The only clear answer I can find in this case is that insofar as the behaviour is verbal, the design of a relatively robust shallow simulation can contribute to our understanding of some of the difficulties of interpreting unconstrained sentences. In particular, it seems likely that this sort of project could unearth previously unnoticed features of English grammar, and pinpoint some of the problems of coping with incomplete, ungrammatical, or misspelt input. How far our understanding of these issues has been advanced is not clear from Colby's paper. In particular, it is not clear whether anything new has been discovered about how to integrate the processing at different levels (syntax, semantics, pragmatics, inferences, and the like) in order to achieve robustness.

As far as understanding paranoia is concerned, I suppose the program can be said to embody some sort of transition net, showing which changes of state can be triggered by various types of input. However, there is no real analysis of the nature of those states, or the processes of change. That would require a far more complex model, showing how beliefs and motives of various sorts are represented, and how they can interact with each other and new information to produce new beliefs and motives and various kinds of disturbances of processing which we recognise as emotions. For instance, feeling afraid involves a complex cognitive state involving beliefs about what is likely to happen and desires that it should not happen, together with disturbances of other cognitive and physiological processes. Being angry involves an even more complex collection of beliefs and interacting motives. Only when we have a deep understanding of such computational states and processes as they occur in "normal" people can we hope to understand what is going on in pathological cases. I shall shortly submit to this journal a long paper surveying some of the processes involving motives, deliberation, decision, and emotions in intelligent animals and artefacts. Perhaps such work can eventually be integrated with Colby's.

COLBY'S RESPONSE (p 557-8)
Sloman, adopting a deep-shallow metaphor, wants a more complex model that includes "physiological processes." At what level should models of mental activity be sealed off, above and below, from other relevant knowledge? A model builder must decide on what to include as significant and what to exclude, not as nonexistent, but as negligible for his purposes. Sloman's slogan is "only when we have a deep understanding of . . . [the] 'normal' can we hope to understand . . . [the] pathological." Without unpacking the deep-shallow metaphor, we can note that there are dozens of examples in medicine in which we understand the pathological quite well and do not understand the normal at all. For instance, we understand the pathogenesis of many illnesses, but no one yet understands what it means to be healthy, unless it is defined only as the absence of illness.

In Sloman's ideal model, "feeling afraid involves a complex cognitive state involving beliefs about what is likely to happen and desires that it should not happen." If this is what is meant by "deep," this is exactly how PARRY operates. It becomes afraid when it believes that the interviewer may be linked to the Mafia and therefore will harm it; it desires not to be harmed, and it takes steps to avoid the anticipated harm. If these interpretation-action patterns become implemented in Sloman's planned model, then knowledge in AI does accumulate in spite of what snappish critics have often claimed.

When Sloman has his model of the normal up and running (it must be more than an armchair concoction of ingredients), he will be faced with the inescapable taxonomic question, What is your definition or criterion of "normal"? The category "normal" is too vague to qualify as a natural kind that can serve as an explanandum. This lack of initial taxonomic "consensibility" is one of the strongest criticisms leveled at the behavioral sciences. The advantage of "paranoia" as a natural kind is that it gains a high concordance of agreement. Notice that even among our commentators it does not gain 100% agreement. The fabric of the world is probabilistic, and there is room for disagreement arising from inherent variation in the phenomena, measurement errors, and alternative interpretation. Until there is a starting workable preliminary classification of "normal" kinds of mental functions, the explanatory models of normal processes will have a hard time gaining acceptance. They lack the first step of taxonomic agreement as to what it is that we are trying to explain.

Maintained by Aaron Sloman
School of Computer Science
The University of Birmingham

--