00300	.SEC PROBLEMS FOR COMPUTER UNDERSTANDING OF NATURAL LANGUAGE
00400	COMMUNICATION IN TELETYPED PSYCHIATRIC INTERVIEWS
00500	
00600		Since the behavior being simulated by this
00700	paranoid  model  is  the language-behavior of a paranoid patient in a
00800	psychiatric interview, the model must have an  ability  to  interpret
00900	and respond to natural language input  sufficient  to demonstrate
01000	linguistic-behavior characteristic of the paranoid mode.  
01100	By "natural language" I shall mean ordinary American English such  as
01200	is  used   in everyday conversations. It is
01300	still difficult to be  explicit  about  the  processes  which  enable
01400	humans  to  interpret and respond to natural language. Philosophers,
01500	linguists and psychologists have investigated natural  language  with
01600	various purposes and yielding ew results useful to model builders. Attempts are being made
01700	currently in artificial intelligence to write  algorithims  which  "understand"
01800	natural   language  expressions. [Enea and Colby,1973].
01900		During the 1960's when
02000	machine processing of natural language  was  dominated  by  syntactic
02100	considerations,  it  became  clear that syntactical information alone
02200	was  insufficient  to  comprehend   the   expressions   of   ordinary
02300	conversations. A current view is that to understand what is said in
02400	linguistic expressions, knowledge of syntax and semantics must  be  combined  with
02500	beliefs  from a  conceptual structure capable of making
02600	inferences. How to achieve this combination efficiently  with  a
02700	large  data-base  represents  a  monumental  task for both theory and
02800	implementation. 
02900		We did not attempt to construct a conventional linguistic parser
03000	to deal with natural language input for several reasons. Parsers to date
03100	have great difficulty in assigning a meaningful interpretation to the
03200	expressions of everyday conversational language using unrestricted English.
03300	A conventional parser may simply halt when it comes across a word not
03400	in its dictionary. Parsers represent tight conjunctions of tests
03500	instead of loose disjunctions needed for everyday language which
03600	may involve misunderstandinga and ununderstandings.
03700		The language analysis utilized by the model first 
03800	puts the input in the form of a list and then determines
03900	the syntactical type for the input expression- question, statement or
04000	imperative. The expression-type  is scanned in an
04100	attempt to form a conceptualization, i.e. a predication of an attribute, 
04200	on an object or a relation between objects. An attribute consists of
04300	something one is or does or possesses. The resultant conceptualization
04400	is then classified according to the rules of Fig. 00 in %000 as malevolent,
04500	benevolent or neutral.
04600		How language is  understood  depends  on  the  intentions  of  the  producers  and
04700	interpreters  in  the  dialogue. Thus  language  is   understood   in
04800	accordance with the participant's view of the situation. Our purpose was to develop a
04900	method for understanding everyday English sufficient for the model to
05000	communicate linguistically in  a  paranoid  way  in the  circumscribed
05100	situation of a psychiatric interview.
05200	We did not try to construct a general-purpose algorithm  which  could
05300	understand  anything  said  in  English  by anybody to anybody in any
05400	dialogue situation. (Does anyone believe it possible?)
05500		We took as  pragmatic measures of "understanding" the ability
05600	(1) to form a conceptualization so that questions can be answered and commands carried out,
05700	(2) to determine the intention of the interviewer, (3) to determine the 
05800	references for pronouns and other anticipated topics.  This
05900	straightforward  approach  to a complex problem has its drawbacks, as
06000	will be shown, but we strove for a highly individualized idiolect sufficient
06100	to  demonstrate  paranoid  processes of an individual in a particular
06200	situation rather than for a general supra-individual or ideal  comprehension
06300	of  English.  If the language-understanding system interfered  with
06400	demonstrating the paranoid processes, we would consider it  defective
06500	and  insufficient  for  our  purposes.             
06600		Some special problems  a dialogue algorithm must handle in a
06700	psychiatric interview will now  be outlined along with a brief description
06800	of how the model deals with them.
06900	
06950	.F
07000	QUESTIONS
07100	
07200		The  principal expression-type used by an interviewer consists
07300	of a question. A question is recognized by its beginning with a wh- or how
07400	form and/or the expression ending with a question-mark.
07500	In  teletyped  interviews a question may
07600	sometimes be put in declarative form followed by a question  mark  as in:
07650	.V
07700		(1) PT.- I LIKE TO GAMBLE ON THE HORSES.             	
07800		(2) DR.- YOU GAMBLE?
07850	.END
07900	Although the verb is missing in (2), the model recognizes that a question
08000	is being asked about its gambling.
08100	
08200	Particularly difficult are `when' questions which  require  a  memory
08300	which  can  assign  each  event a beginning, end and a duration. 
08400	An improved version of the model will have this capacity. Also
08500	troublesome are questions such as `how often',  `how  many',  i.e.  a
08600	`how' followed by a quantifier. If the model has "how often" on its
08700	expectancy list while a topic is under discussion, the appropriate reply
08800	can be made. Otherwise the model ununderstands.
08900		In constructing a simulation  of  a  thought  process  it  is
09000	arbitrary  how  much  information  to represent in memory. Should the
09100	model know what is the capital of Alabama? It is trivial to store a lot of facts. We took the position  that
09200	the  model  should  know  only what we believed it reasonable to know
09300	relevant to a few hundred topics expectable  in  a  psychiatric  interview.
09400	Thus  the  model  performs  badly  when  subjected  to baiting `exam'
09500	questions designed to test its informational limitations rather than to seek useful
09600	psychiatric information.
09650	
09675	.F
09700		IMPERATIVES
09800	
09900		Typical imperatives in a  psychiatric  interview  consist  of
10000	expressions like:
10050	.V
10100		(3) DR.- TELL ME ABOUT YOURSELF.
10200		(4)  DR.-  LETS  DISCUSS  YOUR  FAMILY. 
10250	.END
10300		Such imperatives are
10400	actually interrogatives to the interviewee about the topics they  refer  to.  Since
10500	the  only  physical  action  the  model  can  perform  is to `talk' ,
10600	imperatives  are  treated   as   requests   for   information.
10650	
10675	.F
10700	DECLARATIVES
10800	
10900		In  this  category  is lumped  everything else. It includes
11000	greetings, farewells, yes-no type answers, existence  assertions  and
11100	the usual predications. 
11200	
11250	.F
11300	AMBIGUITIES
11400	
11500		Words have more than  one  sense,  a  convenience  for  human
11600	memories  but  a struggle for language-analysing algorithms. Consider the
11700	word "bug" in the following expressions:
11750	.V
11800		(5) AM I BUGGING YOU?
11900		(6) AFTER A PERIOD OF HEAVY DRINKING HAVE YOU  FELT  BUGS  ON
12000	            YOUR SKIN?
12100		(7) DO YOU THINK THEY PUT A BUG IN YOUR ROOM? 
12150	.END
12200		In expression (5) the term
12300	"bug"  means  to  annoy,  in  (6) it refers to an insect and in (7) it
12400	refers to a microphone used for hidden survellience. Some  words  like
12500	"run" have fifty or more common senses. Context is used to carry
12600	out disambiguation. For example, when the Mafia is under discussion the
12700	model interprets "bug" to mean microphone. Thus we have the advantage
12800	of an idiolect where we can arbitrarily restrict the word senses. One
12900	characteristic of the paranoid mode is that no matter in  what  sense
13000	the interviewer  uses  a word, the  patient may  idiosyncratically
13100	interpret  it in some  sense relevant to his  pathological 
13200	beliefs.   		
13300	
13350	.F
13400	ANAPHORIC REFERENCES
13500		The common anaphoric references consist of the pronouns "it",
13600	"he", "him", "she", "her", "they", "them" as in:
13650	.V
13700		(8) PT.-HORSERACING IS MY HOBBY.
13800		(9) DR.-WHAT DO  YOU  ENJOY  ABOUT  IT?  
13850	.END
13900		The algorithm recognizes "it" as referring to "horseracing" 
14000	because "it" has been placed on the expectancy list when horseracing
14100	was introduced in (8). A more difficult problem arises when the anaphoric
14200	reference points more than one I/O pair back in the dialogue as in:
14250	.V
14300		(10) PT.-THE MAFIA IS OUT TO GET ME.
14400		(11) DR.- ARE YOU AFRAID OF THEM?
14500		(12) PT.- MAYBE.
14600		(13) DR.- WHY IS THAT? 
14650	.END
14700		The "that" of expression (13) does not refer to
14800	(12)  but  to  the  topic  of  being  afraid  which  the  interviewer
14900	introduced in (11). Another  pronominal  confusion  occurs  when  the
15000	interviewer uses `we' in two senses as in:
15050	.V
15100		(14) DR.- WE WANT YOU TO STAY IN THE HOSPITAL.
15200		(15) PT.- I WANT TO BE DISCHARGED NOW.
15300		(16) DR.- WE ARE NOT COMMUNICATING. 
15350	.END
15400		In expression (14) the interviewer
15500	is  using  "we" to refer to psychiatrists or the hospital staff while
15600	in (16) the term refers to the interviewer and patient. Identifying the
15700	correct referent would require beliefs about the dialogue which the
15800	new version of the model will have.
15900	
15950	.F
16000	TOPIC SHIFTS
16100	
16200		In the main a psychiatric interviewer is in control of the
16300	interview. When he has gained sufficient information about a topic,
16400	he shifts to a new topic. Naturally the algorithm must detect this
16500	change of topic as in the following:
16550	.V
16600		(17) DR.- HOW DO YOU LIKE THE HOSPITAL?
16700		(18) PT.- ITS NOT HELPING ME TO BE HERE.
16800		(19) DR.- WHAT BROUGHT YOU TO THE HOSPITAL?
16900		(20) PT.- I AM VERY UPSET AND NERVOUS.
17000		(21) DR.- WHAT TENDS TO MAKE YOU NERVOUS?
17100		(23) PT.- JUST BEING AROUND PEOPLE.
17200		(24) DR.- ANYONE IN PARTICULAR?
17250	.END
17300		In (17) and (19) the topic is the hospital. In (21) the
17400	topic changes to causes of the patient's nervous state.
17500		When a topic is introduced by the patient as in (20),
17600	a number of things can be expected to be asked about it. Thus 
17700	the algorithm has  ready an updated expectancy-anaphora list which 
17800	allows it to determine whether the topic
17900	introduced by the model is being responded to or whether the interviewer
18000	is continuing with the previous topic.
18100		Topics touched upon previously can be re-introduced
18200	at any point in the interview. The model knows that a topic has been 
18300	discussed previously because a topic-flag is set when a topic comes up.
18400	
18450	.F
18500	META-REFERENCES
18600	
18700		These are references, not about a topic directly, but about
18800	what has been said about the topic as in:
18850	.V
18900		(25) DR.- WHY ARE YOU IN THE HOSPITAL?
19000		(26) PT.- I SHOULDNT BE HERE.
19100		(27) DR.-  WHY DO YOU SAY THAT?
19150	.END
19200		The expression (27 ) is about  and meta to expression (26 ).
19300		Sometimes when the patient makes a statement, the doctor replies,
19400	not with a question, but with another statement which constitutes a
19500	rejoinder as in:
19550	.V
19600		(28 ) PT.- I HAVE LOST A LOT OF MONEY GAMBLING.
19700		(29 ) DR.- I GAMBLE QUITE A BIT ALSO.
19750	.END
19800		Here the algorithm interprets (29 ) as a directive to continue
19900	discussing gambling, not as an indication to question the doctor about
20000	gambling. The one exception to this principle occurs when the algorithm
20100	recognizes a chance to add to its model or representation of the interviewer.
20150	
20175	.F
20200	ELLIPSES
20300	
20400	
20500		In dialogues one finds many ellipses, expressions
20600	from which one or more words are omitted as in:
20650	.V
20700		(30 ) PT.- I SHOULDNT BE HERE.
20800		(31) DR.- WHY NOT?
20850	.END
20900		Here the complete construction must be understood as:
20950	.V
21000		(32) DR.- WHY SHOULD YOU NOT BE HERE?
21050	.END
21100	Again this is handled by the expectancy list which anticipates a "why not".
21200		The opposite of ellipsis is redundancy which usually provides no
21300	problem since the same thing is being said more than once as in:
21350	.V
21400		(33 ) DR.- LET ME ASK YOU A QUESTION.
21450	.END
21500		If an analysis were required of this expression (it is not
21600	required here since the expression is a sterotype), it would be recognized
21700	that the verb "ask" takes the noun "question" as direct object and
21800	also a question is something that is asked.
21900	
21950	.F
22000	SIGNALS
22100	
22200		Some fragmentary expressions serve only as directive  signals
22300	to proceed as in:
22350	.V
22400		(34) PT.- I WENT TO THE TRACK LAST WEEK.
22500		(35) DR.- AND?
22550	.END
22600	The fragment of (35) requests a continuation of the story
22700	introduced in (34). The common expressions found in interviews are
22800	"and", "so", "go on", "go ahead", "really", etc. If an input expression
22900	cannot be recognized at all, the lowest level default condition is
23000	to assume it is a signal and either proceed with the next line in a story under discussion
23100	or if the latter is not the case, begin a new story with a prompting
23200	question or statement.
23300	
23350	.F
23400	IDIOMS
23500	
23600		Since so much of conversational language is stereotyped, the task
23700	of recognition is much easier than that of analysis. 
23800	This is particularly true of idioms. Either one knows what an idiom means
23900	or one does not. It is usually hopeless to try to decipher what an
24000	idiom means from an analysis of its constituent parts. If the reader
24100	doubts this, let him ponder the following expressions taken from
24200	actual teletyped interviews.
24250	.V
24300		(36) DR.- WHATS EATING YOU?
24400		(37) DR.- YOU SOUND KIND OF PISSED OFF.
24500		(38) DR.- WHAT ARE YOU DRIVING AT?
24600		(39) DR.- ARE YOU PUTTING ME ON?
24700		(40) DR.- WHY ARE THEY AFTER YOU?
24800		(41) DR.- HOW DO YOU GET ALONG WITH THE OTHER PATIENTS?
24900	 	(42) DR.- HOW DO YOU LIKE YOUR WORK?
25000		(43) DR.- HAVE THEY TRIED TO GET EVEN WITH YOU?
25100		(44) DR.- I CANT KEEP UP WITH YOU.
25150	.END
25200		Understanding idioms is  a matter of rote memory. Hence
25300	an algorithm with a large idiom table is required. As each new idiom
25400	appears in teletyped interviews, it should be added to the idiom table
25500	because what happens once can happen again.
25600		One advantage in constructing an idiolect for a model is that
25700	it understands its own idiomatic expressions which tend to be used
25800	by the interviewer if he understands them as in:
25850	.V
25900		(45) PT.- THEY ARE OUT TO GET ME.
26000		(46) DR.- WHAT MAKES YOU THINK THEY ARE OUT TO GET YOU.
26050	.END
26100		The expression (45 ) is really a double idiom in which "out"
26200	means `intend' and "get" means `harm' in this context. Needless to say. 
26300	an algorithm which tried to pair off the various meanings of "out" with
26400	the various meanings of "get" would have a hard time of it. But an
26500	algorithm which understands what it itself is capable of saying, 
26600	can easily recognize echoed idioms.
26700	
26750	.F
26800	FUZZ TERMS
26900	
27000		In this category fall a large number of expressions which
27100	have little or no meaning and therefore can be ignored by the algorithm.
27200	The lower-case expressions in the following are examples of fuzz:
27250	.V
27300		(47) DR.- well now perhaps YOU CAN TELL ME something ABOUT YOUR FAMILY.
27400		(48) DR.- on the other hand I AM INTERESTED IN YOU.
27500		(49) DR.- hey I ASKED YOU A QUESTION.
27550	.END
27600		It is not the case that in order to ignore  something one must
27700	recognize explicitly what is ignorable. Since pattern-matching allows
27800	for an `anything' slot in many of its patterns, fuzz is thus easily ignored.
27900	
27950	.F
28000	SUBORDINATE CLAUSES
28100	
28200		A subordinate clause is a complete statement inside another statement.
28300	It is most frequently introduced by a relative pronoun, indicated in the
28400	following expressions by lower case:
28450	.V
28500		(50) DR.-  WAS IT THE UNDERWORLD that PUT YOU HERE?
28600		(51) DR.- WHO ARE THE PEOPLE who UPSET YOU?
28700		(52) DR.- HAS ANYTHING HAPPENED which YOU DONT UNDERSTAND?
28750	.END
28800		The words "whether" and "because" serving as conjunctions are less
28900	frequent. A language-analysis also must recognize that subordinate clauses
29000	can function as nouns, adjectives, adverbs, and objects of prepositions.
29100	
29150	.F
29200	VOCABULARY
29300	
29400		How many words should there be in the algorithm's vocabulary?
29500	It is a rare human speaker of English who can recognize 40% of the
29600	415,000 words in the Oxford English Dictionary. In his everyday
29700	conversation an educated person uses perhaps 10,000 words and has
29800	a recognition vocabulary of about 50,000 words. A study of phone
29900	conversations showed that 96 % of the talk employed only 737 words. Of
30000	course the remaining 4% , if not recognized, may be ruinous to the
30100	continuity of a conversation.
30200		In counting the words in 53 teletyped  psychiatric interviews,
30300	we found psychiatrists used only 721 words. Since we are familiar with
30400	psychiatric vocabularies and styles of expression, we believed this
30500	language-algorithm could function adequately with a vocabulary
30600	of at most a few thousand words. There will always be unrecognized words. The
30700	algorithm must be able to continue even if it does not have a particular word 
30800	in its vocabulary. This provision represents one great advantage of
30900	pattern-matching over conventional linguistic parsing.
31000		It is not the number of words which creates difficulties but their
31100	combinations. One thousand factorial is still a very large number. Syntactic
31200	and semantic constraints in stereotypes and in analysis reduce this
31300	number to an indefinitely large one.
31400	
31450	.F
31500	MISSPELLINGS AND EXTRA CHARACTERS
31800		There is really no good defense  against misspellings
31900	in a teletyped interview except having a human monitor retype the correct
32000	versions. Spelling correcting programs are slow, inefficient, and imperfect.
32100	They experience great problems when it is the first character in a word
32200	which is incorrect.
32300		Extra characters sent by the interviewer or by a bad phone
32400	line can be removed by a human monitor.
32500	
32550	.F
32600	META VERBS
32700	
32800		Certain common verbs such as "think", "feel", "believe", etc
32900	take as their objects a clause as in:
32950	.V
33000		(54) DR.- I THINK YOU ARE RIGHT.
33100		(55) DR.- WHY DO YOU FEEL THE GAMBLING IS CROOKED?
33150	.END
33200		The verb "believe" is peculiar since it can also take as
33300	object a noun or noun phrase as in:
33350	.V
33400		(56) DR.- I BELIEVE YOU.
33450	.END
33500		In expression (55) the conjunction "that" can follow
33600	the word "feel" signifying a subordinate clause. This is not the case
33700	after "believe" in expression (56).
33800	
33850	.F
33900	ODD WORDS
34000	
34100		These are words which are odd in the context of a 
34200	teletyped interview while they are quite natural in the usual vis-a-vis
34300	interview in which the participants communicate through speech. This
34400	should be clear from the following examples in which the odd words
34500	appear in lower case:
34550	.V
34600		(57) DR.-YOU sound CONFUSED.
34700		(58) DR.- DID YOU hear MY LAST QUESTION?
34800		(59) DR.- WOULD YOU come in AND sit down PLEASE?
34900		(60) DR.- CAN YOU say WHO?
35000		(61) DR.- I WILL see YOU AGAIN TOMORROW.
35050	.END
35100	
35200	
35250	.F
35300	MISUNDERSTANDING
35400	
35500		It is not fully recognized bt students of language how often people
35600	misunderstand one another in conversation and yet their
35700	dialogues proceed as if understanding and being understood had taken
35800	place.
35900		The classic story involves three partially deaf men cycling
36000	through the English counrtyside:
36050	.V
36100		FIRST - "WHAT TOWN IS THIS?"
36200		SECOND - "THURSDAY"
36300		THIRD - "ME TOO, LETS STOP AND HAVE A DRINK."
36350	.END
36400		Sometimes a psychiatric interviewer realizes when misunderstanding
36500	occurs and tries to correct it. Other times he simply passes it by. It is
36600	characteristic of the paranoid mode to respond idiosyncratically to
36700	particular word-concepts regardless of what the interviewer is saying:
36750	.V
36800		(62) PT.- IT IS NOT HELPING ME TO BE HERE.  	
36810	        (63) DR.- I BET.
36820	        (64) PT.- GAMBLING HAS BEEN NOTHING BUT TROUBLE FOR ME.
36860	.END
36900	
36950	.F
37000	UNUNDERSTANDING
37050	
37100		A dialogue algorithm  must be prepared for situations
37200	in which it simply does not understand i.e. it cannot arrive at any
37300	interpretation as to what the interviewer is saying. An algorithm should
37400	not be faulted for a lack of facts as in:
37450	.V
37500		(65) DR.- WHO IS THE PRESIDENT OF TURKEY?
37550	.END CONTINUE
37600	wherin the memory does not contain the words "president" and "Turkey".
37700	In this default condition it is simplest to reply:
37750	.V
37800		(66) PT.- I DONT KNOW.
37850	.END CONTINUE
37900	and dangerous to reply:
37950	.V
38000		(67) PT.- COULD YOU REPHRASE THE QUESTION?
38050	.END CONTINUE
38100	because of the horrible loops which can result.
38200		Since the main problem in the default condition of ununderstanding
38300	is how to continue, heuristics can be employed such as asking about the 
38400	interviewer's intention as in:
38450	.V
38500		(68) PT.- WHY DO YOU WANT TO KNOW THAT?
38550	.END CONTINUE
38600	or rigidly continuing with a previous topic or introducing a new topic.
38700		These are admittedly desperate measures intended to prompt
38800	the interviewer in directions the algorithm has a better chance of understanding.
38900	Usually it is the interviewer who controls the flow from topic to 
39000	topic but there are times, hopefully few, when control must be assumed
39100	by the algorithm.