From newshub.ccs.yorku.ca!torn!utcsri!rutgers!jvnc.net!yale.edu!ira.uka.de!chx400!bernina!neptune!qvortrup Tue Jul 28 09:41:37 EDT 1992
Article 6475 of comp.ai.philosophy:
Xref: newshub.ccs.yorku.ca rec.arts.sf.science:2785 comp.ai.philosophy:6475
Path: newshub.ccs.yorku.ca!torn!utcsri!rutgers!jvnc.net!yale.edu!ira.uka.de!chx400!bernina!neptune!qvortrup
>From: qvortrup@inf.ethz.ch (Michael Qvortrup)
Newsgroups: rec.arts.sf.science,comp.ai.philosophy
Subject: Re: How do computers fare on scholastic achievement tests?
Message-ID: <1992Jul17.145433.19824@neptune.inf.ethz.ch>
Date: 17 Jul 92 14:54:33 GMT
References: <1992Jul14.014658.4921@newstand.syr.edu> <NICKH.92Jul14141610@VOILA.VENARI.CS.CMU.EDU> <1992Jul16.093057.8880@techbook.com>
Sender: news@neptune.inf.ethz.ch (Mr News)
Organization: Dept. Informatik, Swiss Federal Institute of Technology (ETH), Zurich, CH
Lines: 36
Nntp-Posting-Host: ru3.inf.ethz.ch

In article <1992Jul16.093057.8880@techbook.com> szabo@techbook.com (Nick Szabo) writes:
>[...]
>* Which questions would be easiest for a computer?  Which the most
>  difficult?  For example, I suspect a souped-up version of Mathematica
>  could do pretty well on many of the math questions.  A semantic analyzer 
>  with Mathematica might do well on some of the word problems.  Perhaps 
>  there is a logic to some of the verbal problems that a parser/semantic 
>  analyzer could tackle.  Some questions might be amenable to neural net 
>  training if the computer is fed thousands of related practice questions.  
>  Etc.

I doubt that you would have very much luck in doing that. Mathematica
(and Maple, which I prefer) is very good at solving problems specified
in mathematical form. If the problem isn't given in such a form but, say
in text form ('Farmer A has 5 potatoes ..... What is the profit?') then
you will basically have to solve the human-computer natural language
problem. This problem is as of now unsolved for context dependent texts,
i.e. more dependent sentences with implicit references. (At least the
problem hadn't been solved last time I checked :-)

I have here assumed, that no human is allowed to do any other preprocessing
than scanning the test and starting a program.

>* How difficult for a computer to scan in the questions from a standard
>  SAT test book into a recognizable form?

Depending on the type of question, varying from 'very' to 'intractable'.

Greetings,
--Mike

-- 
#include <std-disclm.h>--"... and there is a small flaw in my character."---
Real Life: Michael Christian Heide Qvortrup      A Dane         ETH, Zuerich
e-mail   : qvortrup@inf.ethz.ch                  abroad         Switzerland
Institut fuer wissenschaftliches Rechnen  /  Inst. of Scientific Computation


