From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!think.com!wupost!uunet!mcsun!uknet!edcastle!aiai!jeff Thu Jan  9 10:34:05 EST 1992
Article 2555 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!think.com!wupost!uunet!mcsun!uknet!edcastle!aiai!jeff
>From: jeff@aiai.ed.ac.uk (Jeff Dalton)
Newsgroups: comp.ai.philosophy
Subject: Re: Searle, again
Message-ID: <5906@skye.ed.ac.uk>
Date: 8 Jan 92 18:25:05 GMT
References: <2127@ucl-cs.uucp> <91338.113617KELLYDK@QUCDN.QueensU.CA> <5796@skye.ed.ac.uk> <40346@dime.cs.umass.edu> <5831@skye.ed.ac.uk> <40453@dime.cs.umass.edu>
Reply-To: jeff@aiai.UUCP (Jeff Dalton)
Distribution: sci.philosophy.tech
Organization: AIAI, University of Edinburgh, Scotland
Lines: 107

In article <40453@dime.cs.umass.edu> orourke@sophia.smith.edu (Joseph O'Rourke) writes:
>[Re: interpreting the Chinese Room symbol manipulations as moves 
>in a chess game] Jeff Dalton writes:
>
>>On the other hand, looked at from the other side, what if someone
>>found a way to encode chess moves as Chinese and "hacked" the rules
>>used in the Chinese Room?  
>
>	I still don't see how this is possible, but perhaps it is
>a failure of my hacking imagination.  

Do you think it's hard to encode the moves?  It ought to be
fairly easy for two people to devise a code so that they could
exchange chess moves while seeming to talk about something else.
Indeed, the "symbols" that refer to moves might be various and
long.  For instance, if someone introduces a walk in a park,
that means to castle on the king's side.  All kinds of different
talk about walks in parks would carry the same meaning, and there
could also be sequences that change the code.  Or there could
be some limitations on the kinds of talk, so that what looked
like completely flexible NL exchanges would actually be very
rigid ones.  It's just that there are, say, 500,000 ways to
say "my piece furthest to the upper left".  So there could be
all this structure that the NL interpretation failed to explain.

Nonsense "talk" could also be used (cf the Nonsense Room).

All of this breaks the idea that the encoding has to be some 
sort of simple, fixed mapping from symbols to objects.

Now someone writes a program for the Room that uses this code,
and the people who talk to the Room all know about it and use
the code too.  Can the person inside the Room discover that
this is what they're doing?

I originally entered this discussion because some people were
claiming that the person inside the room could learn Chinese
by watching the IO.  That seems to me completely wrong, but
your arguments that there's only one right structural
interpretation seem to imply otherwise.

On the other hand, Searle's claim that virtually any interpretation
could be placed on the symbols also seems to be wrong.  Chess happens
to be a bad example for him, if he supposes that the moves have to
form legal games.  The person in the Room couldn't just say "they're
Chess moves" and start reading off legal games.  However, it would
also be difficult to figure out the NL interpretation, and NL
interpretations could be subject to the problems indicated above.

>I assume an interpretation is fixed wrt time:  the symbols cannot
>map to different objects at different times. 

But that certainly isn't true of the NL interpretation.  In a
conversation one could even say "from now on, whenever I use `cat'
I'll actually mean cherry".  Or, in one part of a mathematics
text, X might mean one thing, but something else in another.

>	There is a NL interpretation: a mapping of the symbols to
>various language constructs so that the manipulations make sense for
>a language understanding system.  Now suppose that Searle is right:
>there is in addition a CHESS interpretation: a mapping of the symbols
>to chess pieces and chess square designations so that the manipulations
>correspond to moves in a chess game.  Some symbols must map under CHESS
>to e.g. "white knight," which I will abbreviate N. 

Not necessarily, they could map only to positions.  There's no
need to identify the pieces at all.  Indeed, in the usual
notations for Chess, there aren't any symbols that map to
white knight.  King's knight, yes, and context shows it's
a white one; but not white knight.  

In any case, you ought to be willing to agree that it's easy to
map the symbols to chess moves if we don't require that they form
a legal game (just as it's easy to map them to stock names and
prices).

At this point, I think your argument goes that there will be only
one interpretation that properly accounts for all the structural
features.  For instance, if I interpret the sybols as a series
of digits (easy) but you can find the works of Shakespeare, you
must be right that they really are the works of Shakespeare
(unless I can come up with something even better).  But this
sort of reasoning is not right in general, as the example of
finding backwards messages in rock music shows.

>These same symbols
>must map to something fixed under the NL interpretation; perhaps they
>map to a single natural language concept, say "transitive verb"; or
>perhaps the parsing is less fortuitous, and that symbols that map to N
>under CHESS map to several objects, or part of an object, under NL.
>Regardless, the mapping is fixed.
>
>	So now for the CHESS mapping to be truly a mapping to legal
>chess moves, we have to believe that the manipulations of the symbols
>that map to N by CHESS, are *structurally isomorphic* to the manipulations
>of those same symbols by NL, which might map them to "transitive verb."

I don't see why they have to be structurally isomorphic.

>This is what I find unbelievable:  that there could be such a coincidental
>structural similarity between knight moves and verb manipulations. The
>two semantic domains are just not that similar.

It doesn't have to be coincidental.  Perhaps someone is out to
play a trick.  How does the person in the Room know?

-- jd


