perm filename CHAP4[4,KMC]19 blob sn#050188 filedate 1973-06-22 generic text, type T, neo UTF8
00100	.SEC PROBLEMS FOR COMPUTER UNDERSTANDING OF NATURAL LANGUAGE
00200	COMMUNICATION IN TELETYPED PSYCHIATRIC INTERVIEWS
00300	
00400		Since the behavior being simulated by this paranoid model  is
00500	the   sequential   language-behavior  of  a  paranoid  patient  in  a
00600	psychiatric interview, the model must have an  ability  to  interpret
00700	and  respond  to  natural  language  input  sufficient to demonstrate
00800	conduct characteristic of the paranoid mode.   By "natural  language"
00900	I  shall  mean  ordinary American English such as is used in everyday
01000	conversations. It  is  still  difficult  to  be  explicit  about  the
01100	processes  which  enable  humans  to interpret and respond to natural
01200	language. (A mighty maze !  but not  without  a  plan  -  A.   Pope).
01300	Philosophers,  linguists  and psychologists have investigated natural
01400	language with various purposes and yielding  few  results  useful  to
01500	model  builders.   Attempts  currently  are  being made in artificial
01600	intelligence to write algorithims which "understand" natural language
01700	expressions.    [Enea   and   Colby,1973;   Winograd,1973].   Machine
01800	understanding of natural language is being attempted today but it  is
01900	not  something  to  be completly achieved today or even tomorrow. The
02000	problem at the moment is not to find  immediately  the  best  way  of
02100	doing it but to find any way at all.
02200		During the 1960's when machine processing of natural language
02300	was dominated by  syntactic  considerations,  it  became  clear  that
02400	syntactical  information  alone  was  insufficient  to comprehend the
02500	expressions of ordinary conversations. A  current  view  is  that  to
02600	understand  what  is  said  in  linguistic  expressions, knowledge of
02700	syntax and semantics must be combined with beliefs from a  conceptual
02800	structure   capable   of  making  inferences.  How  to  achieve  this
02900	combination  efficiently  with  a  large   data-base   represents   a
03000	monumental task for both theory and implementation.
03100		We did not attempt to  construct  a  conventional  linguistic
03200	parser  to  analyze  interview  natural  language input for practical
03300	reasons. Parsers  to  date  have  great  difficulty  in  assigning  a
03400	meaningful    interpretation   to   the   expressions   of   everyday
03500	conversational language using unrestricted English.  Purely syntactic
03600	parsers   offer  a  cancerous  proliferation  of  interpretations.  A
03700	conventional parser lacking ignoring mechanisms,may simply halt  when
03800	it comes across a word not in its dictionary. Parsers represent tight
03900	conjunctions of  tests  instead  of  loose  disjunctions  needed  for
04000	gleaning  a  meaninng  from everyday language communication which may
04100	involve misunderstandinga and ununderstandings.  People misunderstand
04200	and  ununderstand  at  times  and  thus  are  partially opaque to one
04300	another.
04400		The  language recognition process utilized by the model first
04500	puts the input in  the  form  of  a  list  and  then  determines  the
04600	syntactical  type  for  the  input expression- question, statement or
04700	imperative. The expression-type is scanned in an attempt  to  form  a
04800	conceptualization,  i.e.  a predication of an attribute, on an object
04900	or a relation between objects. An attribute consists of something one
05000	is  or  does  or possesses.   The resultant conceptualization is then
05100	classified according to the rules of Fig. 00 in %000  as  malevolent,
05200	benevolent or neutral.
05300		How language is understood depends on the intentions  of  the
05400	producers  and  interpreters  in  the  dialogue.   Thus  language  is
05500	understood  in  accordance  with  the  participant's  view   of   the
05600	situation.   Our  purpose  was  to develop a method for understanding
05700	sequences of everyday English sufficient for the model to communicate
05800	linguistically  in a paranoid way in the circumscribed situation of a
05900	psychiatric interview. Such an interview is not small talk; a job  is
06000	to be done.
06100		We did not try to construct a general-purpose algorithm which
06200	could  understand  anything  said in English by anybody to anybody in
06300	any dialogue situation. (Does anyone believe it possible?) We  sought
06400	only  to  extract, distill or cull an idiolectic meaning or gist of a
06500	meaning from the input.
06600		Natural language is not an agree-on universe of discourse
06700	such as arithmetic wherein symbols have the same meaning for everyone
06800	who uses them. What we loosely call "natural language" is actually a set
06900	of idiolects, each being unique to the individual with a unique history.
07000	To be unique does not mean that no property is shared with other
07100	individuals, only that not every property is shared. An overlap of
07200	idiolects allows the communication of shared meanings.
07300		We took as pragmatic measures of "understanding" the
07400	ability  (1)  to  form  a  conceptualization so that questions can be
07500	answered and commands carried out, (2) to determine the intention  of
07600	the  interviewer,  (3)  to  determine the references for pronouns and
07700	other anticipated topics.  This straightforward approach to a complex
07800	problem  has  its  drawbacks,  as  will be shown, but we strove for a
07900	highly individualized idiolect  sufficient  to  demonstrate  paranoid
08000	processes  of an individual in a particular situation rather than for
08100	a general supra-individual or ideal comprehension of English.  If the
08200	language-understanding  system  interfered  with  demonstrating   the
08300	paranoid  processes,  we would consider it defective and insufficient
08400	for our purposes.
08500		Some  special  problems a dialogue algorithm must handle in a
08600	psychiatric interview  will  now  be  outlined  along  with  a  brief
08700	description of how the model deals with them.
08800	
08900	.F
09000	QUESTIONS
09100	
09200		The principal expression-type used by an interviewer consists
09300	of  a  question. A question is recognized by its beginning with a wh-
09400	or how form and/or the expression ending  with  a  question-mark.  In
09500	teletyped  interviews  a question may sometimes be put in declarative
09600	form followed by a question mark as in:
09700	.V
09800		(1) PT.- I LIKE TO GAMBLE ON THE HORSES.             	
09900		(2) DR.- YOU GAMBLE?
10000	.END
10100	Although  the  verb  is  missing  in (2), the model recognizes that a
10200	question is being asked about its gambling.
10300	
10400	Particularly  difficult  are  `when' questions which require a memory
10500	which can assign each event a beginning,  end  and  a  duration.   An
10600	improved   version  of  the  model  will  have  this  capacity.  Also
10700	troublesome are questions such as `how often', `how  many',  i.e.   a
10800	`how'  followed  by a quantifier. If the model has "how often" on its
10900	expectancy list while a topic is under  discussion,  the  appropriate
11000	reply can be made. Otherwise the model ununderstands.
11100		In constructing a simulation  of  a  thought  process  it  is
11200	arbitrary  how  much  information  to represent in memory. Should the
11300	model know what is the capital of Alabama? It is trivial to  store  a
11400	lot  of  facts.  We took the position that the model should know only
11500	what we believed it reasonable to know  relevant  to  a  few  hundred
11600	topics expectable in a psychiatric interview. Thus the model performs
11700	badly when subjected to baiting `exam' questions designed to test its
11800	informational  limitations  rather  than  to  seek useful psychiatric
11900	information.
12000	
12100	.F
12200		IMPERATIVES
12300	
12400		Typical imperatives in a  psychiatric  interview  consist  of
12500	expressions like:
12600	.V
12700		(3) DR.- TELL ME ABOUT YOURSELF.
12800		(4)  DR.-  LETS  DISCUSS  YOUR  FAMILY. 
12900	.END
13000		Such   imperatives   are   actually   interrogatives  to  the
13100	interviewee about the topics they refer to.  Since the only  physical
13200	action  the  model can perform is to `talk' , imperatives are treated
13300	as requests for information.
13400	
13500	.F
13600	DECLARATIVES
13700	
13800		In  this  category  is  lumped  everything  else. It includes
13900	greetings, farewells, yes-no type answers, existence  assertions  and
14000	the usual predications.
14100	
14200	.F
14300	AMBIGUITIES
14400	
14500		Words have more than  one  sense,  a  convenience  for  human
14600	memories  but  a struggle for language-analysing algorithms. Consider
14700	the word "bug" in the following expressions:
14800	.V
14900		(5) AM I BUGGING YOU?
15000		(6) AFTER A PERIOD OF HEAVY DRINKING HAVE YOU  FELT  BUGS  ON
15100	            YOUR SKIN?
15200		(7) DO YOU THINK THEY PUT A BUG IN YOUR ROOM? 
15300	.END
15400		In  expression  (5)  the term "bug" means to annoy, in (6) it
15500	refers to an insect and in (7) it refers to  a  microphone  used  for
15600	hidden  survellience. Some words like "run" have fifty or more common
15700	senses. Context is used to carry  out  disambiguation.  For  example,
15800	when the Mafia is under discussion the model interprets "bug" to mean
15900	microphone. Thus we have the advantage of an idiolect  where  we  can
16000	arbitrarily  restrict  the  word  senses.  One  characteristic of the
16100	paranoid mode is that no matter in what sense the interviewer uses  a
16200	word,  the  patient  may idiosyncratically interpret it in some sense
16300	
16400	.F
16500	ANAPHORIC REFERENCES
16600		The common anaphoric references consist of the pronouns "it",
16700	"he", "him", "she", "her", "they", "them" as in:
16800	.V
16900		(8) PT.-HORSERACING IS MY HOBBY.
17000		(9) DR.-WHAT DO  YOU  ENJOY  ABOUT  IT?  
17100	.END
17200		The algorithm recognizes "it" as referring  to  "horseracing"
17300	because  "it" has been placed on the expectancy list when horseracing
17400	was introduced in (8). A  more  difficult  problem  arises  when  the
17500	anaphoric  reference  points  more  than  one  I/O  pair  back in the
17600	dialogue as in:
17700	.V
17800		(10) PT.-THE MAFIA IS OUT TO GET ME.
17900		(11) DR.- ARE YOU AFRAID OF THEM?
18000		(12) PT.- MAYBE.
18100		(13) DR.- WHY IS THAT? 
18200	.END
18300		The "that" of expression (13) does not refer to (12)  but  to
18400	the  topic  of being afraid which the interviewer introduced in (11).
18500	Another pronominal confusion occurs when the interviewer uses `we' in
18600	two senses as in:
18700	.V
18800		(14) DR.- WE WANT YOU TO STAY IN THE HOSPITAL.
18900		(15) PT.- I WANT TO BE DISCHARGED NOW.
19000		(16) DR.- WE ARE NOT COMMUNICATING. 
19100	.END
19200		In expression (14) the interviewer is using "we" to refer  to
19300	psychiatrists  or the hospital staff while in (16) the term refers to
19400	the interviewer and patient. Identifying the correct  referent  would
19500	require beliefs about the dialogue which the new version of the model
19600	will have.
19700	
19800	.F
19900	TOPIC SHIFTS
20000	
20100		In the main a psychiatric interviewer is in control of the
20200	interview. When he has gained sufficient information about a topic,
20300	he shifts to a new topic. Naturally the algorithm must detect this
20400	change of topic as in the following:
20500	.V
20600		(17) DR.- HOW DO YOU LIKE THE HOSPITAL?
20700		(18) PT.- ITS NOT HELPING ME TO BE HERE.
20800		(19) DR.- WHAT BROUGHT YOU TO THE HOSPITAL?
20900		(20) PT.- I AM VERY UPSET AND NERVOUS.
21000		(21) DR.- WHAT TENDS TO MAKE YOU NERVOUS?
21100		(23) PT.- JUST BEING AROUND PEOPLE.
21200		(24) DR.- ANYONE IN PARTICULAR?
21300	.END
21400		In (17) and (19) the topic is the hospital. In (21) the topic
21500	changes to causes of the patient's nervous state.
21600		When a topic is introduced by  the  patient  as  in  (20),  a
21700	number  of  things  can  be  expected  to be asked about it. Thus the
21800	algorithm has ready an updated expectancy-anaphora list which  allows
21900	it  to  determine  whether the topic introduced by the model is being
22000	responded to or  whether  the  interviewer  is  continuing  with  the
22100	previous topic.
22200		Topics touched upon previously can be  re-introduced  at  any
22300	point  in  the  interview.  The  model  knows  that  a topic has been
22400	discussed previously because a topic-flag is set when a  topic  comes
22500	up.
22600	
22700	.F
22800	META-REFERENCES
22900	
23000		These are references, not about a topic directly, but about
23100	what has been said about the topic as in:
23200	.V
23300		(25) DR.- WHY ARE YOU IN THE HOSPITAL?
23400		(26) PT.- I SHOULDNT BE HERE.
23500		(27) DR.-  WHY DO YOU SAY THAT?
23600	.END
23700		The expression (27 ) is about  and meta to expression (26 ).
23800		Sometimes when the patient  makes  a  statement,  the  doctor
23900	replies,  not  with  a  question,  but  with  another statement which
24000	constitutes a rejoinder as in:
24100	.V
24200		(28 ) PT.- I HAVE LOST A LOT OF MONEY GAMBLING.
24300		(29 ) DR.- I GAMBLE QUITE A BIT ALSO.
24400	.END
24500		Here  the  algorithm  interprets  (29  )  as  a  directive to
24600	continue discussing gambling, not as an indication  to  question  the
24700	doctor  about  gambling.  The  one exception to this principle occurs
24800	when the algorithm recognizes  a  chance  to  add  to  its  model  or
24900	representation of the interviewer.
25000	
25100	.F
25200	ELLIPSES
25300	
25400	
25500		In dialogues one finds many ellipses, expressions from  which
25600	one or more words are omitted as in:
25700	.V
25800		(30 ) PT.- I SHOULDNT BE HERE.
25900		(31) DR.- WHY NOT?
26000	.END
26100		Here the complete construction must be understood as:
26200	.V
26300		(32) DR.- WHY SHOULD YOU NOT BE HERE?
26400	.END
26500	Again this is handled by the expectancy list which anticipates a "why
26600	not".
26700		The opposite of ellipsis is redundancy which usually provides
26800	no problem since the same thing is being said more than once as in:
26900	.V
27000		(33 ) DR.- LET ME ASK YOU A QUESTION.
27100	.END
27200		If  an  analysis  were required of this expression (it is not
27300	required here since the expression  is  a  sterotype),  it  would  be
27400	recognized  that  the  verb "ask" takes the noun "question" as direct
27500	object and also a question is something that is asked.
27600	
27700	.F
27800	SIGNALS
27900	
28000		Some fragmentary expressions serve only as directive  signals
28100	to proceed as in:
28200	.V
28300		(34) PT.- I WENT TO THE TRACK LAST WEEK.
28400		(35) DR.- AND?
28500	.END
28600	The  fragment of (35) requests a continuation of the story introduced
28700	in (34). The common expressions found in interviews are "and",  "so",
28800	"go  on", "go ahead", "really", etc. If an input expression cannot be
28900	recognized at all, the lowest level default condition is to assume it
29000	is  a  signal  and either proceed with the next line in a story under
29100	discussion or if the latter is not the case, begin a new story with a
29200	prompting question or statement.
29300	
29400	.F
29500	IDIOMS
29600	
29700		Since so much of conversational language is stereotyped,  the
29800	task  of  recognition  is much easier than that of analysis.  This is
29900	particularly true of idioms. Either one knows what an idiom means  or
30000	one does not. It is usually hopeless to try to decipher what an idiom
30100	means from an analysis of its constituent parts. If the reader doubts
30200	this,  let  him  ponder  the  following expressions taken from actual
30300	teletyped interviews.
30400	.V
30500		(36) DR.- WHATS EATING YOU?
30600		(37) DR.- YOU SOUND KIND OF PISSED OFF.
30700		(38) DR.- WHAT ARE YOU DRIVING AT?
30800		(39) DR.- ARE YOU PUTTING ME ON?
30900		(40) DR.- WHY ARE THEY AFTER YOU?
31000		(41) DR.- HOW DO YOU GET ALONG WITH THE OTHER PATIENTS?
31100	 	(42) DR.- HOW DO YOU LIKE YOUR WORK?
31200		(43) DR.- HAVE THEY TRIED TO GET EVEN WITH YOU?
31300		(44) DR.- I CANT KEEP UP WITH YOU.
31400	.END
31500		Understanding  idioms  is  a  matter of rote memory. Hence an
31600	algorithm with a large idiom table is required.  As  each  new  idiom
31700	appears  in  teletyped  interviews,  it  should be added to the idiom
31800	table because what happens once can happen again.
31900		One advantage in constructing an idiolect for a model is that
32000	it understands its own idiomatic expressions which tend to be used by
32100	the interviewer if he understands them as in:
32200	.V
32300		(45) PT.- THEY ARE OUT TO GET ME.
32400		(46) DR.- WHAT MAKES YOU THINK THEY ARE OUT TO GET YOU.
32500	.END
32600		The expression (45 ) is really a double idiom in which  "out"
32700	means  `intend'  and  "get" means `harm' in this context. Needless to
32800	say.  an algorithm which tried to pair off the  various  meanings  of
32900	"out"  with  the  various meanings of "get" would have a hard time of
33000	it. But an algorithm which understands what it itself is  capable  of
33100	saying, can easily recognize echoed idioms.
33200	
33300	.F
33400	FUZZ TERMS
33500	
33600		In this category fall a large  number  of  expressions  which
33700	have  little  or  no  meaning  and  therefore  can  be ignored by the
33800	algorithm. The lower-case expressions in the following  are  examples
33900	of fuzz:
34000	.V
34100		(47) DR.- well now perhaps YOU CAN TELL ME something ABOUT YOUR FAMILY.
34200		(48) DR.- on the other hand I AM INTERESTED IN YOU.
34300		(49) DR.- hey I ASKED YOU A QUESTION.
34400	.END
34500		It is not the case that in order to ignore something one must
34600	recognize explicitly what is ignorable. Since pattern-matching allows
34700	for an `anything' slot in many of its patterns, fuzz is  thus  easily
34800	ignored.
34900	
35000	.F
35100	SUBORDINATE CLAUSES
35200	
35300		A subordinate clause is a complete statement  inside  another
35400	statement.  It  is  most frequently introduced by a relative pronoun,
35500	indicated in the following expressions by lower case:
35600	.V
35700		(50) DR.-  WAS IT THE UNDERWORLD that PUT YOU HERE?
35800		(51) DR.- WHO ARE THE PEOPLE who UPSET YOU?
35900		(52) DR.- HAS ANYTHING HAPPENED which YOU DONT UNDERSTAND?
36000	.END
36100		The words "whether" and "because" serving as conjunctions are
36200	less  frequent.  A  language-analysis  also   must   recognize   that
36300	subordinate  clauses  can function as nouns, adjectives, adverbs, and
36400	objects of prepositions.
36500	
36600	.F
36700	VOCABULARY
36800	
36900		How many words should there be in the algorithm's vocabulary?
37000	It is a rare human speaker of English who can recognize  40%  of  the
37100	415,000  words  in  the  Oxford  English  Dictionary. In his everyday
37200	conversation an educated person uses perhaps 10,000 words and  has  a
37300	recognition  vocabulary  of  about  50,000  words.  A  study of phone
37400	conversations showed that 96 % of the talk employed only  737  words.
37500	(French, Carter, and Koening, 1930).
37600	Of course the remaining 4% , if not recognized, may be ruinous to the
37700	continuity of a conversation.
37800		In counting the words in 53 teletyped psychiatric interviews,
37900	we found psychiatrists used only 721 words.  Since  we  are  familiar
38000	with  psychiatric  vocabularies and styles of expression, we believed
38100	this language-algorithm could function adequately with  a  vocabulary
38200	of  at  most  a few thousand words. There will always be unrecognized
38300	words. The algorithm must be able to continue even  if  it  does  not
38400	have  a  particular word in its vocabulary. This provision represents
38500	one great advantage of pattern-matching over conventional  linguistic
38600	parsing.
38700		It is not the number of words which creates difficulties  but
38800	their  combinations.  One  thousand  factorial  is still a very large
38900	number. Syntactic and semantic  constraints  in  stereotypes  and  in
39000	analysis reduce this number to an indefinitely large one.
39100	
39200	.F
39300	MISSPELLINGS AND EXTRA CHARACTERS
39400		There is really no good defense  against  misspellings  in  a
39500	teletyped  interview except having a human monitor retype the correct
39600	versions. Spelling correcting programs  are  slow,  inefficient,  and
39700	imperfect.  They  experience  great  problems  when  it  is the first
39800	character in a word which is incorrect.
39900		Extra characters sent by the interviewer or by a bad phone
40000	line can be removed by a human monitor.
40100	
40200	.F
40300	META VERBS
40400	
40500		Certain common verbs such as "think", "feel", "believe", etc
40600	take as their objects a clause as in:
40700	.V
40800		(54) DR.- I THINK YOU ARE RIGHT.
40900		(55) DR.- WHY DO YOU FEEL THE GAMBLING IS CROOKED?
41000	.END
41100		The  verb  "believe"  is  peculiar  since it can also take as
41200	object a noun or noun phrase as in:
41300	.V
41400		(56) DR.- I BELIEVE YOU.
41500	.END
41600		In expression (55) the conjunction "that" can follow the word
41700	"feel" signifying a subordinate clause. This is not  the  case  after
41800	"believe" in expression (56).
41900	
42000	.F
42100	ODD WORDS
42200	
42300		These are words which are odd in the context of  a  teletyped
42400	interview  while  they  are  quite  natural  in  the  usual vis-a-vis
42500	interview in which the participants communicate through speech.  This
42600	should  be  clear  from the following examples in which the odd words
42700	appear in lower case:
42800	.V
42900		(57) DR.-YOU sound CONFUSED.
43000		(58) DR.- DID YOU hear MY LAST QUESTION?
43100		(59) DR.- WOULD YOU come in AND sit down PLEASE?
43200		(60) DR.- CAN YOU say WHO?
43300		(61) DR.- I WILL see YOU AGAIN TOMORROW.
43400	.END
43500	
43600	
43700	.F
43800	MISUNDERSTANDING
43900	
44000		It is not fully recognized bt students of language how often people
44100	misunderstand one another in conversation and yet their
44200	dialogues proceed as if understanding and being understood had taken
44300	place.
44400		The classic story involves three partially deaf men cycling
44500	through the English counrtyside:
44600	.V
44700		FIRST - "WHAT TOWN IS THIS?"
44800		SECOND - "THURSDAY"
44900		THIRD - "ME TOO, LETS STOP AND HAVE A DRINK."
45000	.END
45100		Sometimes a psychiatric interviewer realizes when misunderstanding
45200	occurs and tries to correct it. Other times he simply passes it by. It is
45300	characteristic of the paranoid mode to respond idiosyncratically to
45400	particular word-concepts regardless of what the interviewer is saying:
45500	.V
45600		(62) PT.- IT IS NOT HELPING ME TO BE HERE.  	
45700	        (63) DR.- I BET.
45800	        (64) PT.- GAMBLING HAS BEEN NOTHING BUT TROUBLE FOR ME.
45900	.END
46000	
46100	.F
46200	UNUNDERSTANDING
46300	
46400		A dialogue algorithm  must be prepared for situations
46500	in which it simply does not understand i.e. it cannot arrive at any
46600	interpretation as to what the interviewer is saying. An algorithm should
46700	not be faulted for a lack of facts as in:
46800	.V
46900		(65) DR.- WHO IS THE PRESIDENT OF TURKEY?
47000	.END CONTINUE
47100	wherin the memory does not contain the words "president" and "Turkey".
47200	In this default condition it is simplest to reply:
47300	.V
47400		(66) PT.- I DONT KNOW.
47500	.END CONTINUE
47600	and dangerous to reply:
47700	.V
47800		(67) PT.- COULD YOU REPHRASE THE QUESTION?
47900	.END CONTINUE
48000	because of the horrible loops which can result.
48100		Since the main problem in the default condition of ununderstanding
48200	is how to continue, heuristics can be employed such as asking about the 
48300	interviewer's intention as in:
48400	.V
48500		(68) PT.- WHY DO YOU WANT TO KNOW THAT?
48600	.END CONTINUE
48700	or rigidly continuing with a previous topic or introducing a new topic.
48800		These are admittedly desperate measures  intended  to  prompt
48900	the  interviewer  in  directions the algorithm has a better chance of
49000	understanding. Usually it is the interviewer who  controls  the  flow
49100	from  topic to topic but there are times, hopefully few, when control
49200	must be assumed by the algorithm.