paper.txt

megahal is the conversation simulators conversing with a user in natural language. The program will

                            INTRODUCING MEGAHAL

Jason L. Hutchens                                       Michael D. Alder
Dept. of E&E Engineering                            Dept. of Mathematics
University of Western Australia          University of Western Australia
Nedlands W.A. 6907, Australia              Nedlands W.A. 6907, Australia
[e-mail]                          [e-mail]



Abstract

Conversation simulators are computer programs which give the appearance of
conversing with a user in natural language.  Alan Turing devised a simple
test in order to decide whether such programs are intelligent.  In 1991,
the Cambridge Centre for Behavioural Studies held the first formal
instantiation of the Turing Test.  In this incarnation the test was known as
the Loebner contest, as Dr. Hugh Loebner pledged a $100,000 grand prize for
the first computer program to pass the test.  In this paper we give a brief
background to the contest, before describing in detail the workings of
MegaHAL, the primary author's entry to the 1998 Loebner contest.



1. Introduction

Alan Turing was a brilliant British mathematician who played a great role in
the development of the computer.  The imitation game, nowadays known as the
Turing test, was devised by Turing as a method for deciding whether or not a
computer program is intelligent.

The Turing test takes place between an interrogator and two subjects.  The
interrogator communicates with these subjects via a computer terminal, and
must decide which is a human being and which is a computer program.  The
human being helps the interrogator to make the correct identification, while
the computer program attempts to trick the interrogator into making the
wrong identification.  If the latter case occurs, the computer program is
said to be exhibiting intelligence (Turing, 1992).

One of the great advantages of the Turing test is that it allows the
interrogator to evaluate almost all of the evidence that we would assume to
constitute thinking (Moor, 1976).  For instance, the interrogator can pose
hypothetical situations in order to ask the subjects how they would react.

Alan Turing died in 1954, a decade before conversation simulators such as
ELIZA emerged.  It is indeed unfortunate that he did not live to witness
his test being performed.  One cannot help but think that he would have
been disappointed.



2. The Loebner Contest

Apart from a few limited tests performed by programmers of conversation
simulators (Colby, 1981), the Turing test was not formally conducted until
1995.  Although the inaugural Loebner contest, held in 1991, was touted as
the first formal instantiation of the Turing test, it was not until 1995
that it truly satisfied Turing's original specifications (Hutchens, 1996).

The first Loebner contest was held on the 8th of November 1991 in Boston's
Computer Museum.  Because this was a contest rather than an experiment, six
computer programs were accepted as subjects.  Four human subjects and ten
judges were selected from respondents to a newspaper advertisement; none of
them had any special expertise in Computer Science (Epstein, 1992).

The original Turing test involved a binary decision between two subjects by
a single judge.  With ten subjects and ten judges, the situation was
somewhat more complex.  After months of deliberation, the prize committee
developed a suitable scoring mechanism.  Each judge was required to rank
the subjects from least human-like to most human-like, and to mark the
point at which they believed the subjects switched from computer programs
to human beings.

If the median rank of a computer program exceeded the median rank of at
least one of the human subjects, then that computer program would win the
grand prize of $100,000[1].  If there was no grand prize winner, the
computer program with the highest median rank would win the contest with a
prize of $2000.



3. Conversation Simulators

Since its inception, the Loebner contest has primarily attracted hobbyist
entries which simulate conversation using template matching; a method
employed by Joseph Weizenbaum in his ELIZA conversation simulator, developed
at MIT between 1964 and 1966.  Put simply, these programs look for certain
patterns of words in the user's input, and reply with a pre-determined
output, which may contain blanks to be filled in with details such as the
user's name.

Such programs are effective because they exploit the fact that human beings
tend to read much more meaning into what is said than is actually there; we
are fooled into reading structure into chaos, and we interpret non-sequitur
as whimsical conversation (Shieber, 1994).

Weizenbaum was shocked at the reaction to ELIZA.  He noticed three main
phenomenon which disturbed him greatly (Weizenbaum, 1976):

1. A number of practising psychiatrists believed that ELIZA could grow into
   an almost completely automatic form of psychotherapy.

2. Users very quickly became emotionally involved---Weizenbaum's secretary
   demanded to be left alone with the program, for example.

3. Some people believed that the program demonstrated a general solution to
   the problem of computer understanding of natural language.

Over three decades have passed since ELIZA was created.  Computers have
become significantly more powerful, while storage space and memory size
have increased exponentially.  Yet, at least as far as the entrants of the
Loebner contest go, the capabilities of conversation simulators have
remained exactly where they were thirty years ago.  Indeed, judges in the
1991 contest said that they felt let down after talking to the computer
entrants, as they had had their expectations raised when using ELIZA during
the selection process.



4. MegaHAL

In 1996 the primary author entered the Loebner contest with an ELIZA
variant named HeX, which was written during his spare time in under a
month.  Apart from the lure of the prize money, a major motivation for the
entry was a desire to illustrate the shortcomings of the contest (Hutchens,
1996).  A considerably more powerful program, SEPO, was entered the
following year, where it was placed second.  We believe this to be
indicative of a gradual improvement in the quality of the contestants.

The program submitted to this year's contest, MegaHAL, uses a significantly
different method of simulating conversation than either HeX or SEPO, and we
dedicate the remainder of this paper to describing its workings.

4.1 Markov Modelling

MegaHAL is able to construct a model of language based on the evidence it
encounters while conversing with the user.  To begin with, the input
received from the user is parsed into an alternating sequence of words and
non-words, where a word is a series of alphanumeric characters and a
non-word is a series of other characters.  This is done to ensure not only
that new words are learned, but that the separators between them are
learned as well.  If the user has a habit of putting a double space after a
full stop, for instance, MegaHAL will do just the same.

The resulting string of symbols[2] is used to train two 4th-order Markov
models (Jelinek, 1986).  One of these models can predict which symbol will
following any sequence of four symbols, while the other can predict which
symbol will precede any such sequence.  Markov models express their
predictions as a probability distribution over all known symbols, and are
therefore capable of choosing likely words over unlikely ones.  Models of
order 4 were chosen to ensure that the prediction is based on two words;
this has been found necessary to produce output resembling natural language
(Hutchens, 1994).
4.2 Generating Candidate Replies

Using a Markov model to generate replies is easy; Shannon was doing much
the same thing by flipping through books back in 1949 (Shannon, 1949).
However, such replies will often be nonsensical, and will bear no
relationship to the user's input.

MegaHAL therefore attempts to generate suitable replies by basing them on
one or more keywords from the user's input.  This explains why two Markov
models are necessary; the first model generates a sentence from the keyword
on, while the second model generates the remainder of the sentence, from
the keyword back to the beginning.

Keywords are obtained from the users input.  Frequently occurring words,
such as "the", "and" and "what", are discarded, as their presence in the
input does not mean they need to be present in the output.  The remaining
words are transformed if necessary---"my" becomes "your" and "why" becomes
"because", for example.  What remains is used to seed the output.

4.3 Selecting a Reply

MegaHAL is able to generate many hundreds of candidate replies per second,
each of which contain at least one keyword.  Once a small time period has
elapsed, the program must display a reply to the user.  A method is needed
for selecting a suitable reply out of the hundreds of candidates.

   I(w|s) = -log P(w|s)       - Equation 1

MegaHAL chooses the reply which assigns the keywords the highest
information.  The information of a word is defined in Equation 1 as the
surprise it causes the Markov model.  Hence the most surprising reply is
selected, which helps to guarantee its originality.  Note that P(w|s) is
the probability of word w following the symbol sequence s, according to the
Markov model.

The algorithm for MegaHAL proceeds as follows: