perm filename CHAP4[4,KMC]18 blob sn#049370 filedate 1973-06-15 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   language-behavior  of  a  paranoid  patient  in  a  psychiatric
00600	interview, the model must have an ability to interpret and respond to
00700	natural    language   input   sufficient   to   demonstrate   conduct
00800	characteristic of the paranoid mode.   By "natural language" I  shall
00900	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]. Machine understanding of natural
01800	language is being attempted today but  it  is  not  something  to  be
01900	completly achieved today or even tomorrow. The problem is not to find
02000	immediately the best way of doing it but to find any way at all.
02100		During the 1960's when machine processing of natural language
02200	was dominated by  syntactic  considerations,  it  became  clear  that
02300	syntactical  information  alone  was  insufficient  to comprehend the
02400	expressions of ordinary conversations. A  current  view  is  that  to
02500	understand  what  is  said  in  linguistic  expressions, knowledge of
02600	syntax and semantics must be combined with beliefs from a  conceptual
02700	structure   capable   of  making  inferences.  How  to  achieve  this
02800	combination  efficiently  with  a  large   data-base   represents   a
02900	monumental task for both theory and implementation.
03000		We did not attempt to  construct  a  conventional  linguistic
03100	parser  to  deal  with  natural  language  input for several reasons.
03200	Parsers to date have  great  difficulty  in  assigning  a  meaningful
03300	interpretation to the expressions of everyday conversational language
03400	using  unrestricted  English.  Purely  syntactic  parsers   offer   a
03500	cancerous proliferation of interpretations. A conventional parser may
03600	simply halt when it comes  across  a  word  not  in  its  dictionary.
03700	Parsers  represent  tight  conjunctions  of  tests  instead  of loose
03800	disjunctions  needed  for  everyday  language   which   may   involve
03900	misunderstandinga  and  ununderstandings.  People  misunderstand  and
04000	ununderstand at times and thus are partially opaque to one another.
04100		The  language  analysis  utilized by the model first puts the
04200	input in the form of a list and then determines the syntactical  type
04300	for  the  input  expression-  question,  statement or imperative. The
04400	expression-type is scanned in an attempt to form a conceptualization,
04500	i.e.  a  predication  of  an  attribute,  on  an object or a relation
04600	between objects. An attribute consists of something one is or does or
04700	possesses.   The   resultant  conceptualization  is  then  classified
04800	according to the rules of Fig. 00 in %000 as  malevolent,  benevolent
04900	or neutral.
05000		How language is understood depends on the intentions  of  the
05100	producers   and  interpreters  in  the  dialogue.  Thus  language  is
05200	understood  in  accordance  with  the  participant's  view   of   the
05300	situation.  Our  purpose  was  to  develop a method for understanding
05400	everyday  English   sufficient   for   the   model   to   communicate
05500	linguistically  in a paranoid way in the circumscribed situation of a
05600	psychiatric interview. Such an interview is not small talk; a job  is
05700	to be done.
05800		We did not try to construct a general-purpose algorithm which
05900	could  understand  anything  said in English by anybody to anybody in
06000	any dialogue situation. (Does anyone believe it possible?) We  sought
06100	only to extract or distill an idiolectic meaning or gist of a meaning
06200	from the input. We took as pragmatic measures of "understanding"  the
06210		Natural language is not an agree-on universe of discourse
06220	such as arithmetic wherein symbols have the same meaning for everyone
06221	who uses them. What we loosely call "natural language" is actually a set
06222	of idiolects, each being unique to the individual with a unique history.
06223	To be unique does not mean that no property is shared with other
06224	individuals, only that not every property is shared. An overlap of
06225	idiolects allows the communication of shared meanings.
06227		We took as pragmatic measures of "understanding" the
06300	ability  (1)  to  form  a  conceptualization so that questions can be
06400	answered and commands carried out, (2) to determine the intention  of
06500	the  interviewer,  (3)  to  determine the references for pronouns and
06600	other anticipated topics.  This straightforward approach to a complex
06700	problem  has  its  drawbacks,  as  will be shown, but we strove for a
06800	highly individualized idiolect  sufficient  to  demonstrate  paranoid
06900	processes  of an individual in a particular situation rather than for
07000	a general supra-individual or ideal comprehension of English.  If the
07100	language-understanding  system  interfered  with  demonstrating   the
07200	paranoid  processes,  we would consider it defective and insufficient
07300	for our purposes.
07400		Some  special  problems a dialogue algorithm must handle in a
07500	psychiatric interview  will  now  be  outlined  along  with  a  brief
07600	description of how the model deals with them.
07700	
07800	.F
07900	QUESTIONS
08000	
08100		The principal expression-type used by an interviewer consists
08200	of  a  question. A question is recognized by its beginning with a wh-
08300	or how form and/or the expression ending  with  a  question-mark.  In
08400	teletyped  interviews  a question may sometimes be put in declarative
08500	form followed by a question mark as in:
08600	.V
08700		(1) PT.- I LIKE TO GAMBLE ON THE HORSES.             	
08800		(2) DR.- YOU GAMBLE?
08900	.END
09000	Although  the  verb  is  missing  in (2), the model recognizes that a
09100	question is being asked about its gambling.
09200	
09300	Particularly  difficult  are  `when' questions which require a memory
09400	which can assign each event a beginning,  end  and  a  duration.   An
09500	improved   version  of  the  model  will  have  this  capacity.  Also
09600	troublesome are questions such as `how often', `how  many',  i.e.   a
09700	`how'  followed  by a quantifier. If the model has "how often" on its
09800	expectancy list while a topic is under  discussion,  the  appropriate
09900	reply can be made. Otherwise the model ununderstands.
10000		In constructing a simulation  of  a  thought  process  it  is
10100	arbitrary  how  much  information  to represent in memory. Should the
10200	model know what is the capital of Alabama? It is trivial to  store  a
10300	lot  of  facts.  We took the position that the model should know only
10400	what we believed it reasonable to know  relevant  to  a  few  hundred
10500	topics expectable in a psychiatric interview. Thus the model performs
10600	badly when subjected to baiting `exam' questions designed to test its
10700	informational  limitations  rather  than  to  seek useful psychiatric
10800	information.
10900	
11000	.F
11100		IMPERATIVES
11200	
11300		Typical imperatives in a  psychiatric  interview  consist  of
11400	expressions like:
11500	.V
11600		(3) DR.- TELL ME ABOUT YOURSELF.
11700		(4)  DR.-  LETS  DISCUSS  YOUR  FAMILY. 
11800	.END
11900		Such   imperatives   are   actually   interrogatives  to  the
12000	interviewee about the topics they refer to.  Since the only  physical
12100	action  the  model can perform is to `talk' , imperatives are treated
12200	as requests for information.
12300	
12400	.F
12500	DECLARATIVES
12600	
12700		In  this  category  is  lumped  everything  else. It includes
12800	greetings, farewells, yes-no type answers, existence  assertions  and
12900	the usual predications.
13000	
13100	.F
13200	AMBIGUITIES
13300	
13400		Words have more than  one  sense,  a  convenience  for  human
13500	memories  but  a struggle for language-analysing algorithms. Consider
13600	the word "bug" in the following expressions:
13700	.V
13800		(5) AM I BUGGING YOU?
13900		(6) AFTER A PERIOD OF HEAVY DRINKING HAVE YOU  FELT  BUGS  ON
14000	            YOUR SKIN?
14100		(7) DO YOU THINK THEY PUT A BUG IN YOUR ROOM? 
14200	.END
14300		In  expression  (5)  the term "bug" means to annoy, in (6) it
14400	refers to an insect and in (7) it refers to  a  microphone  used  for
14500	hidden  survellience. Some words like "run" have fifty or more common
14600	senses. Context is used to carry  out  disambiguation.  For  example,
14700	when the Mafia is under discussion the model interprets "bug" to mean
14800	microphone. Thus we have the advantage of an idiolect  where  we  can
14900	arbitrarily  restrict  the  word  senses.  One  characteristic of the
15000	paranoid mode is that no matter in what sense the interviewer uses  a
15100	word,  the  patient  may idiosyncratically interpret it in some sense
15200	
15300	.F
15400	ANAPHORIC REFERENCES
15500		The common anaphoric references consist of the pronouns "it",
15600	"he", "him", "she", "her", "they", "them" as in:
15700	.V
15800		(8) PT.-HORSERACING IS MY HOBBY.
15900		(9) DR.-WHAT DO  YOU  ENJOY  ABOUT  IT?  
16000	.END
16100		The algorithm recognizes "it" as referring  to  "horseracing"
16200	because  "it" has been placed on the expectancy list when horseracing
16300	was introduced in (8). A  more  difficult  problem  arises  when  the
16400	anaphoric  reference  points  more  than  one  I/O  pair  back in the
16500	dialogue as in:
16600	.V
16700		(10) PT.-THE MAFIA IS OUT TO GET ME.
16800		(11) DR.- ARE YOU AFRAID OF THEM?
16900		(12) PT.- MAYBE.
17000		(13) DR.- WHY IS THAT? 
17100	.END
17200		The "that" of expression (13) does not refer to (12)  but  to
17300	the  topic  of being afraid which the interviewer introduced in (11).
17400	Another pronominal confusion occurs when the interviewer uses `we' in
17500	two senses as in:
17600	.V
17700		(14) DR.- WE WANT YOU TO STAY IN THE HOSPITAL.
17800		(15) PT.- I WANT TO BE DISCHARGED NOW.
17900		(16) DR.- WE ARE NOT COMMUNICATING. 
18000	.END
18100		In expression (14) the interviewer is using "we" to refer  to
18200	psychiatrists  or the hospital staff while in (16) the term refers to
18300	the interviewer and patient. Identifying the correct  referent  would
18400	require beliefs about the dialogue which the new version of the model
18500	will have.
18600	
18700	.F
18800	TOPIC SHIFTS
18900	
19000		In the main a psychiatric interviewer is in control of the
19100	interview. When he has gained sufficient information about a topic,
19200	he shifts to a new topic. Naturally the algorithm must detect this
19300	change of topic as in the following:
19400	.V
19500		(17) DR.- HOW DO YOU LIKE THE HOSPITAL?
19600		(18) PT.- ITS NOT HELPING ME TO BE HERE.
19700		(19) DR.- WHAT BROUGHT YOU TO THE HOSPITAL?
19800		(20) PT.- I AM VERY UPSET AND NERVOUS.
19900		(21) DR.- WHAT TENDS TO MAKE YOU NERVOUS?
20000		(23) PT.- JUST BEING AROUND PEOPLE.
20100		(24) DR.- ANYONE IN PARTICULAR?
20200	.END
20300		In (17) and (19) the topic is the hospital. In (21) the topic
20400	changes to causes of the patient's nervous state.
20500		When a topic is introduced by  the  patient  as  in  (20),  a
20600	number  of  things  can  be  expected  to be asked about it. Thus the
20700	algorithm has ready an updated expectancy-anaphora list which  allows
20800	it  to  determine  whether the topic introduced by the model is being
20900	responded to or  whether  the  interviewer  is  continuing  with  the
21000	previous topic.
21100		Topics touched upon previously can be  re-introduced  at  any
21200	point  in  the  interview.  The  model  knows  that  a topic has been
21300	discussed previously because a topic-flag is set when a  topic  comes
21400	up.
21500	
21600	.F
21700	META-REFERENCES
21800	
21900		These are references, not about a topic directly, but about
22000	what has been said about the topic as in:
22100	.V
22200		(25) DR.- WHY ARE YOU IN THE HOSPITAL?
22300		(26) PT.- I SHOULDNT BE HERE.
22400		(27) DR.-  WHY DO YOU SAY THAT?
22500	.END
22600		The expression (27 ) is about  and meta to expression (26 ).
22700		Sometimes when the patient  makes  a  statement,  the  doctor
22800	replies,  not  with  a  question,  but  with  another statement which
22900	constitutes a rejoinder as in:
23000	.V
23100		(28 ) PT.- I HAVE LOST A LOT OF MONEY GAMBLING.
23200		(29 ) DR.- I GAMBLE QUITE A BIT ALSO.
23300	.END
23400		Here  the  algorithm  interprets  (29  )  as  a  directive to
23500	continue discussing gambling, not as an indication  to  question  the
23600	doctor  about  gambling.  The  one exception to this principle occurs
23700	when the algorithm recognizes  a  chance  to  add  to  its  model  or
23800	representation of the interviewer.
23900	
24000	.F
24100	ELLIPSES
24200	
24300	
24400		In dialogues one finds many ellipses, expressions from  which
24500	one or more words are omitted as in:
24600	.V
24700		(30 ) PT.- I SHOULDNT BE HERE.
24800		(31) DR.- WHY NOT?
24900	.END
25000		Here the complete construction must be understood as:
25100	.V
25200		(32) DR.- WHY SHOULD YOU NOT BE HERE?
25300	.END
25400	Again this is handled by the expectancy list which anticipates a "why
25500	not".
25600		The opposite of ellipsis is redundancy which usually provides
25700	no problem since the same thing is being said more than once as in:
25800	.V
25900		(33 ) DR.- LET ME ASK YOU A QUESTION.
26000	.END
26100		If  an  analysis  were required of this expression (it is not
26200	required here since the expression  is  a  sterotype),  it  would  be
26300	recognized  that  the  verb "ask" takes the noun "question" as direct
26400	object and also a question is something that is asked.
26500	
26600	.F
26700	SIGNALS
26800	
26900		Some fragmentary expressions serve only as directive  signals
27000	to proceed as in:
27100	.V
27200		(34) PT.- I WENT TO THE TRACK LAST WEEK.
27300		(35) DR.- AND?
27400	.END
27500	The  fragment of (35) requests a continuation of the story introduced
27600	in (34). The common expressions found in interviews are "and",  "so",
27700	"go  on", "go ahead", "really", etc. If an input expression cannot be
27800	recognized at all, the lowest level default condition is to assume it
27900	is  a  signal  and either proceed with the next line in a story under
28000	discussion or if the latter is not the case, begin a new story with a
28100	prompting question or statement.
28200	
28300	.F
28400	IDIOMS
28500	
28600		Since so much of conversational language is stereotyped,  the
28700	task  of  recognition  is much easier than that of analysis.  This is
28800	particularly true of idioms. Either one knows what an idiom means  or
28900	one does not. It is usually hopeless to try to decipher what an idiom
29000	means from an analysis of its constituent parts. If the reader doubts
29100	this,  let  him  ponder  the  following expressions taken from actual
29200	teletyped interviews.
29300	.V
29400		(36) DR.- WHATS EATING YOU?
29500		(37) DR.- YOU SOUND KIND OF PISSED OFF.
29600		(38) DR.- WHAT ARE YOU DRIVING AT?
29700		(39) DR.- ARE YOU PUTTING ME ON?
29800		(40) DR.- WHY ARE THEY AFTER YOU?
29900		(41) DR.- HOW DO YOU GET ALONG WITH THE OTHER PATIENTS?
30000	 	(42) DR.- HOW DO YOU LIKE YOUR WORK?
30100		(43) DR.- HAVE THEY TRIED TO GET EVEN WITH YOU?
30200		(44) DR.- I CANT KEEP UP WITH YOU.
30300	.END
30400		Understanding  idioms  is  a  matter of rote memory. Hence an
30500	algorithm with a large idiom table is required.  As  each  new  idiom
30600	appears  in  teletyped  interviews,  it  should be added to the idiom
30700	table because what happens once can happen again.
30800		One advantage in constructing an idiolect for a model is that
30900	it understands its own idiomatic expressions which tend to be used by
31000	the interviewer if he understands them as in:
31100	.V
31200		(45) PT.- THEY ARE OUT TO GET ME.
31300		(46) DR.- WHAT MAKES YOU THINK THEY ARE OUT TO GET YOU.
31400	.END
31500		The expression (45 ) is really a double idiom in which  "out"
31600	means  `intend'  and  "get" means `harm' in this context. Needless to
31700	say.  an algorithm which tried to pair off the  various  meanings  of
31800	"out"  with  the  various meanings of "get" would have a hard time of
31900	it. But an algorithm which understands what it itself is  capable  of
32000	saying, can easily recognize echoed idioms.
32100	
32200	.F
32300	FUZZ TERMS
32400	
32500		In this category fall a large  number  of  expressions  which
32600	have  little  or  no  meaning  and  therefore  can  be ignored by the
32700	algorithm. The lower-case expressions in the following  are  examples
32800	of fuzz:
32900	.V
33000		(47) DR.- well now perhaps YOU CAN TELL ME something ABOUT YOUR FAMILY.
33100		(48) DR.- on the other hand I AM INTERESTED IN YOU.
33200		(49) DR.- hey I ASKED YOU A QUESTION.
33300	.END
33400		It is not the case that in order to ignore something one must
33500	recognize explicitly what is ignorable. Since pattern-matching allows
33600	for an `anything' slot in many of its patterns, fuzz is  thus  easily
33700	ignored.
33800	
33900	.F
34000	SUBORDINATE CLAUSES
34100	
34200		A subordinate clause is a complete statement  inside  another
34300	statement.  It  is  most frequently introduced by a relative pronoun,
34400	indicated in the following expressions by lower case:
34500	.V
34600		(50) DR.-  WAS IT THE UNDERWORLD that PUT YOU HERE?
34700		(51) DR.- WHO ARE THE PEOPLE who UPSET YOU?
34800		(52) DR.- HAS ANYTHING HAPPENED which YOU DONT UNDERSTAND?
34900	.END
35000		The words "whether" and "because" serving as conjunctions are
35100	less  frequent.  A  language-analysis  also   must   recognize   that
35200	subordinate  clauses  can function as nouns, adjectives, adverbs, and
35300	objects of prepositions.
35400	
35500	.F
35600	VOCABULARY
35700	
35800		How many words should there be in the algorithm's vocabulary?
35900	It is a rare human speaker of English who can recognize  40%  of  the
36000	415,000  words  in  the  Oxford  English  Dictionary. In his everyday
36100	conversation an educated person uses perhaps 10,000 words and  has  a
36200	recognition  vocabulary  of  about  50,000  words.  A  study of phone
36300	conversations showed that 96 % of the talk employed only  737  words.
36310	(French, Carter, and Koening, 1930).
36400	Of course the remaining 4% , if not recognized, may be ruinous to the
36500	continuity of a conversation.
36600		In counting the words in 53 teletyped psychiatric interviews,
36700	we found psychiatrists used only 721 words.  Since  we  are  familiar
36800	with  psychiatric  vocabularies and styles of expression, we believed
36900	this language-algorithm could function adequately with  a  vocabulary
37000	of  at  most  a few thousand words. There will always be unrecognized
37100	words. The algorithm must be able to continue even  if  it  does  not
37200	have  a  particular word in its vocabulary. This provision represents
37300	one great advantage of pattern-matching over conventional  linguistic
37400	parsing.
37500		It is not the number of words which creates difficulties  but
37600	their  combinations.  One  thousand  factorial  is still a very large
37700	number. Syntactic and semantic  constraints  in  stereotypes  and  in
37800	analysis reduce this number to an indefinitely large one.
37900	
38000	.F
38100	MISSPELLINGS AND EXTRA CHARACTERS
38200		There is really no good defense  against  misspellings  in  a
38300	teletyped  interview except having a human monitor retype the correct
38400	versions. Spelling correcting programs  are  slow,  inefficient,  and
38500	imperfect.  They  experience  great  problems  when  it  is the first
38600	character in a word which is incorrect.
38700		Extra characters sent by the interviewer or by a bad phone
38800	line can be removed by a human monitor.
38900	
39000	.F
39100	META VERBS
39200	
39300		Certain common verbs such as "think", "feel", "believe", etc
39400	take as their objects a clause as in:
39500	.V
39600		(54) DR.- I THINK YOU ARE RIGHT.
39700		(55) DR.- WHY DO YOU FEEL THE GAMBLING IS CROOKED?
39800	.END
39900		The  verb  "believe"  is  peculiar  since it can also take as
40000	object a noun or noun phrase as in:
40100	.V
40200		(56) DR.- I BELIEVE YOU.
40300	.END
40400		In expression (55) the conjunction "that" can follow the word
40500	"feel" signifying a subordinate clause. This is not  the  case  after
40600	"believe" in expression (56).
40700	
40800	.F
40900	ODD WORDS
41000	
41100		These are words which are odd in the context of  a  teletyped
41200	interview  while  they  are  quite  natural  in  the  usual vis-a-vis
41300	interview in which the participants communicate through speech.  This
41400	should  be  clear  from the following examples in which the odd words
41500	appear in lower case:
41600	.V
41700		(57) DR.-YOU sound CONFUSED.
41800		(58) DR.- DID YOU hear MY LAST QUESTION?
41900		(59) DR.- WOULD YOU come in AND sit down PLEASE?
42000		(60) DR.- CAN YOU say WHO?
42100		(61) DR.- I WILL see YOU AGAIN TOMORROW.
42200	.END
42300	
42400	
42500	.F
42600	MISUNDERSTANDING
42700	
42800		It is not fully recognized bt students of language how often people
42900	misunderstand one another in conversation and yet their
43000	dialogues proceed as if understanding and being understood had taken
43100	place.
43200		The classic story involves three partially deaf men cycling
43300	through the English counrtyside:
43400	.V
43500		FIRST - "WHAT TOWN IS THIS?"
43600		SECOND - "THURSDAY"
43700		THIRD - "ME TOO, LETS STOP AND HAVE A DRINK."
43800	.END
43900		Sometimes a psychiatric interviewer realizes when misunderstanding
44000	occurs and tries to correct it. Other times he simply passes it by. It is
44100	characteristic of the paranoid mode to respond idiosyncratically to
44200	particular word-concepts regardless of what the interviewer is saying:
44300	.V
44400		(62) PT.- IT IS NOT HELPING ME TO BE HERE.  	
44500	        (63) DR.- I BET.
44600	        (64) PT.- GAMBLING HAS BEEN NOTHING BUT TROUBLE FOR ME.
44700	.END
44800	
44900	.F
45000	UNUNDERSTANDING
45100	
45200		A dialogue algorithm  must be prepared for situations
45300	in which it simply does not understand i.e. it cannot arrive at any
45400	interpretation as to what the interviewer is saying. An algorithm should
45500	not be faulted for a lack of facts as in:
45600	.V
45700		(65) DR.- WHO IS THE PRESIDENT OF TURKEY?
45800	.END CONTINUE
45900	wherin the memory does not contain the words "president" and "Turkey".
46000	In this default condition it is simplest to reply:
46100	.V
46200		(66) PT.- I DONT KNOW.
46300	.END CONTINUE
46400	and dangerous to reply:
46500	.V
46600		(67) PT.- COULD YOU REPHRASE THE QUESTION?
46700	.END CONTINUE
46800	because of the horrible loops which can result.
46900		Since the main problem in the default condition of ununderstanding
47000	is how to continue, heuristics can be employed such as asking about the 
47100	interviewer's intention as in:
47200	.V
47300		(68) PT.- WHY DO YOU WANT TO KNOW THAT?
47400	.END CONTINUE
47500	or rigidly continuing with a previous topic or introducing a new topic.
47600		These are admittedly desperate measures  intended  to  prompt
47700	the  interviewer  in  directions the algorithm has a better chance of
47800	understanding. Usually it is the interviewer who  controls  the  flow
47900	from  topic to topic but there are times, hopefully few, when control
48000	must be assumed by the algorithm.