Knowledge processing and commonsense1

Knowledge processing and commonsense1

Knowledge-Based Systems 10 (1997) 147–151 Knowledge processing and commonsense1 R. Narasimhan C/o CMC Ltd, Bangalore 560 001, India Received 15 May 1...

89KB Sizes 7 Downloads 100 Views

Knowledge-Based Systems 10 (1997) 147–151

Knowledge processing and commonsense1 R. Narasimhan C/o CMC Ltd, Bangalore 560 001, India Received 15 May 1997; accepted 29 May 1997

Abstract Symbolic expert systems have been developed during the last three decades to model knowledge-based human intelligent behaviour. A general criticism of such expert systems is that they lack commonsense. But there is little consensus among AI workers on what constitutes commonsense. In this paper a characterization of commonsense is attempted. The limitations and open problems relating to current approaches to expert systems design are discussed. In addition, open problems that need to be studied to adequately model commonsense behaviour are discussed. Our basic arguments hinge on the distinctions between tacit knowledge and propositionizable knowledge. The thesis is that commonsense behaviour is essentially underpinned by tacit knowledge. q 1997 Elsevier Science B.V. Keywords: Symbolic expert systems; Commonsense intelligence; Tacit knowledge; Skill-based expertise

1. Rule-based knowledge processing and commonsense The core received view of artificial intelligence (AI) has recently been succinctly stated by Kirsh [1] as follows: One’s commitment is that knowledge and conceptualization lie at the heart of AI; that a major goal of the field is to discover the basic knowledge units of cognition... The basic idea that knowledge and conceptualization lie at the heart of AI stems from the seductive view that cognition is inference. Intelligence skills...are composed of two parts: a declarative knowledge base and an inference engine. It is further believed that while the declarative knowledge base has to be necessarily domain-specific, it is possible to construct the inference engine to be universal (i.e., domain independent). In conformity with the above belief, symbolic expert systems methodology has been developed and brought to a high level of competence during the last two decades. Symbolic expert systems are engineered systems that try to exploit the strength of AI techniques through the use of the following strategy: 1. choose a well-delimited application area (i.e., task domain) of practical value; 2. propositionize the knowledge of this area;

3. organize this knowledge in a usable form; 4. work out strategies to use the knowledge to go from typical problem statements to plausible solutions. Many critics—both from within the AI community and from outside—point out that symbolic expert systems as presently constituted lack commonsense. However, there is no general agreement on what constitutes commonsense. Sometimes it has been interpreted very narrowly as the ability to reason from insufficient knowledge of the problem situation [2]. Further discussion has been concentrated on determining whether suitable augmentations to logical reasoning could model this activity. McCarthy [3] identifies a much longer list of issues relating to commonsense. Some of the more significant ones—apart from issues such as reasoning from insufficient knowledge, representation and use of meta-knowledge, etc.—are: • • • • •

dealing with situations that change in time; dealing with simultaneous occurrences of several events with mutual interactions; dealing with physical events of the everyday world (i.e., implementing aspects of the naive physical knowledge about the world); dealing with the agentive aspects of oneself and of others (i.e., intentions, beliefs, wants, abilities, etc.); and so on.

1

Revised version of the article originally published in Knowledge-based Computer Systems: Research and Applications. (Eds K.S.R. Anjaneyulu, M. Sasikumar and S. Ramani) Narosa Publishing House, New Delhi, 1997.

0950-7051/97/$17.00 q 1997 Elsevier Science B.V. All rights reserved PII S 09 50 - 70 5 1( 9 7) 0 00 2 5- 7

Analogously there are issues relating to commonsense reasoning. The basic issue again is whether commonsense

148

R. Narasimhan/Knowledge-Based Systems 10 (1997) 147–151

knowledge and reasoning, interpreted in this manner, can be accommodated in some suitable extension to traditional first-order logic. Davis [4] gives an excellent account of the extensive work that has been, and is being done, to extend formal logic and reasoning suitably to cope with the representation and use of commonsense knowledge. However, to address the issues involved here more systematically, it is useful to group knowledge underpinning human behaviour in two broad categories as shown in Table 1. In terms of the two varieties of knowledge illustrated in Table 1, to say that symbolic expert systems lack commonsense is to say the following: •

• • •

Tacit knowledge underpins our behaviour in the perceptual-motor domain, and plausibly also much of our communication competence in natural language (see Ref. [5] for further elaborations). All this knowledge is taken for granted when human beings plan and solve problems in given problem domains. This tacit knowledge base is normally missing in mainstream AI systems and if any part of it is needed for planning and problemsolving, it has to be fully articulated and explicitly represented in the (declarative) knowledge base that underpins the AI system’s performance.

Two serious problems arise at this stage. First, we do not know what is the nature of this tacit knowledge and, therefore, we are unable to articulate it successfully. Second, at the representational level, we are unable to unify this tacit knowledge effectively with the problem-domain knowledge (usually expressed in predicate logic formalism and/or as condition–action rules). However, CYC [6] is an ambitious long-term research project attempting to accomplish precisely this. It remains to be seen how natural and successful the outcome will be for general purpose use in AI. Referring to Table 1 again, the division of knowledge into two categories has significance and value in accounting for

the pragmatics of behaviour. Literacy (i.e., writing, reading, the use of symbols, notations, etc.), in general, is a prerequisite to the acquisition of knowledge that is propositionizable. This is not the case for the acquisition of tacit knowledge. These two kinds of knowledge are, however, interdependent in the sense that one’s perceptual-motor competence may be modified by one’s professional knowledge. An expert may, and quite often does, see the world differently from the way an ordinary person sees it. On the basis of the above differentiation, commonsense—in so far as it is generally available to all human beings—must underpin an individual’s perceptual-motor competence and his/her ability to acquire and use skills. Articulation of knowledge and reasoning in the commonsense mode is accomplished through the use of natural language (in most cases, through the use of one’s own mother tongue). As indicated in Table 1, increase in one’s competence to function successfully in the commonsense mode must be accounted for in terms of the following aspects: 1. 2. 3. 4.

developmental factors (in the early stages); exposure to examples; rehearsal and practice; apprenticeship (in the case of complex skill acquisition).

Notice that all the above features apply not only to the acquisition of sensori-motor competence, but also to acquisition of competence in informal natural language use (see Ref. [5] for further elaboration of this viewpoint). For computational modelling, the open problems concerning commonsense behaviour, then, are accounting for the competence of ordinary (i.e., non-professional) human beings in coping with their day-to-day interactions with the physical world and with other human beings in the perceptual-motor and natural language modalities. These, clearly, are the fundamental issues that need to be addressed if we are to be able to deploy information processing systems to function successfully as free-agents navigating, exploring, manipulating objects, and interacting with other similar agents or with human beings.

Table 1 Two kinds of knowledge Knowledge that is tacit

Knowledge that is propositionizable

Underpins what behaviour?

Perceptual-motor competence Skill acquisition and use Naive natural language behaviour

Puzzles, games defined through explicit rules Problem-solving in an articulated task domain

How acquired?

Through informal means: exposure to examples apprenticeship rehearsal and practice

Through formal means: systematic formal tuition learning based on theories text books

What is it called?

Commonsense knowledge Craft knowledge

Professional or expert knowledge

Who has it?

Everybody (artisans and craftsmen when skill-based)

Professionals (experts)

R. Narasimhan/Knowledge-Based Systems 10 (1997) 147–151

What is the nature of commonsense knowledge (i.e., what is the nature of its internal representation) and what is the mode of commonsense reasoning? In other words, how are knowledge and control deployed to support behaviour in the commonsense mode? One way of approaching these issues is to clarify our notion of what constitutes intelligent behaviour in the tacit or commonsense mode.

2. Commonsense intelligence: Vision and language behaviour What is ‘‘intelligence’’? Here is one definition of it by McCarthy and Hayes [7]: We shall say an entity is intelligent if it has an adequate model of the world (including the intellectual world of mathematics), understanding of its own goals and other mental processes; if it is clever enough to answer a wide variety of questions on the basis of this model; if it can get additional information from the external world when required and can perform such tasks in the external world as its goals demand and its physical activities permit. If this is meant to be a minimal definition of intelligence, most of us in the world are unlikely to be classified as intelligent! Surely, as mentioned earlier, our primary concern must be with the ordinary activities of ordinary people and not with the deployment of professional skills by experts. Everything that an ordinary person does in his/her normal course of living in this world, in so far as it succeeds at all, must be underpinned by intelligence. Behaviour mediated by vision and natural language play key roles here. Understanding commonsense must start with understanding the nature of the mediating roles involved in these two modalities and how they come about. What are relevant questions to ask to this end? It is significant to note that behaviour precisely in these two modalities—vision and natural language—have been extremely difficult to model computationally. How would one account for this? I would identify at least two major reasons: 1. our attempts to base modelling in these two modalities on propositionized knowledge; in other words, opting for formal rule-based approaches; 2. our attempts to process information in these two modalities in a wholly de-contextualized fashion. Consider visually-mediated behaviour first. The received view stemming from the forcefully argued theses by Marr [8] is that the primary task of perception is generating descriptions. According to him: ‘‘...perception is construction of description... That is the core of the thing and a really important point to come to terms with’’. The crucial question, even granting this viewpoint is: ‘‘What kind of a description?’’ Surely, this cannot be answered unless one

149

takes into account the end-use of the description. Marr’s whole work is preoccupied with a totally de-contextualized task: ‘‘How can the visual system generate a complete description of a visual scene in terms of its constituent objects and their layout?’’ The implicit assumption is that once such a (task-independent) complete description has been generated, the information compiled could be put to a variety of task-dependent uses as needed. Many vision specialists, however, are beginning to move away from this wholly de-contextualized approach. See, for instance, the relevant position statements included in Ref. [9]. Second, visually-mediated behaviour spans a very wide spectrum. We can list some of these as follows: 1. recognizing, identifying: resulting in naming, categorizing, etc.; 2. discriminating: resulting in perception of similarity/difference; 3. describing, interpreting: resulting in verbal statements (formal/informal) or in non-verbal representations (pictorial/discrete symbolic); 4. exploring, searching; 5. navigating. There would seem to be little reason to assume that vision mediates in all these cases in a standard way. In contrast to Marr, Ramachandran [10] has recently argued that vision is really opportunistic, rather than rulebased and highly systematic. According to him: It may not be too far-fetched to suggest that the visual system uses a bewildering array of special purpose tailor-made tricks and rules of thumb to solve its problems. If this pessimistic view of perception is correct, then the task of vision researchers ought to be to uncover these rules rather than to attribute to the system a degree of sophistication that it simply does not possess. Seeking over-arching principles may be an exercise in futility. Crick [11] has endorsed this viewpoint by noting that This approach is at least compatible with what we know of the organization of the cortex in monkeys, and with Franc¸ois Jacob’s idea that Evolution is a tinkerer. We find an analogous state of affairs when we come to studying natural language behaviour. In almost all current computational approaches dealing with natural language, the strategy employed is to delink the language processing stage from the stage where the processed output is put to some well-defined use (e.g., to carry out some action). Language processing is thus dealt with as a wholly decontextualized (i.e., unsituated) autonomous activity. On the other hand, our focus must always be on the endbehaviour and how this could be mediated through the language input. Language processing should, clearly, be determined by the end-use to be addressed. The manner of

150

R. Narasimhan/Knowledge-Based Systems 10 (1997) 147–151

processing the input, as well as the extent of processing, should be determined by the end-use of the processed output.

3. Knowledge and control in commonsense behaviour So, what can we say about knowledge and control in commonsense behaviour? Some indications to answer this question may be found in the fact that in human beings and other animals, sensori-motor behaviour is underpinned by dedicated mechanisms—in the visual, auditory, tactile and manipulatory modalities. In all these modalities hard-wired programs carry out low-level (i.e., initial stage) processing of the sensori-motor information. It has been persuasively argued that low-level processing is data-driven and not knowledge-based [8]. We can perhaps argue that knowledge is tacit at this level because it is implicit in the mechanisms that support performance. Knowledge and control are not separated in perceptual-motor behaviour. Knowledge is implicit in the control structure. The mechanism is itself ultimately the representation of the expertise in these modalities. Looked at in this perspective, computational issues like the ‘‘frame problem’’ [12] for a navigating, manipulating robot assume quite different formulations from the ones that are usually presented in the AI literature. Motor actions are controlled by—among other factors—perceptual inputs. And, as the world changes due to the actions performed, the perceptual inputs change accordingly and the control for subsequent actions is suitably modified. So, while action is being performed, changes in the world need not be reflected in a fully articulated knowledge base which is assumed to control and drive action. The resulting changes in the perceptual inputs should be able to provide the needed monitoring and control information for performing further action. These observations are similar in spirit to Brooks’ theses arguing for ‘‘intelligence without representation’’ [13], ‘‘intelligence without reason’’ [14], and so on. To quote him: When we examine very simple level intelligence, we find that explicit representations and models of the world simply get in the way. It turns out to be better to use the world as its own model... It may be the case that our introspective descriptions of our internal representations are completely different from what we really use...[13] Brooks has been able to demonstrate successfully the deployment of reasonably complex behaviour by insectlike robotic creatures in their interaction with the real world. The subsumption architecture used in the design of these creatures is made up of several layers, each layer consisting of a fixed-topology network of simple finite state machines (FSMs). There is no central locus of control. The FSMs are data-driven by the messages they receive from the real world, as well as from within and between

layers. The real unanswered question is how complex can the behaviour of such creatures be without the use of a central representation. More specifically, it is unclear how a subsumption architecture of this sort can come to grips with the language modality of behaviour which is a prerequisite to planning, ratiocination, instruction-based learning, and so on. Connectionist architecture has been proposed and extensively studied as another alternative to mainstream symbolic AI to come to grips with tacit knowledge and learning without representation. Despite the large variety of connectionist systems that have been built (see, for example, Refs. [15,16]), there is as yet no clear indication how higher-level intelligent activities could be coherently modelled in terms of connectionism. For instance, how does one link high-level vision (e.g., recognition of a scene, or even a complex object) to the low-level information processing layers? How does one relate articulation in the language modality to tacit abstractions in the perceptualmotor modalities? These are crucial issues that need to be addressed before acceptable computational models of intelligent behaviour can be worked out. Referring to Table 1, at the knowledge level what is needed to distinguish commonsense behaviour from expert (or professional) behaviour is to understand better the difference between skill-based knowledge and theorybased knowledge. It must be emphasized that this distinction is not exactly equivalent to that between procedural knowledge and declarative knowledge. Deployment of skill involves both unarticulated know-how and articulation of aspects of this know-how and related situational details in ordinary natural language. Analogously, as we shall presently see, deployment of theory-based knowledge by professional experts is not equivalent to operating strictly (or exclusively) within a deductive formalism. Skill-based expertise is more akin to commonsense behaviour than to explicit, theory-based, symbol-manipulation behaviour. Skill-based expertise should, therefore, be a valuable domain of study for a computational-level understanding of commonsense behaviour. Finally, what about commonsense reasoning? Since according to our definition the language of commonsense is natural language, the mode of reasoning must be the mode that people use in their normal interactions with the world (including other people). There are persuasive arguments to show that the actual reasoning process of commonsense is not the inferential process as the logicians define it. In other words, it is not deduction in a well-defined logic [17]. Computationally describing the nature of this process is still very much an open problem. It is worth noting that commonsense underpins creativity and generation of new knowledge. This is sufficient indication that reasoning strictly within standard logical formalisms would not help. Dreyfus [18] argues that expertise is based more on recognition than on laborious inferential reasoning. This is certainly true of skill-based expertise and may equally well

R. Narasimhan/Knowledge-Based Systems 10 (1997) 147–151

be true of the intuitive grasp of problem situations that experienced professionals exhibit. In other words, as experience builds up, situational aspects are recognized at sufficiently high global levels and not laboriously built out of condition–action primitives from bottom up. Such globally recognized situational aspects directly suggest appropriate (i.e., plausible) strategies, plans or actions. The ‘‘expertise’’ of an expert would, then, seem to consist in being able, at the perceptual level, to narrow the problem situation to its salient aspects and, based on this and prior experience, to delimit the solution space (i.e., decision– action space) to a small, potentially profitable one. It is well established that experienced chess players function in this mode [19].

4. Concluding remarks We set out by discussing symbolic expert systems methodology that has been developed within AI during the last two or three decades to model knowledge-based human intelligent behaviour. We saw that in the design of these expert systems, typically, knowledge is explicitly propositionized and reasoning for purposes of monitoring and control is handled through making inferences in a welldefined logic. A general criticism of such expert systems is that they lack commonsense. We then tried to characterize the nature of commonsense and argued that the knowledge that underpins commonsense is tacit and not propositionized. And also reasoning in commonsense is based on the use of natural language and does not, for the most part, conform to any strict formalized logical inference. Computational modelling of commonsense behaviour cannot, therefore, be handled through extensions to current symbolic expert systems methodologies. Connectionism and subsumption architectures are being promoted as alternative frameworks for modelling behaviour based on tacit knowledge. In these models knowledge and control are intertwined and there is no attempt to base behaviour on a propositionized knowledge base. But we saw that such modelling attempts are at a very rudimentary stage and several fundamental open problems remain to be formulated and solved before these models can be deployed to handle significant aspects of commonsense behaviour. Mainstream symbolic-level AI has so far been exclusively preoccupied with attempts to simulate human performance based on propositionizable knowledge. However, we would be approaching behaviour modelling from the wrong end if we started out by assuming that ‘‘propositionizing’’ is a central principle of behaviour. Scientists and logicians are intensely preoccupied with propositionizing and propositionizable knowledge and tend to forget that no other animal propositionizes. But all animals, one would suppose, are able to build up a tacit knowledge of their worlds. Much of everyday, informal human behaviour, we have argued in this paper, is also based on tacit knowledge.

151

Phylogenetic continuity (i.e., continuity at the level of evolution) of knowledge-based behaviour, hence, must be sought in the domains of tacit knowledge. Human commonsense behaviour qualitatively differs from the tacit knowledge-based behaviour of other animals through the availability and use of the (natural) language modality exclusively to humans. Understanding, at the computational-level, these similarities and differences between human behaviour and the behaviour of other animals is really the most significant challenge to behaviour modelling and, by extension, to AI. References [1] D. Kirsh, Foundations of AI: The big issues, Artificial Intelligence 47 (1991) 3–30. [2] R.E. Moore, The role of logic in knowledge representation and commonsense reasoning, Proceedings of AAAI-82, 1982, pp. 428– 433. [3] J. McCarthy, Some expert systems need commonsense, in: V. Lifschitz (Ed.), Formalizing Commonsense: Papers by John McCarthy, Ablex, Norwood, NJ, 1990, pp. 189–197. [4] E. Davis, Representations of Commonsense Knowledge, Morgan Kaufmann, San Mateo, CA, 1990. [5] R. Narasimhan, Language Behaviour: Acquistion and Evolutionary History, National Centre for Software Technology (NCST), Bombay, 1996. [6] D.B. Lenat, R.V. Guha, Building Large Knowledge-based Systems: Representation and Inference in the CYC Project, Addison-Wesley, Reading, MA, 1989. [7] J. McCarthy, P. Hayes, Some philosophical problems from the standpoint of artificial intelligence, in: B. Meltzer, D. Michie (Eds.), Machine Intelligence 4, Edinburgh University Press, Edinburgh, UK, 1969, pp. 463–502. [8] D. Marr, Vision, Freeman, San Francisco, CA, 1982. [9] AI Symposium: Position Statements, ACM Computing Surveys, September 1995. [10] V.S. Ramachandran, Interactions between motion, depth, color and form: The utilitarian theory of perception, in: C. Blackmore (Ed.), Vision: Coding and Efficiency, Cambridge University Press, Cambridge, UK, 1990. [11] F. Crick, What Mad Pursuit, Penguin Books, London, 1990, p. 156. [12] P. Hayes, The frame problem and related problems in AI, in: A. Elithorn, D. Jones (Eds.), Artificial and Human Thinking, Elsevier, Amsterdam, 1973, pp. 45–59. [13] R.A. Brooks, Intelligence without representation, Artificial Intelligence 47 (1991) 139–159. [14] R.A. Brooks, Intelligence without reason, Proceedings of IJCAI, 1991, pp. 569–595. [15] D.E. Rumelhart, J.L. McClelland (Eds.), Parallel Distributed Processing 1: Foundations, MIT Press/Bradford Books, Cambridge, MA, 1986. [16] J.L. McClelland, D.E. Rumelhart (Eds.), Parallel Distributed Processing 2: Psychological and Biological Models, MIT Press/Bradford Books, Cambridge, MA, 1986. [17] P.N. Johnson-Laird, Reasoning without logic, in: T. Myers, K. Brown, B. McGonigle (Eds.), Reasoning and Discourse Processes, New York, Academic Press, 1986, pp. 13–50. [18] S.E. Dreyfus, The nature of expertise, Panel Discussion, Proceedings of IJCAI, vol. 2, Morgan Kaufmann, San Mateo, CA, 1985, pp. 1306– 1309. [19] W.G. Chase, S.A. Simon, The mind’s eye in chess, in: W.G. Chase (Ed.), Visual Information Processing, New York, Academic Press, 1973, pp. 251–281.