When Seeing's Not Believing: Language, Magic, and AI by Naomi S. Baron

Abstract

This paper is rooted in the problem of how to interpret signs. The focus is on special class of signs: those used for communicating with computers. Especially, this article looks at the area of artificial intelligence (AI) known as natural language processing.

How do we interpret the signs used by (1)the natural language program designer and (2)the end user running such a program?

The main goal is to find out whether we can design computers that think in natural language and why we think machines might be intelligent and linguistic instead of wondering whether machines can think and talk.

Although artificial intelligence are critiqued due to its cost, dehumanizing effect, millions of research dollars are pouring into developing so called "fifth generation" computers.

To be human is to seek coherence. When confronted with stimuli, we attempt to make sense of them. Constructing interpretation --just as in the example of the computer ELIZA which emulates conversation of a Rogerian psychotherapist (generated by Joseph Weinzenbaum at MIT in mid 60s.)-- has been studied in various field and generated some common questions.

1) Can a phenomenon be interpreted in isolation, must its meaning derive from context?

2) Is knowledge an individual or a social product?

3) How will interpretations vary, depending upon the nature of our interaction, with the phenomenon we are interpreting?

Conversational Imperative: Ordinary language user comes face to face with another sentient being (includes humans, babies, non-fluent foreigners, pets). Although a baby or a non-fluent foreigner can not fully understand what we say, we attempt to make sense-- to provide a meaningful interpretation for social interaction.

Magic and Magicians: Then the author talks about magic, the art of misdirection, the rules for good magic. By exploring the conversational base of human language and the foundations of recreational magic, author recast the problem whether machines think and use language and what makes us think that machines might be intelligent and linguistic?

Computer scientist acts like a magician and employs the rules of a good magic. Natural language front ends are successful largely because users believe that the front end is a legitimate interlocutor, accepting natural language input ad responding in natural language in return. In actual fact, both comprehension and production are skillfully orchestrated illusions.

One critical difference between illusion and natural language front end is the matter of change. The purpose of recreational magic is always to entertain, never to be an instrument of change.

The author suggests that the average user needs to be able to function on two planes: to employ front ends as long as such programs are adequate for his/ her needs, but to understand enough about the underlying program structure so that be ready if a gap occurs between the linguistic illusion and the user's pragmatic need.