Robotics Reference
In-Depth Information
Vice-president: Please dial me up on 491-1850 [1]
At this point in the conversation, and contrary to the instructions in
the note left on the teletype, the vice-president did
not
terminate his
last remark with a period, as a result of which the computer did not
respond to him because it did not know that he had completed what he
was typing. This lack of response from “Bobrow” so infuriated the vice-
president, who believed that Bobrow was playing games with him, that
he called Bobrow on the telephone, woke him from a deep sleep, and
said: “Why are you being so snotty to me?”, to which Bobrow replied,
with all sincerity, “What do you mean I am being snotty to you?” By
now of course the vice-president was quite angry, and read Bobrow the
dialog that “they” had been having, which caused Bobrow to burst into
uncontrollable laughter. It then took Bobrow a while to convince the
vice-president that he had actually been in conversation with a computer.
ELIZA and several other programs of the early era of conversational
software worked by employing simple matching techniques to recognize
the structure of a phrase or sentence in the user's input, and then to
respond with an utterance that corresponded with the form of the user's
input. For example, a program might recognize the sentence or phrase
I like cats
to be of the form
I
<
verb
><
plural noun
>
and it might have already stored the response
<
>
<
>
I
verb
them too, but why do you
verb
them?
which would be converted, in this case, to
I like them too, but why do you like them?
Thus, programs such as ELIZA had absolutely no understanding of the
language. They were merely outputting responses that conformed in style
to the type of response the user might have been expecting, and thereby
giving the impression that the program understood the conversation.
And by being endowed with a measure of knowledge about a particular
domain, programs were able not only to conduct simple conversations
but also to answer simple questions. In fact, question-answering has long
been one of the better performing tasks for NLP systems, principally
