An Assessment of the Evolution of Research and Systems - Man-Machine Dialogue: Design and Challenges - page 7

you are listening to my advice”, generated after “I am listening to your
advice”. With this example, we can note that the system does not understand
much, but it is able to switch the persons around and frame the user's
utterance in a question “what makes you believe that”, a deliberately open
question. The techniques implemented by the input utterances are word
sequence detection and keyword detection. Those implemented for the
system's output utterance generation are the direct production of typical
sentences, the concatenation of text span, whether they are typical spans or
spans obtained through a user's utterance. The system also has the beginning
of memory, inasmuch as it is able to return to a familiar term used a few
speaking turns prior.
A few years after ELIZA, the PARRY [COL 71] system had an impact due
to its supplementary techniques. This time the machine simulates a paranoid
subject during his first (written) interview with the user who is supposed to
play the role of a psychiatrist, a profession to which the main author
incidentally belongs. The claimed scientific approach is the studying and
modeling of paranoia, and this goes so far as the funding that comes in part
from the National Institute of Mental Health, and the methodology that
includes not only the modeling and computer science development of the
model, but also its assessment by mental health professionals: a total of 25
psychiatrists were involved, and the overwhelming majority of them (23)
diagnosed the system as paranoid, making it pass the Turing test with flying
colors. The dialogues are carried out as interviews and start with the factual
questions that the user asked the system: name, age, occupation. Thus,
PARRY has in his memory a set of answers to these typical questions: his
name is Frank Smith, he is 28 years old, and interned in a hospital. He also
has in his memory various questions that the system can ask, thus inverting
the dialogue orientation: “who are you?”, “what do you want with me?”, as
well as anecdotes, and especially words around a relatively well-elaborated
concept, such as that of mafia. The techniques implemented are also
techniques of text span research, keyword detection, first and second person
pronoun management, but all with more finesse than ELIZA had. For
example, the word “fear” has a set of predefined spans, and verbs such as “to
believe” have specific processes. Moreover, the system is characterized by an
attempt at personality or mental states through variables: fear, anger and
distrust. The values of these variables increase or decrease as the dialogue
unfolds, according to what the user says. The system's behavior evolves in a