Eliza - der Computer als Gesprächspartner


Autor: Andre Adrian
Version: 22.Apr.2007

Der Eliza-Effekt

Das Programm Eliza wurde 1966 von Joseph Weizenbaum geschrieben. Eliza analysiert die Texteingabe des Gesprächspartners nach einfachen Regeln und produziert aus Teilen der Eingabe und eigenen Satzbausteinen eine Antwort. Berühmt wurde Eliza mit einem Regelsatz welcher die Gesprächpsychotherapie von Carl Rogers nachbildet.
Joseph Weizenbaum hat mit Eliza ein Künstliches Intelligenz Programm geschaffen welches nach 40 Jahren immer noch beachtet wird. Der Erfolg von Eliza war dem Autor wohl etwas unheimlich. Die Eliza Manie zeigt doch nur, daß ein Mensch alles vermenschlicht. Blitz und Donner wurden als Götter vermenschlicht. In Fabeln, Märchen und Romanen wie z.B. "Watership Down" treten sprechende, vermenschlichte Tiere auf.
Es scheint dem Menschen unmöglich zu sein den menschlichen Blickwinkel abzulegen. Alles nicht-menschliche in der Umwelt wird als mehr oder minder seltsamer Mitmensch angesehen. Vielleicht ist dies die Funktionsweise der Spiegelneuronen: Wir können uns in ein rollendes Rad welches kurz vor dem Umkippen ist hineinversetzen. Wir können sinnvoll die Zukunft "gleich kippt es um" vorhersehen. Und weil wir das nötige "physikalische Gefühl" für das Rad haben, funktioniert die menschliche Empathie (Einfühlungsvermögen). Die Gedankenkette wird mit der Schlußfolgerung: "Wenn ich mich hineinversetzen kann, dann muß es ein Mitmensch sein" abgeschlossen.
Alle Tamagotchis dieser Welt und wohl ein Großteil der Dackel funktionieren für ihre Besitzer nach diesem Muster: Das Verhalten ist verständlich weil vorhersehbar, der Besitzer kann sich hineinversetzen. Die Spiegelneuronen funktionieren wie wenn ein Mitmensch die beobachtete Handlung ausführt. Und weil die Spiegelneuronen arbeiten muß das beobachtete Objekt ein Mitmensch sein. Vielleicht aus Plastik, mit drei Knöpfen und einem LCD Bildschirm, aber trotzdem.
Eigentlich ist der Eliza-Effekt nichts Neues und vor allem nichts Unheimliches.
Leider gab es im Bereich Künstliche Intelligenz die Verwechslung zwischen "der Programm Autor ist intelligent" und "das Programm ist intelligent". Das Eliza Programm zeigt keine Intelligenz im Sinne von "lernfähiges Programm".

Klientenzentrierte Psychotherapie

Die Therapie von Carl Rogers ist heute fest etabliert. Einmal in der Besserung oder Heilung von psychischen und psychosomatischen Störungen. Viel wichtiger ist ihre Bedeutung für die Kommunikation zwischen "Gesunden". Niccolò Machiavelli sagt "der Mensch ist schlecht". Carl Rogers sagt "der Mensch ist gut". Beide Autoren können sich auf empirische Beweise stützen. Vielleicht zeigen Machiavelli und Rogers die beiden Extreme des menschlichen Verhaltens.
Die klientenzentrierte Psychotherapie macht den Therapeuten zum Geburtshelfer der Selbsthilfe. Eigentlich weiss der Klient ja selbst, was am besten für ihn ist. Nur kommt dieses Wissen ohne Therapeut nicht heraus. Für die Geburtshelfer Funktion des Therapeuten sind drei Verhaltenselemente nötig:
Das Eliza Programm hat diese Therapeuten Anforderungen erfüllt, oder besser gesagt: Eliza hat die Gesprächspartner nicht enttäuscht. Wenn im Regelsatz von Eliza nur Aussagen mit positiver Wertschätzung sind kann Eliza nur eine positive Wertschätzung zeigen. Wenn die Sätze des Klienten als Fragen gespiegelt werden, wird eine "Seelenverwandtschaft" zwischen Klienten und Programm suggeriert. Sind die Antwortphrasen auf Schlüsselwörter des Klienten gut gewählt, entsteht sogar der Eindruck von "Eigeninitiative" des Programmes. Werden die Antwortphrasen passend zu den Schlüsselwörtern gewählt, "lügt" Eliza nicht.
Joseph Weizenbaum schreibt hierzu: "It is important to note that this assumption is one made by the speaker. Whether it is realistic or not is an altogether separate question. In any case, it has a crucial psychological utility in that it serves the speaker to maintain his sense of being heard and understood. The speaker furher defends his impression (which even in real life may be illusory) by attributing to his conversational partner all sorts of background knowledge, insights and reasoning ability. But again, these are the speaker's contribution to the conversation."

Gespräch zwischen zwei KIs

Neben Eliza sind noch weitere ähnliche Künstliche Intelligenz Programm geschrieben worden. RFC439 enthält das Gespräch zwischen Eliza und Parry vom 18 September 1972. Dabei ist Eliza der Doktor und Parry der Patient. Das Programm Parry simuliert einen paranoiden Schizophrenen. Der Autor von Parry ist Kenneth Colby.  Ein Auszug (Parry beginnt):
   *Do you know anything about bookies?

   What makes you think I know anything about bookies?

   *A bookie didn't pay me off once.

   What does that suggest to you?

   *I beat up the bookie who did not pay off, and I thought he might use
   his friends in the underworld to get even with me.

Die Eliza Regeln

Joseph Weizenbaum hat die Eliza Regeln in dem Papier "ELIZA--A Computer Program For the Study of Natural Language Communication Between Man and Machine" beschrieben. Schlüsselwort und Satz Transformation sind die beiden Kernelemente von Eliza

Eliza Schlüsselwörter

In der Eingabe des Gesprächspartners wird nach Schlüsselworten gesucht. Die Schlüsselwörter in dem kleine Eliza Program therapist von Joe Strout in Python sind:

Eliza Transformationen

Eine Eliza Sitzung

>>> import eliza
>>> eliza.command_interface()
Therapist
---------
Talk to the program by typing in plain English, using normal upper-
and lower-case letters and punctuation.  Enter "quit" when done.
========================================================================
Hello.  How are you feeling today?
>I am happy.
Did you come to me because you are happy?
>Yes.
OK, but can you elaborate a bit?
>I think I like to share my happyness.
Do you really think so?
>Why not?
Why don't you tell me the reason why not?
>Because positive thinking wins the world.
If positive thinking wins the world, what else must be true?
>That all men are brothers.
How does that make you feel?
>Happy.
Please tell me more.
>You should try happyness yourself.
Why do you say that about me?
>Because you are so negative.
Does that reason apply to anything else?
>Computers are negative, too.
Why do you say that computers am negative, too?
>Because computers can not talk proper english.
What other reasons come to mind?
>That this chat is long enough.
Can you elaborate on that?
>quit
Thank you for talking with me.
>>>

Der Quelltext

Der Quelltext ist kompatibel zu Python 2.5.1.

#----------------------------------------------------------------------
#  eliza.py
#
#  a cheezy little Eliza knock-off by Joe Strout <joe@strout.net>
#  with some updates by Jeff Epler <jepler@inetnebr.com>
#  hacked into a module and updated by Jez Higgins <jez@jezuk.co.uk>
#  changed whrandom to random by Andre Adrian
#  last revised: 21 April 2007
#----------------------------------------------------------------------

import string
import re
import random

class eliza:
  def __init__(self):
    self.keys = map(lambda x:re.compile(x[0], re.IGNORECASE),gPats)
    self.values = map(lambda x:x[1],gPats)

  #----------------------------------------------------------------------
  # translate: take a string, replace any words found in dict.keys()
  #  with the corresponding dict.values()
  #----------------------------------------------------------------------
  def translate(self,str,dict):
    words = string.split(string.lower(str))
    keys = dict.keys();
    for i in range(0,len(words)):
      if words[i] in keys:
        words[i] = dict[words[i]]
    return string.join(words)

  #----------------------------------------------------------------------
  #  respond: take a string, a set of regexps, and a corresponding
  #    set of response lists; find a match, and return a randomly
  #    chosen response from the corresponding list.
  #----------------------------------------------------------------------
  def respond(self,str):
    # find a match among keys
    for i in range(0,len(self.keys)):
      match = self.keys[i].match(str)
      if match:
        # found a match ... stuff with corresponding value
        # chosen randomly from among the available options
        resp = random.choice(self.values[i])
        # we've got a response... stuff in reflected text where indicated
        pos = string.find(resp,'%')
        while pos > -1:
          num = string.atoi(resp[pos+1:pos+2])
          resp = resp[:pos] + \
            self.translate(match.group(num),gReflections) + \
            resp[pos+2:]
          pos = string.find(resp,'%')
        # fix munged punctuation at the end
        if resp[-2:] == '?.': resp = resp[:-2] + '.'
        if resp[-2:] == '??': resp = resp[:-2] + '?'
        return resp

#----------------------------------------------------------------------
# gReflections, a translation table used to convert things you say
#    into things the computer says back, e.g. "I am" --> "you are"
#----------------------------------------------------------------------
gReflections = {
  "am"   : "are",
  "was"  : "were",
  "i"    : "you",
  "i'd"  : "you would",
  "i've"  : "you have",
  "i'll"  : "you will",
  "my"  : "your",
  "are"  : "am",
  "you've": "I have",
  "you'll": "I will",
  "your"  : "my",
  "yours"  : "mine",
  "you"  : "me",
  "me"  : "you"
}

#----------------------------------------------------------------------
# gPats, the main response table.  Each element of the list is a
#  two-element list; the first is a regexp, and the second is a
#  list of possible responses, with group-macros labelled as
#  %1, %2, etc. 
#----------------------------------------------------------------------
gPats = [
  [r'I need (.*)',
  [  "Why do you need %1?",
    "Would it really help you to get %1?",
    "Are you sure you need %1?"]],
 
  [r'Why don\'?t you ([^\?]*)\??',
  [  "Do you really think I don't %1?",
    "Perhaps eventually I will %1.",
    "Do you really want me to %1?"]],
 
  [r'Why can\'?t I ([^\?]*)\??',
  [  "Do you think you should be able to %1?",
    "If you could %1, what would you do?",
    "I don't know -- why can't you %1?",
    "Have you really tried?"]],
 
  [r'I can\'?t (.*)',
  [  "How do you know you can't %1?",
    "Perhaps you could %1 if you tried.",
    "What would it take for you to %1?"]],
 
  [r'I am (.*)',
  [  "Did you come to me because you are %1?",
    "How long have you been %1?",
    "How do you feel about being %1?"]],
 
  [r'I\'?m (.*)',
  [  "How does being %1 make you feel?",
    "Do you enjoy being %1?",
    "Why do you tell me you're %1?",
    "Why do you think you're %1?"]],
 
  [r'Are you ([^\?]*)\??',
  [  "Why does it matter whether I am %1?",
    "Would you prefer it if I were not %1?",
    "Perhaps you believe I am %1.",
    "I may be %1 -- what do you think?"]],
 
  [r'What (.*)',
  [  "Why do you ask?",
    "How would an answer to that help you?",
    "What do you think?"]],
 
  [r'How (.*)',
  [  "How do you suppose?",
    "Perhaps you can answer your own question.",
    "What is it you're really asking?"]],
 
  [r'Because (.*)',
  [  "Is that the real reason?",
    "What other reasons come to mind?",
    "Does that reason apply to anything else?",
    "If %1, what else must be true?"]],
 
  [r'(.*) sorry (.*)',
  [  "There are many times when no apology is needed.",
    "What feelings do you have when you apologize?"]],
 
  [r'Hello(.*)',
  [  "Hello... I'm glad you could drop by today.",
    "Hi there... how are you today?",
    "Hello, how are you feeling today?"]],
 
  [r'I think (.*)',
  [  "Do you doubt %1?",
    "Do you really think so?",
    "But you're not sure %1?"]],
 
  [r'(.*) friend (.*)',
  [  "Tell me more about your friends.",
    "When you think of a friend, what comes to mind?",
    "Why don't you tell me about a childhood friend?"]],
 
  [r'Yes',
  [  "You seem quite sure.",
    "OK, but can you elaborate a bit?"]],
 
  [r'(.*) computer(.*)',
  [  "Are you really talking about me?",
    "Does it seem strange to talk to a computer?",
    "How do computers make you feel?",
    "Do you feel threatened by computers?"]],
 
  [r'Is it (.*)',
  [  "Do you think it is %1?",
    "Perhaps it's %1 -- what do you think?",
    "If it were %1, what would you do?",
    "It could well be that %1."]],
 
  [r'It is (.*)',
  [  "You seem very certain.",
    "If I told you that it probably isn't %1, what would you feel?"]],
 
  [r'Can you ([^\?]*)\??',
  [  "What makes you think I can't %1?",
    "If I could %1, then what?",
    "Why do you ask if I can %1?"]],
 
  [r'Can I ([^\?]*)\??',
  [  "Perhaps you don't want to %1.",
    "Do you want to be able to %1?",
    "If you could %1, would you?"]],
 
  [r'You are (.*)',
  [  "Why do you think I am %1?",
    "Does it please you to think that I'm %1?",
    "Perhaps you would like me to be %1.",
    "Perhaps you're really talking about yourself?"]],
 
  [r'You\'?re (.*)',
  [  "Why do you say I am %1?",
    "Why do you think I am %1?",
    "Are we talking about you, or me?"]],
 
  [r'I don\'?t (.*)',
  [  "Don't you really %1?",
    "Why don't you %1?",
    "Do you want to %1?"]],
 
  [r'I feel (.*)',
  [  "Good, tell me more about these feelings.",
    "Do you often feel %1?",
    "When do you usually feel %1?",
    "When you feel %1, what do you do?"]],
 
  [r'I have (.*)',
  [  "Why do you tell me that you've %1?",
    "Have you really %1?",
    "Now that you have %1, what will you do next?"]],
 
  [r'I would (.*)',
  [  "Could you explain why you would %1?",
    "Why would you %1?",
    "Who else knows that you would %1?"]],
 
  [r'Is there (.*)',
  [  "Do you think there is %1?",
    "It's likely that there is %1.",
    "Would you like there to be %1?"]],
 
  [r'My (.*)',
  [  "I see, your %1.",
    "Why do you say that your %1?",
    "When your %1, how do you feel?"]],
 
  [r'You (.*)',
  [  "We should be discussing you, not me.",
    "Why do you say that about me?",
    "Why do you care whether I %1?"]],
   
  [r'Why (.*)',
  [  "Why don't you tell me the reason why %1?",
    "Why do you think %1?" ]],
   
  [r'I want (.*)',
  [  "What would it mean to you if you got %1?",
    "Why do you want %1?",
    "What would you do if you got %1?",
    "If you got %1, then what would you do?"]],
 
  [r'(.*) mother(.*)',
  [  "Tell me more about your mother.",
    "What was your relationship with your mother like?",
    "How do you feel about your mother?",
    "How does this relate to your feelings today?",
    "Good family relations are important."]],
 
  [r'(.*) father(.*)',
  [  "Tell me more about your father.",
    "How did your father make you feel?",
    "How do you feel about your father?",
    "Does your relationship with your father relate to your feelings today?",
    "Do you have trouble showing affection with your family?"]],

  [r'(.*) child(.*)',
  [  "Did you have close friends as a child?",
    "What is your favorite childhood memory?",
    "Do you remember any dreams or nightmares from childhood?",
    "Did the other children sometimes tease you?",
    "How do you think your childhood experiences relate to your feelings today?"]],
   
  [r'(.*)\?',
  [  "Why do you ask that?",
    "Please consider whether you can answer your own question.",
    "Perhaps the answer lies within yourself?",
    "Why don't you tell me?"]],
 
  [r'quit',
  [  "Thank you for talking with me.",
    "Good-bye.",
    "Thank you, that will be $150.  Have a good day!"]],
 
  [r'(.*)',
  [  "Please tell me more.",
    "Let's change focus a bit... Tell me about your family.",
    "Can you elaborate on that?",
    "Why do you say that %1?",
    "I see.",
    "Very interesting.",
    "%1.",
    "I see.  And what does that tell you?",
    "How does that make you feel?",
    "How do you feel when you say that?"]]
  ]

#----------------------------------------------------------------------
#  command_interface
#----------------------------------------------------------------------
def command_interface():
  print "Therapist\n---------"
  print "Talk to the program by typing in plain English, using normal upper-"
  print 'and lower-case letters and punctuation.  Enter "quit" when done.'
  print '='*72
  print "Hello.  How are you feeling today?"
  s = ""
  therapist = eliza();
  while s != "quit":
    try: s = raw_input(">")
    except EOFError:
      s = "quit"
      print s
    while s[-1] in "!.": s = s[:-1]
    print therapist.respond(s)


if __name__ == "__main__":
  command_interface()