00100	    		CHAPTER FOUR
00200	 SPECIAL PROBLEMS FOR COMPUTER UNDERSTANDING OF NATURAL LANGUAGE 
00300		 IN TELETYPED PSYCHIATRIC INTERVIEWS
00400	
00500	
00600	By `natural language` I shall mean everyday American English such  as
00700	is  used  by  readers  of  this book in ordinary conversations. It is
00800	still difficult to be  explicit  about  the  processes  which  enable
00900	hummans  to  interpret and respond to natural language. Philosophers,
01000	linguists and psychologists have investigated natural  language  with
01100	various purposes and few useful results.  Now attempts are being made
01200	in artificial intelligence to write  algorithims  which  `understand'
01300	natural   language  expressions.     
01400		During the 1960's when
01500	machine processing of natural language  was  dominated  by  syntactic
01600	considerations,  it  became  clear that syntactical information alone
01700	was  insufficient  to  comprehend   the   expressions   of   ordinary
01800	conversations. The current view is that to understand what is said in
01900	linguistic expressions, syntax and semantics must  be  combined  with
02000	beliefs  from an underlying conceptual structure having an ability to
02100	draw inferences. How to achieve this combination efficiently  with  a
02200	large  data-base  represents  a  monumental  task for both theory and
02300	implementation.  
02400		Since the behavior being simulated by our
02500	paranoid  model  is  the language-behavior of a paranoid patient in a
02600	psychiatric interview, the model must have an  ability  to  interpret
02700	and respond to natural language input  sufficient only to demonstrate
02800	language-behavior characteristic of the paranoid mode.  How  language
02900	is  understood  depends  on  the  intentions  of  the  producers  and
03000	interpreters  in  the  dialogue. Thus  language  is   understood   in
03100	accordance with the participant's view of the game being played. Our purpose was to develop a
03200	method for understanding everyday English sufficient for the model to
03300	communicate linguistically in  a  paranoid  way  in the  circumscribed
03400	situation of a psychiatric interview.
03500	We did not try to construct a general-purpose algorithm  which  could
03600	understand  anything  said  in  English  by anybody to anybody in any
03700	dialogue situation. (Does anyone believe it possible?)
03800		We took as a pragmatic measure of "understanding" the ability
03900	of the algorithm to `get the message' of an expression by trying to classify
04000	the imperative or directive intent of the interviewer,i.e.what effect he is
04100	trying to bring about in the interviewee relative to the topic.  This
04200	straightforward  approach  to a complex problem has its drawbacks, as
04300	will be shown, but we strove for a highly individualized idiolect sufficient
04400	to  demonstrate  paranoid  processes of an individual in a particular
04500	situation rather than for a general supra-individual or ideal  comprehension
04600	of  English.  If the language-understanding algorithm interfered  with
04700	demonstrating the paranoid processes, we would consider it  defective
04800	and  insufficient  for  our  purposes.             (Insert from Machr
04900	here)
05000		Some special problems  a dialogue algorithm must cope with in a
05100	psychiatric      interview      will      now      be      discussed.
05200	
05300	QUESTIONS
05400	
05500		The  principal  sentence-type used by an interviewer consists
05600	of a question. The usual wh- and yes-no questions must be  recognized
05700	by  the  language-algorithm. In  teletyped  interviews a question may
05800	sometimes be put in declarative form followed by a question  mark  as in:
05900		(1) PT.- I LIKE TO GAMBLE ON THE HORSES.             	
06000		   DR.- YOU GAMBLE?
06100	
06200	Particularly difficult are `when' questions which  require  a  memory
06300	which  can  assign  each  event a beginning, end and a duration. Also
06400	troublesome are questions such as `how often',  `how  many',  i.e.  a
06500	`how' followed by a quantifier.
06600		In constructing a simulation  of  a  thought  process  it  is
06700	arbitrary  how  much  information  to represent in memory. Should the
06800	model know what is the capital of Alabama? It is trivial to store a lot of facts. We took the position  that
06900	the  model  should  know  only what we believed it reasonable to know
07000	relevant to a few hundred topics expectable  in  a  psychiatric  interview.
07100	Thus  the  model  performs  badly  when  subjected  to baiting `exam'
07200	questions designed to test its informational limitations rather than to seek useful
07300	psychiatric information.
07400		IMPERATIVES
07500	
07600		Typical imperatives in a  psychiatric  interview  consist  of
07700	expressions like:
07800		(2) DR.- TELL ME ABOUT YOURSELF.
07900		(3)  DR.-  LETS  DISCUSS  YOUR  FAMILY. 
08000		Such imperatives are
08100	actually interrogatives to the interviewee about the topics they  refer  to.  Since
08200	the  only  physical  action  the  model  can  perform  is to `talk' ,
08300	imperatives  should  be  treated   as   requests   for   information.
08400	DECLARATIVES
08500	
08600		In  this  category  we lump everything else. It includes
08700	greetings, farewells, yes-no type answers, existence  assertions  and
08800	predications made upon a subject. 
08900	
09000	AMBIGUITIES
09100	
09200		Words have more than  one  sense,  a  convenience  for  human
09300	memories  but  a struggle for language-analysing algorithms. Consider the
09400	word `bug' in the following expressions:
09500		(4) AM I BUGGING YOU?
09600		(5) AFTER A PERIOD OF HEAVY DRINKING HAVE YOU  FELT  BUGS  ON
09700	YOUR SKIN?
09800		(6) DO YOU THINK THEY PUT A BUG IN YOUR ROOM? 
09900		In expression (4) the term
10000	`bug'  means  to  annoy,  in  (5) it refers to an insect and in (6) it
10100	refers to a microphone used for hidden survellience. Some common words  like
10200	`run' have fifty or more common senses. Context must be used to carry
10300	out disambiguation, as described in 00.0. Also we have the  advantage
10400	of an idiolect where we can arbitrarily restrict the word senses. One
10500	characteristic of the paranoid mode is that no matter in  what  sense
10600	the interviewer  uses  a word, the  patient may  idiosyncratically
10700	interpret  it in some  sense relevant to his  pathological  malevolence
10800	beliefs.   		
10900	
11000	ANAPHORIC REFERENCES
11100		The common anaphoric references consist of the pronouns `it',
11200	`he', `him', `she', `her', `they', `them' as in:
11300		(7) PT.-HORSERACING IS MY HOBBY.
11400		(8) DR.-WHAT DO  YOU  ENJOY  ABOUT  IT?  
11500		The algorithm must
11600	recognize  that the 'it' refers to `horseracing'. More difficult is a
11700	reference more than one I/O pair back in the dialogue as in:
11800		(9) PT.-THE MAFIA IS OUT TO GET ME.
11900		(10) DR.- ARE YOU AFRAID OF THEM?
12000		(11) PT.- MAYBE.
12100		(12) DR.- WHY IS THAT? 
12200		The "that" of expression (12) does not refer to
12300	(11)  but  to  the  topic  of  being  afraid  which  the  interviewer
12400	introduced in (10). Another  pronominal  confusion  occurs  when  the
12500	interviewer uses `we' in two senses as in:
12600		(13) DR.- WE WANT YOU TO STAY IN THE HOSPITAL.
12700		(14) PT.- I WANT TO BE DISCHARGED NOW.
12800		(15) DR.- WE ARE NOT COMMUNICATING. 
12900		In expression (13) the interviewer
13000	is  using  "we" to refer to psychiatrists or the hospital staff while
13100	in (15) the term refers to the interviewer and patient.
13200	
13300	TOPIC SHIFTS
13400	
13500		In the main a psychiatric interviewer is in control of the
13600	interview. When he has gained sufficient information about a topic,
13700	he shifts to a new topic. Naturally the algorithm must detect this
13800	change of topic as in the following:
13900		(16) DR.- HOW DO YOU LIKE THE HOSPITAL?
14000		(17) PT.- ITS NOT HELPING ME TO BE HERE.
14100		(18) DR.- WHAT BROUGHT YOU TO THE HOSPITAL?
14200		(19) PT.- I AM VERY UPSET AND NERVOUS.
14300		(20) DR.- WHAT TENDS TO MAKE YOU NERVOUS?
14400		(22) PT.- JUST BEING AROUND PEOPLE.
14500		(23) DR.- ANYONE IN PARTICULAR?
14600		In (16) and (18) the topic is the hospital. In (20) the
14700	topic changes to causes of the patient's nervous state.
14800		When a topic is introduced by the patient as in (19),
14900	a number of things can be expected to be asked about it. Thus 
15000	the algorithm can have ready an expectancy-anaphora list which 
15100	allows it to determine whether the topic
15200	introduced by the model is being responded to or whether the interviewer
15300	is continuing with the previous topic.
15400		Topics touched upon previously can be re-introduced
15500	at any point in the interview. The memory of the model is responsible
15600	for knowing what has been discussed.
15700	
15800	META-REFERENCES
15900	
16000		These are references, not about a topic directly, but about
16100	what has been said about the topic as in:
16200		(24) DR.- WHY ARE YOU IN THE HOSPITAL?
16300		(25) PT.- I SHOULDNT BE HERE.
16400		(26)DR.-  WHY DO YOU SAY THAT?
16500		The expression (26 ) is about  and meta to expression (25 ).
16600		Sometimes when the patient makes a statement, the doctor replies,
16700	not with a question, but with another statement which constitutes a
16800	rejoinder as in:
16900		(27 ) PT.- I HAVE LOST A LOT OF MONEY GAMBLING.
17000		(28 ) DR.- I GAMBLE QUITE A BIT ALSO.
17100		Here the algorithm should interpret (28 ) as a directive to continue
17200	discussing gambling, not as an indication to question the doctor about
17300	gambling. The one exception to this principle occurs when the algorithm
17400	recognizes a chance to add to its model or representation of the interviewer.
17500	ELLIPSES
17600	
17700	
17800		In dialogues one finds many ellipses, expressions
17900	from which one or more words are omitted as in:
18000		(29 ) PT.- I SHOULDNT BE HERE.
18100		(30) DR.- WHY NOT?
18200		Here the complete construction must be understood as:
18300		(31) DR.- WHY SHOULD YOU NOT BE HERE?
18400		By saving the previous surface expression and the belief it mapped
18500	into in memory, the algorithm can recognize either what the missing words
18600	are or the concepts they refer to.
18700		The opposite of ellipsis is redundancy which usually provides no
18800	problem since the same thing is being said more than once as in:
18900		(32 ) DR.- LET ME ASK YOU A QUESTION.
19000		If an analysis were required of this expression (it is not
19100	required here since the expression is a sterotype), it would be recognized
19200	that the verb "ask" takes the noun "question" as direct object and
19300	also a question is something that is asked.
19400	
19500	SIGNALS
19600	
19700		Some fragmentary expressions serve only as directive  signals
19800	to proceed as in:
19900		(33) PT.- I WENT TO THE TRACK LAST WEEK.
20000		(34) DR.- AND?
20100	The fragment of (34) requests a continuation of the story
20200	introduced in (33). The common expressions found in interviews are
20300	"and", "so", "go on", "go ahead", "really", etc. If an input expression
20400	cannot be recognized at all, the lowest level fedault condition is
20500	to assume it is a signal and either proceed with the next line in a story under discussion
20600	or if the latter is not the case, begin a new story with a prompting
20700	question or statement.
20800		This strategy can fail as in:
20900		(FIND GOOD EXAMPLE)
21000	
21100	IDIOMS
21200	
21300		Since so much of conversational language is stereotyped, the task
21400	of recognition is much easier than that of analysis. 
21500	This is particularly true of idioms. Either one knows what an idiom means
21600	or one does not. It is usually hopeless to try to decipher what an
21700	idiom means from an analysis of its constituent parts. If the reader
21800	doubts this, let him ponder the following expressions taken from
21900	actual teletyped interviews.
22000		(35) DR.- WHATS EATING YOU?
22100		(36) DR.- YOU SOUND KIND OF PISSED OFF.
22200		(37) DR.- WHAT ARE YOU DRIVING AT?
22300		(38) DR.- ARE YOU PUTTING ME ON?
22400		(39) DR.- WHY ARE THEY AFTER YOU?
22500		(40) DR.- HOW DO YOU GET ALONG WITH THE OTHER PATIENTS?
22600	 	(41) DR.- HOW DO YOU LIKE YOUR WORK?
22700		(42) DR.- HAVE THEY TRIED TO GET EVEN WITH YOU?
22800		(43) DR.- I CANT KEEP UP WITH YOU.
22900		Understanding idioms is  a matter of rote memory. Hence
23000	an algorithm with a large idiom table is required. As each new idiom
23100	appears in teletyped interviews, it should be added to the idiom table
23200	because what happens once can happen again.
23300		One advantage in constructing an idiolect for a model is that
23400	it understands its own idiomatic expressions which tend to be used
23500	by the interviewer if he understands them as in:
23600		(44) PT.- THEY ARE OUT TO GET ME.
23700		(45) DR.- WHAT MAKES YOU THINK THEY ARE OUT TO GET YOU.
23800		The expression (45 ) is really a double idiom in which "out"
23900	means `intend' and "get" means `harm' in this context. Needless to say. 
24000	an algorithm which tried to pair off the various meanings of "out" with
24100	the various meanings of "get" would have a hard time of it. But an
24200	algorithm which understands what it itself is capable of saying, should
24300	be able to recognize echoed idioms.
24400	
24500	FUZZ TERMS
24600	
24700		In this category we group a large number of expressions which
24800	have little or no meaning and therefore can be ignored by the algorithm.
24900	The lower-case expressions in the following are examples of fuzz:
25000		(46) DR.- well now perhaps YOU CAN TELL ME something ABOUT YOUR FAMILY.
25100		(47) DR.- on the other hand I AM INTERESTED IN YOU.
25200		(48) DR.- hey I ASKED YOU A QUESTION.
25300		It is not the case that in order to ignore  something one must
25400	recognize explicitly what is ignorable. Since pattern-matching allows
25500	for an `anything' slot in many of its patterns, fuzz is thus easily ignored.
25600	
25700	SUBORDINATE CLAUSES
25800	
25900		A subordinate clause is a complete statement inside another statement.
26000	It is most frequently introduced by a relative pronoun, indicated in the
26100	following expressions by lower case:
26200		(49) DR.-  WAS IT THE UNDERWORLD that PUT YOU HERE?
26300		(50) DR.- WHO ARE THE PEOPLE who UPSET YOU?
26400		(51) DR.- HAS ANYTHING HAPPENED which YOU DONT UNDERSTAND?
26500		The words "whether" and "because" serving as conjunctions are less
26600	frequent. A language-algorithm also must recognize that subordinate clauses
26700	can function as nouns, adjectives, adverbs, and objects of prepositions.
26800	
26900	VOCABULARY
27000	
27100		How many words should there be in the algorithm's vocabulary?
27200	It is a rare human speaker of English who can recognize 40% of the
27300	415,000 words in the Oxford English Dictionary. In his everyday
27400	conversation an educated person uses perhaps 10,000 words and has
27500	a recognition vocabulary of about 50,000 words. A study of phone
27600	conversations showed that 96 % of the talk employed only 737 words. Of
27700	course the remaining 4% , if not recognized, may be ruinous to the
27800	continuity of a conversation.
27900		In counting the words in 53 teletyped  psychiatric interviews,
28000	we found psychiatrists used only 721 words. Since we are familiar with
28100	psychiatric vocabularies and styles of expression, we believed this
28200	language-algorithm could function adequately with a vocabulary
28300	of a few thousand words. There will always be unrecognized words. The
28400	algorithm must be able to continue even if it does not have a particular word 
28500	in its vocabulary. This provision represents one great advantage of
28600	pattern-matching over conventional linguistic parsing.
28700		It is not the number of words which creates difficulties but their
28800	combinations. One thousand factorial is still a very large number. Syntactic
28900	and semantic constraints in stereotypes and in analysis reduce this
29000	number to an indefinitely large one.
29100	
29200	MISSPELLINGS AND EXTRA CHARACTERS
29300	
29400	
29500		There is really no good defense  against misspellings
29600	in a teletyped interview except having a human monitor retype the correct
29700	versions. Spelling correcting programs are slow, inefficient, and imperfect.
29800	They experience great problems when it is the first character in a word
29900	which is incorrect.
30000		Extra characters sent by the interviewer or by a bad phone
30100	line can be removed by a human monitor.
30200	
30300	META VERBS
30400	
30500		Certain common verbs such as "think", "feel", "believe", etc
30600	take as their objects a clause as in:
30700		(53) DR.- I THINK YOU ARE RIGHT.
30800		(54) DR.- WHY DO YOU FEEL THE GAMBLING IS CROOKED?
30900		The verb "believe" is peculiar since it can also take as
31000	object a noun or noun phrase as in:
31100		(55) DR.- I BELIEVE YOU.
31200		In expression (54) the conjunction "that" can follow
31300	the word "feel" signifying a subordinate clause. This is not the case
31400	after "believe" in expression (55).
31500	
31600	ODD WORDS
31700	
31800		These are words which are odd in the context of a 
31900	teletyped interview while they are quite natural in the usual vis-a-vis
32000	interview in which the participants communicate through speech. This
32100	should be clear from the following examples in which the odd words
32200	appear in lower case:
32300		(56) DR.-YOU sound CONFUSED.
32400		(57) DR.- DID YOU hear MY LAST QUESTION?
32500		(58) DR.- WOULD YOU come in AND sit down PLEASE?
32600		(59) DR.- CAN YOU say WHO?
32700		(60) DR.- I WILL see YOU AGAIN TOMORROW.
32800	
32900	
33000	MISUNDERSTANDING
33100	
33200		It is not fully recognized bt students of language how often people
33300	misunderstand one another in conversation and yet their
33400	dialogues proceed as if understanding and being understood had taken
33500	place.
33600		The classic story involves three partially deaf men cycling
33700	through the English counrtyside:
33800		FIRST - "WHAT TOWN IS THIS?"
33900		SECOND - "THURSDAY"
34000		THIRD - "ME TOO, LETS STOP AND HAVE A DRINK."
34100		Sometimes a psychiatric interviewer realizes when misunderstanding
34200	occurs and tries to correct it. Other times he simply passes it by. It is
34300	characteristic of the paranoid mode to respond idiosyncratically to
34400	particular word-concepts regardless of what the interviewer is saying:
34500		(FIND GOOD EXAMPLE)
34600	
34700	UNUNDERSTANDING
34800		A dialogue algorithm  must be prepared for situations
34900	in which it simply does not understand i.e. it cannot arrive at any
35000	interpretation as to what the interviewer is saying. An algorithm should
35100	not be faulted for a lack of facts as in:
35200		(61) DR.- WHO IS THE PRESIDENT OF TURKEY?
35300	wherin the memory does not contain the words "president" and "Turkey".
35400	In this default condition it is simplest to reply:
35500		(62) PT.- I DONT KNOW.
35600	and dangerous to reply:
35700		(63) PT.- COULD YOU REPHRASE THE QUESTION?
35800	because of the horrible loops which can result.
35900		Since the main problem in the default condition of ununderstanding
36000	is how to continue, heuristics can be employed such as asking about the 
36100	interviewer's intention as in:
36200		(64) PT.- WHY DO YOU WANT TO KNOW THAT?
36300	or rigidly continuing with a previous topic or introducing a new topic.
36400		These are admittedly desperate measures intended to prompt
36500	the interviewer in directions the algorithm has a better chance of understanding.
36600	Usually it is the interviewer who controls the flow from topic to 
36700	topic but there are times, hopefully few, when control must be assumed
36800	by the algorithm.