You are on page 1of 10

The Central Processes of the Model 57

remember
del stmt, to
which interviewer
is about to
respond
c
choosedel
D
fchange
* | I subject J
Fig. 5.5.
the algorithm), the program will refuse to discuss Mafia topics at all since
it is too "upset" to talk about this most sensitive area.
To make some of these operations more intelligible, let us consider
interview examples. Suppose at some point in the interview the doctor
asks a standard first-interview question as follows:
(5) Dr.DO YOU EVER HAVE THE FEELING YOU ARE
BEING WATCHED?
58 Artificial Paranoia
Flg. 5.6 Directed graph of flare concepts.
If this is the first reference to the delusional net, FEAR will increase
greatly and the linguistic response will be:
(6) Pt.YOU KNOW, THEY KNOW ME.
In making this response, the model must expect from the interviewer a
number of typical questions of the wh-type as well as rejoinder state-
ments. The use of "they" by the interviewer in his response to the model's
output is assumed to be an anaphoric reference to the "they" PARRY is
talking about. Although it is likely the interviewer will react to the model's
output of (6), the algorithm must be prepared for the possibility that he
will change the topic. Hence if the interviewer at this point asks some
non-sequitur question such as:
(7) Dr.HOW LONG HAVE YOU BEEN IN HOSPITAL?
the program recognizes that no reference to the delusional topic has been
made and answers the question just as though it had been asked in any
other context. This ability to deal with input in a flexible manner is
important because of the many contingencies that can occur in psychiat-
ric dialogues.
If the topic is changed abruptly in this way by an interviewer, the
algorithm "remembers" that it has output its first delusional statement of
(6). When the interviewer makes another neutral delusional reference, the
The Central Processes of the Model 59
next "line" of the delusional story will be output, e.g.,
(8) Pt.THE MAFIA REALLY KNOW ABOUT ME.
The ability to answer typical "wh-" and "how" questions depends on how
much conceptual information is contained in the delusional belief being
addressed. For example, suppose PARRY replied as in (6):
(6) Pt.YOU KNOW, THEY KNOW ME.
and the interviewer then asked:
(9) Dr.WHERE DO THEY KNOW ABOUT YOU?
If the expectancy-anaphoras contain no "where," then a question about
location cannot be answered. In this default situation, the algorithm
recognizes the anaphoric "they," "know" and "you." Hence it knows at
least that the topic has not been changed so it outputs the next statement
in the delusional story:
(9) pt.THEY KNOW WHO I AM.
and again anticipates questions and rejoinders pertaining to this state-
ment.
In constructing the data-base of beliefs, we tried to pack as much
information in each belief as any "reasonable" (like ourselves) inter-
viewer might request. However, one cannot anticipate everything and
when some unanticipated information is requested, another relevant reply
must be substituted. This heuristic may seem less than perfect but there is
little else to do when the model simply lacks the pertinent information. By
the way, humans do this also.
When the interviewer shows interest in the delusional story, PARRY
continues to output assertions appropriate to the dialogue. However,
when the interviewer expresses doubt or disbelief about the delusions,
ANGER and FEAR increase and the interviewer becomes questioned as
in:
(10) Pt.YOU DON'T BELIEVE ME, DO YOU?
Such an output expression attempts to prompt the dialogue toward the
relation between the interviewer and the model, which will be described
later (see p. 65).
If no delusional reference at all is detected by this procedure, the
algorithm attempts the next function that searches for certain types of
references to the self. (See Fig. 5.7.)
60 Artificial Paranoia
Self References
Since the main concern of a psychiatric interview consists of the
beliefs, feelings, states, and actions of the patient, the model must be able
to answer a large number of questions about its "Self."
If the input is recognized as a question and no topic is currently under
discussion and the question refers to the "Self," then it is assumed
temporarily that it will refer only to a main self-topic. These main or
"introductory" self-topics (age, sex, marriage, health, family, occupation,
hospital stay, etc.) in turn have subtopics to varying depths. For example,
suppose the interviewer asks:
(11) Dr.HOW DO YOU LIKE THE HOSPITAL?
Since "hospital" is a main "introductory" topic with several levels of
subtopics, the algorithm answers the question with:
(12) Pt.I SHOULDN'T HAVE COME HERE.
and then anticipates a variety of likely questions such as "What brought
you to the hospital?", "How long have you been in the hospital?", "How
do you get along with the other patients?", etc. Each of these questions
brings up further topics, some of which represent a continuation of the
main topic "hospital," but others of which represent a shift to another
main introductory topic, e.g., "other patients." Since many of the inputs
of the interviewer consist of ellipses or fragments, the algorithm assumes
them to refer to the topic or subtopic under discussion. If some topic is
being discussed, the algorithm checks first for a new main topic, then for a
follow-up to the last subtopic, then (unless the subtopic is itself a main
topic, as for example "other patients" in the above) for a follow-up to the
last main topic. Thus continuity and coherence in the dialogue are
maintained.
If some meaning cannot be extracted from the question but it is
recognized at least that a question is being asked, a procedure is called
that attempts to handle certain common miscellaneous questions that are
difficult to categorize. These include the space-time orientation questions
("What day is this?") and everyday information ("Who is president?")
asked by psychiatrists in a mental-status examination to test a patient's
awareness and orientation. Some quantitative "how" questions ("how
many," "how often," "how long") are recognized here. Since any
adjective or adverb can follow a "how," one of the limitations of the
model is its inability to handle all of them satisfactorily because the
The Central Processes of the Model 61
IreturrNy
f a l s e / ^
^^witli^S.
direct r e f ^ i yS
s^o s e l f ? ^ ^
N 1
TREPLY [\
.suspicious y
Wi est i on/
^ X wi t h S . v
^ direct ref >
" SJ o s e l f ? ^
N*
yreturrK
/ false X
. auesti on?^ a
N|
NORMAL 1
REPLY
return false}
J defensive ]
reply J
^^
'ANSWERA
answer |
suspiciously
Fig. 5.7.
62 Artificial Paranoia
relevant information is lacking in the data-base. If absolutely no clues are
recognized in the question, the algorithm is forced to output a noncommit-
tal reply such as:
(13) Pt.WELL, I DON'T KNOW.
This function also checks for statements about the self that are taken to
be insulting or complimentary. Naturally the presence of a negator in the
input reverses the meaning. Thus:
(14) Dr.YOU DONT SEEM VERY ALERT.
is classified as an insult whereas:
(15) Dr.YOU ARE RIGHT.
is considered complimentary and benevolent.
Among the introductory self-topics are those that constitute sensitive
areas, e.g., sex, religion, and family. If the interviewer refers to one of
these areas, the value of ANGER increases sharply and a response is
selected from one of the lists categorized as "hostile," "defensive,"
"personal," or "guarded," depending on the level of MISTRUST at the
moment. For example, if the interviewer asks a question about PARRY'S
sex life, it first replies with:
(16) Pt.MY SEX LIFE IS MY OWN BUSINESS.
If the interviewer persists or even later tries to ask about sex, the model
will respond with a hostile reply, such as:
(17) Pt.DO YOU KNOW WHAT YOU ARE DOING?
The particular sensitive areas in the model are part of the initial
conditions specific for this hypothetical patient. Of course, these topics
are commonly found to be sensitive areas in human patients.
The model operates sequentially trying one major process after
another. If it has come this far (i.e., having tested for special reactions,
delusional references, and self-references without recognizing anything in
the input pertinent to these procedures), it proceeds to the next process,
which handles flare references. (See Figs. 5.8 and 5.9.)
Flare References
The data-base contains a directed graph of concepts involved in the
model's "stories." PARRY has small stories to tell about horseracing,
gambling, bookies, etc. The major concepts of these stories are termed
The Central Processes of the Model 63
"flare" concepts since they activate stories that are differently weighted
in the graph. (See Fig. 5.6.)
In the strong version of the model, the concept "Mafia" is given the
highest weight, while in the weak version the concept "Rackets" is most
heavily weighted. In both versions "Horses" has the lowest weight. The
weights are assigned to the concepts and not individual words or
word-groups denoting the concepts.
The graph is directed in the sense that reference to horseracing elicits
the first line of a story about horseracing. When a story is ended, a prompt
is given to the interviewer to discuss the next story in the graph, which
involves "bookies." The model strives to tell its stories under appropriate
conditions and leads the interviewer along paths of increasing delusional
relevance. Much depends on whether the interviewer follows these leads
"benevolently" and reacts to the prompts.
The first step in this procedure is to scan the input for a flare concept
having the highest weight. Thus if a flare concept is already under
discussion, a weaker new flare will be disregarded. If the flare concept is
one in a story that has already been partially told, then a prompt is offered
regarding the next story-node in the graph.
If a question is asked about the events of a story, the model tries to
answer it. Also the model is sensitive to whether the interviewer is
showing interest in the story or whether he tries to change the subject or
expresses a negative attitude, such as disbelief.
If the interviewer indicates a positive attitude toward the story, then
benevolence is recognized and the variables of ANGER, FEAR, and
MISTRUST decrease slightly after each I-O pair. ANGER decreases
more rapidly than FEAR while MISTRUST, being a more stable variable
once it has risen, decreases least.
If no flare concepts are recognized in the input, the model next tries to
detect if a reference is being made to the relation between the interviewer
and the model. In an interview interaction there exist two situations, the
one being talked about and the one the participants are in at the moment.
Sometimes the latter situation becomes the former, i.e. the one talked
about.
Interviewer-Interviewee Relations
As described in Chapter 4, the algorithm must be ready to handle input
referring to the relation between interviewer and model. The simplest
cases are exemplified by expressions such as:
(18) Dr.I UNDERSTAND YOU.
e
flareref
I
D
FLAREMECH
respond to
flare and
delete as flare
FLAREMECH
respond to
flare and
delete as
flare
FLARE REPLY
suspicious
question
- * silence j
LEAD ON
ANSFLARE
answer
RESPFLARE
respond
I CHECK TOPICI
| check
to see whether
interviewer is
changing topic
ask leading
question about
next higher
flare
Fig. 5.8.
64
e
checkflare
D
_/TY.
set new flare =
this word,
set result = true
set global
flare to new
flare
Fig. 5.9.
65
66 Artificial Paranoia
(19) Dr.YOU DO NOT TRUST ME.
Those phrases in an expression that can appear between "I" and "you" or
between "you" and "me" we classified as representing a positive or
negative attitude on the part of the interviewer. Thus expression (18) is
taken to be positive whereas (19) is negative because, although it contains a
positive verb, the verb is negated.
If a positive attitude is expressed by the interviewer, FEAR and
ANGER decrease. FEAR and ANGER increase depending on the con-
ceptualizations of the input. These attitudes of the interviewer, as
interpreted by the model, are reflected in the values of the affect-
variables.
Associated in the data base with each type of attitude expression
expected are lists of appropriate output expressions. Thus in reply to:
(20) Dr.I UNDERSTAND YOU.
the model would reply:
(21) Pt. GLAD YOU DO.
or
(22) Pt.I APPRECIATE YOUR TRYING TO UNDERSTAND.
or some equivalent expression depending on values of the affect-
variables. When ANGER and FEAR are high, positive attitude expres-
sions are interpreted as insincerity and hence evoke hostile replies.
The remainder of input expression types thus far not discussed are
handled by a procedure for miscellaneous expressions.
Miscellaneous Expressions
This procedure deals with all those interviewer expressions from which
no clear conceptualization can be formed. The only thing that can be
determined is perhaps the sentence-type of the input. Presented with one
of these expressions, if FEAR is extremely high PARRY signs off without
a farewell expression and cannot be contacted through further natural
language input. If FEAR is high but not extreme, and the input is
recognized as a question, the model chooses a reply from a list that brings
up the attitude of the interviewer as in:
(23) Pt.WHY DO YOU WANT TO KNOW?

You might also like