IRE Laboratory tests: automatic systems chapter Robert N. Oddy Butterworth & Company Karen Sparck Jones All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, including photocopying and recording, without the written permission of the copyright holder, application for which should be addressed to the Publishers. Such written permission must also be obtained before any part of this publication is stored in a retrieval system of any nature. 172 Laboratory tests: automatic Systems I from conventional indexed documents and a thesaurus. The form that its image of the man takes is a subgraph (from its own `knowledge') together with a few notes about his reactions to information displayed during the dialogue. Now, we turn to the program. The processes in the program were refined in a hierarchical fashion as the meanings of the program's goals were gradually elaborated. The description of one cycle in the dialogue (i.e. at a very high-level in the model) is as follows: (1) read a message from the man; (2) use it to influence the state of Thomas' meta- image; (3) use the meta-image to respond to the man (For readers who are programmers, I should say that each of the three steps is coded as a procedure call.) I shall not elaborate the first step here, except to say that it includes an interpretation of the text of the man's message which makes reference to Thomas' images. Thus the communication can be regarded as having a cognitive basis46. Let us proceed just one more level down in the (informal) program definition with the description of step 2: (1) update Thomas' self-assessment according to the man's reaction to the last display; (2) `prune' Thomas' meta-image in the region of nodes which the man does not appear to like; (3) `enrich' the meta-image in the region of nodes in which the man shows interest; (4) incorporate new material in the meta-image if the man has explicitly mentioned new words; (5) make sure the meta-image is not fragmented I am not attempting to make this specific program fully comprehensible to the reader (the details can be found elsewhere33), but to give him or her a feel for the way a model may be elaborated through programming. For this purpose, I hope that I have now carried the illustration far enough. The process of elaboration of the model is by no means automatic, and the result is not unique. Some decisions must be made by the programmer in an ad hoc way: he uses intuition, introspection and, when he can, the dictates of theories which, he feels, are applicable. This is a well-known method in Artificial Intelligence work, and has been defended ably by Lindsay56 and Schank64. A computer model rarely achieves its final form at the first attempt. Indeed, if it does, one can conclude that it has been abandoned, due to failure or lack of interest. The modeller will try to correct defects in its performance by making what he hopes are appropriate modifications to the program, and hence the model. This is a type of theory which can be very readily changed1 and re-tested65' 66 The advantage of this facility becomes clear when we consider that, even in the case of a fairly simple procedural model, a8 incorporated into Thomas, it is extremely difficult to account for its behaviour (i.e. relate output to input) mathematically. The computer modellin8 U- I 1 I