From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!ispd-newsserver!psinntp!scylla!daryl Thu Jan 16 17:22:04 EST 1992
Article 2750 of comp.ai.philosophy:
Newsgroups: comp.ai.philosophy
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!ispd-newsserver!psinntp!scylla!daryl
>From: daryl@oracorp.com
Subject: Re:
Message-ID: <1992Jan15.174826.28569@oracorp.com>
Organization: ORA Corporation
Date: Wed, 15 Jan 1992 17:48:26 GMT

>>The fact that there are a huge number of conversations consistent with
>>what has been said so far is irrelevant: the computer only needs to
>>select a response that occurs in *some* conversation that agrees with
>>the current conversation so far. If more than one response is possible,
>>the computer just picks one. Conversations don't have to have a unique
>>outcome.

>Oh, come now.  Such a system would almost immediately fail the Turing test.
>Picture it in operation.  Your first utterance is "Hello."  The computer
>races through its database looking for conversations that begin with this
>sentence.  There are rather a lot of them; it picks one at random:
>"Ah, so we meet again, Professor Chung!!"  You are taken aback, not least
>because your name is Daryl.

I meant *complete* conversations, starting from the first time the
conversants meet. That is why I allowed conversations of up to 100
years in length. By this definition, "Ah, so we meet again, Professor
Chung!" is *not* a sensible response to an opening sentence "Hello."

Or perhaps it is. I can always reply with "You must be mistaken, I am
not Professor Chung. My name is Daryl." His response would then have
to be part of a valid conversation beginning with:

   Q. Hello.
   A. Ah, so we meet again, Professor Chung!!
   Q. You must be mistaken, I am not Professor Chung. My name is Daryl.

His response might be:
   A. You can't fool me, I know you are really Professor Chung.

It's a matter of opinion whether you want to consider such a
conversation "sensible".

When I said there should be a database that contains only sensible
conversations, I mean conversations that would pass the Turing Test.
The assumptions behind the Turing Test is that there is no
communication between the participants other than through the
keyboard, so there is no way to communicate the fact that I am Daryl
other than by saying so.

> Now, the conversation "Hello."  "Ah, so we meet again, Professor Chung!"
> (or rather its translation into Chinese) is indeed a possible conversation,
> so it lives in the database.

It is not a possible complete conversation, where a complete conversation
includes everything that has ever been said between the two people.

> But it's not appropriate. The table lookup algorithm is incapable of any
> sense of context.  There are simply too many possible conversations at
> every point; all its replies will be nonsense.

If a reply is inappropriate, given the exchange so far, then it is
*not* a sensible reply, and I am assuming a database of only sensible
conversations. A sensible complete conversation will (usually) start
with our saying hello and introducing ourselves and telling a little
bit about our backgrounds.

As to whether a table lookup is capable of taking context into
account, I am assuming that there is no interaction between the human
and the program except through their conversation. Therefore, the only
context that can come into play is the context provided by what they
have said to each other in the past. Taking that context into account
is part of what I mean by a "sensible conversation".

> And no, don't talk to me about assigning probabilities to the
> conversations, either.  The next conversation you have is almost
> guaranteed to be unique, and thus can't be distinguished by its
> probability from any other conversation.

The way I have defined conversation, you only have *one* extended
conversation with each person. Everything you said in the past is part
of the single, on-going conversation. The relevant probabilities
involved are not the probabilities of entire conversations, but the
probability of the next reply, given what has gone on before.

> This algorithm is hardly worth discussing, except that I already see
> you using it in another reply to argue that a machine that can
> converse in Chinese is theoretically possible.

I thought that it would be uncontroversial. By whatever criterion you
use for "sensible conversations", the fact that the conversations are
only finite in length immediately implies that there exists a finite
state machine that has only sensible conversations. If you are only
interested in i/o behavior, then every finite state machine can be
implemented by a table lookup. I don't see anything at all
controversial in any of this. Note that I am *not* claiming that such
a machine could actually be built---it may very well have more states
than there are atoms in the universe---I am only saying that such a
machine exists in the mathematical sense that there exists a prime number
bigger than 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000.

> It's kind of fun, by the way, to see how big the database might be.  Let's
> limit our attention to conversations of 1 hour, and assume 1 syllable can
> be pronounced per second, and that Chinese has about 500 distinct syllables
> (we'll ignore tone).  I calculate about 10^2500 possible conversations.
> Will our database even fit in this universe?  Can the lookup be accomplished
> in the lifetime of the universe?

At the beginning of my article, I explicitly said that if we ignored
space and time limitations, then such a table lookup was possible. It
is certainly possible that we will never have a 20 Meg RAM, 33 MHZ
processor that can pass the Turing Test, so if you want to argue about
the practical impossibility of a machine that could pass the Turing Test,
I am sympathetic.

> The number of reasonable conversations would of course be smaller.
> However, there is no way to characterize the set of reasonable
> conversations without introducing Chinese grammatical information and
> real world knowledge, the sorts of things the table-lookup machine is
> not supposed to be worrying about.

I am assuming that grammatical and real world knowledge have already
been taken care of in the choice of what conversations to put in the
table: only sensible (both grammatically and semantically)
conversations are in the table, by assumption. Whether or not there is
a nice way to characterize the set of sensible conversations, since it
is a finite set, it is a computable set.

Daryl McCullough
ORA Corp.
Ithaca, NY








