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