ELIZA - A Computer Program for the Study of Natural Language Communication Between Man and Machine

What the paper sets out to describe

Joseph Weizenbaum's paper appeared in Communications of the ACM, volume 9, issue 1, January 1966, pages 36–45, with DOI 10.1145/365153.365168.^[1] Its title presents ELIZA as a computer program for the study of natural-language communication between man and machine.^[1] The abstract describes the program as operating within the MAC time-sharing system at MIT and as making certain kinds of natural-language conversation between man and computer possible.^[1] The paper is therefore organized around a concrete program and around the conditions under which that kind of conversation can be produced.

How ELIZA processes a sentence

In the account given by the paper, an input sentence is analysed on the basis of decomposition rules.^[1] These rules are triggered by key words appearing in the input text.^[1] Responses are generated by reassembly rules associated with the selected decomposition rules.^[1] The description can be read as a chain: a key word activates a decomposition rule; that rule provides the analysis of the input; and the reassembly rule associated with it produces the response. The same rule selection links the analysis of the input to the generation of the reply.

Five technical problems

The paper names five fundamental technical problems.^[1] The first is the identification of key words.^[1] The second is the discovery of minimal context.^[1] The third is the choice of appropriate transformations.^[1] The fourth is the generation of responses in the absence of key words.^[1] The fifth is the provision of an editing capability for ELIZA scripts.^[1] These are presented as technical problems, not as solved details; the list sets out what the program and its scripts must address for the kind of conversation described in the abstract. Identifying key words concerns which words in the input can trigger a rule. Minimal context concerns how much surrounding text a rule needs. Appropriate transformations concern the selection of the transformation to be used. Generation without key words concerns what happens when no keyword is present. Editing capability concerns work on the ELIZA scripts themselves.

Discussion and closing concerns

The paper concludes with a discussion of some psychological issues relevant to the ELIZA approach, as well as future developments.^[1] This closing discussion is part of the paper's scope alongside the description of the program. The paper thus combines a technical account of rule-based input analysis and response generation with a discussion of psychological issues and possible further work.

Why the paper is important

The paper's importance lies in how explicitly it separates the parts of a conversational program: key-word identification, decomposition, rule selection, and reassembly. It also records five recurring technical difficulties in a compact list. By describing ELIZA as a program within a time-sharing system and by setting out the rule mechanism and the open problems, the paper gives a clear statement of what a natural-language conversation program has to manage. Its closing discussion of psychological issues and future developments shows that the technical design was presented together with questions about the approach and about what might come next.^[1]