, hierarchy, instantiation, proof, primal, primitives, specification, symbolics, truth, and universals.

      From logic and mathematics: abduction, algorithm, attribute, categorical vs. hypothetical reasoning, certainty, confidence, consolidation, classification, deduction, format logic, fuzzy sets, game theory, induction, inference, modus ponens, modus tollens, monotonic vs. nonmonotonic logics, probability, plausibility, problem solving, predicate calculus, propositional logic, quantification theory, symbolic logic, truth tables.

      From linguistics and psychology: case grammar, concept formation, creativity, deep vs. surface structures, discourse analysis, discrimination, empathy, immediate constituent analysis, intelligence, learning, mental representation, meaning, narrative structures, parsing, pattern recognition, reminding, rewriting rules, semantics and syntax, speech acts, thought models, tutorials, understanding, visual perception.

      From computer science: best-first vs. best few search, bottom-up vs. top-down search, computer-aided human translation vs. human-aided machine translation, database vs. knowledge base, dataflow architecture, default values, debugging, declarative vs. procedural languages, diagnostic, garbage collectors, graphs, if-then rules, lists and trees, machine language, man-machine interface, memory organization, massively parallel processing, program synthesis, software engineering vs. knowledge engineering, task-level programming.

      Some of these concepts are found in more than one member of the "cognitive hexagon" and may correspondingly take different expressions. Thus the concept designated in computer science by the term if-then rule (règle si .. alors; is designated by the term production rule (règle de production) in cognitive psychology and AI as well as by the term condition-action rule (règle conditionnelle) in AI.

      An interesting case in point is the tree-structure terminology used in AI. This may take the form of kinship terms such as the ones used in componential analysis: ancestor (ancêtre), father (père), grandfather (grand-père), child (fils), descendent (successeur), parent (précurseur), siblings, daughters (filles), twins, sisters (soeurs), etc. There is a parallel set of "botanical" terms which are used in linguistics and mathematics to refer to the same concepts. These are used in AI to describe hierarchically ordered links and problem-solving goals in a more graphic form: root (racine), node (noeud), leaf (feuille), branch (branche), tree (arbre), etc. However, AI trees grow upside down: their root is at the top node (tête) and their leaves are tipnodes (nœuds terminaux).

      Most of these borrowings from other fields do not acquire new meanings in the new field, a situation similar to the case of terms borrowed across such closely related fields as philosophy and logic, or psychology and linguistics. We may find a similar situation in French, where AI terms display few meaning differences from their original fields.

      In what follows, a number of AI terms, both English and French, are examined and some linguistic problems arising in the translation process are discussed.

   2. Comparative Terminology
      1. Expert systems, knowledge systems and other AI systems

         As originally used, the term expert system (système expert) refers to a computer program that could perform at or near the level of a human expert by using knowledge and inference rules to solve difficult problems. As currently used, the term applies to any computer program developed by means of AI techniques, even when it is so narrow as to never rival a human expert. Some AI people would rather use the term knowledge system or knowledge-based system (système à base de connaissances) to refer to AI programs that capture the expertise of skilled practitioners and derive their utility more from their user-friendliness than from their knowledge-capturing ability. The same AI people would tend to reserve expert system for systems having large knowledge bases that truly rival human experts.

         In French, a distinction is also made between système expert and système à base de connaissances, though less categorically. Terms such as système expert de grande taille and petit système expert are frequently used instead. Shorter expressions, such as supersystème expert and minisystème expert seem also acceptable.

         In both languages, the words program (programme) and system (système) are used interchangeably in software-related expressions. And these words are gradually being replaced by the suffix -er (-eur), as can be seen in the following attested examples: problem solving system / problem solver (système de résolution de problèmes / résolveur de problèmes); theorem proving system / theorem prover (programme de démonstration de théorèmes / démonstrateur de théorèmes); automatic programming system / automatic programmer (système de programmation automatique / programmeur automatique); inferencing program / inferencer (système inferential); natural language interpreting system / NL interpreter (interpréteur LN); speech recognition system / speech recognizer (système de reconnaissance de la parole / reconnaisseur de parole). In other expressions, e.g. vision system, the term "system" retains his more general meaning, i.e. hardware and software.

         Speech-recognition systems are only 25 years old, but one can already distinguish speaker-trained or single-speaker systems (appareils monolocuteurs) that recognize the grammar and pronunciation of a single command-giving person, from the multispeaker systems, speaker-independent or user-tuned systems (systèmes multilocuteurs), also called unrestricted or unconstrained speech recognizers (appareils multilocuteurs-multiréférences), able to be used by a few hundred speakers.

         Besides these voice-driven or vocal-input systems (appareils à commande vocale, systèmes commandés à la voix), there are also voice-output systems (appareils à sortie vocale), based on speech synthesis techniques.

         Another useful distinction is drawn between AI learning systems and AI teaching systems. Both in French and in English, the instruction process may be viewed from either the student’s or the teacher’s point of view. Thus, in computer terminology outside of AI, there are frequent references to computer-aided learning (or CAL) (apprentissage assisté par ordinateur) and its counterpart, computer-aided teaching (enseignement assisté par ordinateur). There would have been no reason for this useful distinction to become blurred if research in that field had not led to machines that learn from their own past experiences. It so happens that these latter are designated by the terms learning AI system, learning program (programme d’apprentissage) or learning machine (machine autodidacte). However, in what is called intelligent computer-aided instruction (or ICAI), the teacher being an AI program, the predominant point of view is that of the tutor. Consequently, the common terms for these programs are intelligent tutoring system and intelligent tutor (didacticiel intelligent, didacticiel IA). In this manner, "learning" has come to qualify the machine that learns by itself, and "tutoring" - the machine that teaches. One may wonder what term will be used to refer to tutoring systems capable of learning from their past experiences.**

      2. Knowledge representations

         The basic AI unit or conceptual entity in any knowledge representation scheme is the object (objet), defined as a collection of attributes, parameters or rules. Knowledge bases are often represented by object trees that graphically show how their different objects relate to each other.

         In some object-oriented representations, units of knowledge are called frames (cadres), and their attributes or values are stored in slots (tiroirs). Frames are sets of slots related to specific objects, and as such are similar to the property lists used in conventional programming.

         Object representations are also called contexts (contextes) in some medical systems (see MYCIN); schemes (schémas) in cognitive psychology (Bartlett 1932) and scripts, scenes, or subscripts (scripts, scènes, sous-scénarios) in computer understanding systems (Schank 1984).

         In AI terminology, object-oriented languages (langages orientés objets) are therefore not to be confused with the object languages (langages-objets) used in conventional programming and translated from a source language (Tremblay & Sorensen 1985: 4).

         Another distinction inherited from conventional programming is that between procedural and declarative representations as two complementary views of a computer program. If knowledge consists of facts and their relationships, then its representation, i.e.,the way in which a system stores it, may be either procedural (based on rules that tell the system what to do) or declarative (based on statements telling the system what to know); it may also be a mixture of the two. The French equivalents are respectivelyreprésentation procédurale, représentation déclarative, and représentation mixte.

      3. From machine languages to language machines

         First-generation computer languages were called machine languages (langages-machine) because they could be used directly by a given machine or computer (Rosenberg 1984: 299).

         Second, third and fourth generation languages have since been developed in classical computer science with varying degrees of flexibiiity and sophistication. More recently, fifth-generation computers have introduced the concept of language machines (machines-langages). In the high-level language-machine, the instruction set is designed for the high-level language itself. A translation process still exists but it is now simpler. More efficient coding may be accomplished during the same translation. We can thus see that the distinction between what is designated by "machine language" and what is designated by "language machine" goes beyond what is suggested by a simple inversion of the lexical items in these terms.

   3. Some Linguistic Aspects of Al Terminology
      1. Semantics

         As it happens in almost every other highly specialized field, oral and written communication in AI is characterized by the occurrence of commonly used keywords in the place of less frequently used but more "proper" terms in certain contexts: these substitutions may occur at the expense of discourse clarity and coherence.

         For instance, the term machine was used first as a synonym of "automaton" (e.g., in "finite-state machine"), but soon replaced computer in the sense of "hardware" (e.g., in "machine-aided translation"), and "software" (e.g., in "inference machine" and "medical advice machine").

         In the same text, expressions such as machine vision and computer vision may freely and frequently alternate with artificial vision, AI vision, automated vision and automatic vision. The same situation occurs in machine cognition, machine perception and machine reasoning, to name only a few. One can safely predict that the oncoming AI applications to robotics as underlined by Feigenbaum and McCorduck (1984) will further expand the meaning of "machine" to include intelligent robots as well (e.g., reading machines, machine doctors and military spy machines). However, for the time being, "machine" means either just hardware as in "16-bit machine" or hardware-software as in "learning machine."

         As we have seen previously, semantically contrasted concepts have lead to the emergence in AI of antonym pairs which may occur in other related fields: man vs. machine, natural vs. artificial, machine-aided vs. human-aided, data vs. fact, data-driven vs. goal-directed, dumb vs. smart, etc.

      2. Terms formation

         Lexical creativity in AI terminology seems to rely much more on word composition than on derivation. So far, the only truly productive suffix is -er, as used in "prover," "recognizer," "solver," and "inferencer." The more frequently used word composition model is based on the form of terms used elsewhere in computer technology. Thus, "bottom-up vs. top-down programming" inspired "bottom-up vs. top-down inferencing strategies"; "computer-based, computer-directed, and computer-driven techniques" have yielded "knowledge-based, example-directed, failure-driven and object-oriented" AI approaches. However, the most eloquent example is given by two terms describing memory allocation techniques: the "first-fit" and the "best-fit" (Ralston 1983: 477). This particular word-structure quite unexpectedly proliferated in AI terminology: metric-first, best-first vs. best-few, breadth-first vs. depth-first, etc. It is on this point that French displays a less uniform series of lexical equivalents (meilleure métrique, le meilleur d’abord, les plus prometteurs, largeur d’abord vs. profondeur d’abord). In other instances, French-speaking AI specialists have come up with a number of variously inspired expressions. Thus tree representation becomes arborisation, gapping grammars are grammaires discontinues, garbage collector is ramasse-miettes, shells are coquilles vides, and knowledge chunk is referred to either as fragment de savoir (Bonnet, 1984, p. 99) or granule de connaissances (Farreny, 1985: 7).

      3. Spelling

         AI terminology is created and adopted at such a pace that many spelling variants may coexist in both French and English before a shorter, single form can supersede the other. We may mention the half-hearted use of hyphenation in adverbial constructions (as in common sense / common-sense reasoning) and in other compound words (as in meta-rule / metarule, metametarule, and meta-meta-database). In French, one finds métabase de données, méta-connaissance, and métarègle or méta-règle.

3. Conclusion

   Despite its rapid evolution, AI remains a fairly recent discipline and the AI specialists themselves feel that it is too early to start work on terminology standardization (AFNOR 1985: 42).

   At this stage of our research, it would therefore seem inappropriate to submit concrete recommendations for AI terminology unification, even if the lack of usage consensus among AI specialists may sometimes be confusing for the translator or the layman. Terminological description seems at this point more useful than prescriptive terminologies.

   There is no doubt in our mind that the terminological research that makes possible the publishing of AI vocabularies and the dissemination of up-to-date AI terminology in as many languages as possible can only lead to a greater uniformity and disambiguation of terminological usage in this field.

   A better understanding of the terminological problems helps fulfill the communication needs of AI specialists within a given language. It may also benefit the translator who is called upon to transfer knowledge concerning artificial intelligence across language barriers. And ultimately it helps the layman to understand in what ways his daily life will be changed by the AI revolution.

Bibliography

AFNOR, 1985, "L’intelligence artificielle" in Enjeux n° 61, Paris: Association française de normalisation.

Bartlett, F., 1932, Remembering: a Study in Experimental and Social Psychology, London: Cambridge University Press.

Bobrow. D., Winograd, 1977, "KRL another perspective", in Cognitive Science 3, pp. 29­42.

Bonnet, A., 1984, L’intelligence artificielle : promesses et réalités, Paris : InterEditions.

Edmunds. R.A., 1985, Standard Glossary of Computer Terminology, Englewood Cliffs: N.Y: Prentice-Hall.

Farreny, H., Les systèmes experts : principes et exemples, Toulouse : Cépadues Ed.

Feigenbaum, E.A., McCorduck P., 1984, The Fifth Generation: Artificial Intelligence and Japan’s Computer Challenge to the World, New York: New American Library.

Gardner, H., 1985, The Mind’s New Science: a History of the Cognitive Revolution, New York: Basic Books. Inc.

Harmon P., King. D., 1985, Expert systems: Artificial Intelligence in Business, New York: J. Wiley and Sons Inc.

McCorduck. P., 1979, Machines Who Think, San Francisco: W.H. Freemer.

Parent, R., 1984, Point de vue québécois sur l’intelligence artificielle, Québec : Ministère des communications.

Ralston, A., (ed), 1983, Encyclopedia of Computer Science and Engineering, New York: Van Nostrand Reinhold Co.

Rosenberg., J.M., 1984, Dictionary of Computers, Data Processing and Telecommunications New York: J. Wiley and Sons.

Schank, R., 1985, The Cognitive Computer, Reading, Mass.: Addison-Wesley Publ. Co.

Science Council of Canada, 1983, Atelier sur l’intelligence artificielle organisé par le Conseil des Sciences du Canada, Ottawa.

Tremblay, J.-P., Sorensen, G., The Theory and Practice of Compilers, New York: McGraw Hill

Wilensky, R., 1983, Planning and Understanding: A Computational Approach to Human Reasoning, Reading, Mass.: Addison-Wesley Publ. Co.

Winston, P.H., P. Prendergast (eds), 1983, The AI Business: The Commercial Uses of Artificial Intelligence, Cambridge, Mass.: MIT Press.

(In the next issue: Part Two — English-French Glossary of Artificial Intelligence.)

* I should like to thank my colleague Patrick F. McNamer and Professors T. Oren and D. Skuce (Dept. of Computer Science, University of Ottawa) for their comments and suggestions on this paper.

** Prof. T. Oren has already suggested the terms experiential learning tutor (didacticiel autodidacte).