00100	PROBLEMS OF NATURAL LANGUAGE UNDERSTANDING IN TELETYPED INTERVIEW DIALOGUES.
00200	
00300	
00400	     By `natural language` I shall mean everyday American English
00500	such as is used by readers of this book in ordinary conversations.
00600	It is still difficult to be explicit about the processes which
00700	enable hummans to interpret and respond to natural language.
00800	Philosophers, linguists and psychologists have speculated about
00900	and investigated natural language with various purposes and few
01000	useful results.  Now attempts are being made in artificial intelligence to write        
01100	algorithims which `understand' what is being expressed in natural
01200	language utterances.
01300	     During the 1960's when machine processing of natural language
01310	
01400	was dominated by syntactic considerations, it became clear that
01500	this approach was insufficient.  The current view is that to understand
01600	what utterances say, knowledGe about linguistic syntax and semantics
01700	must be combined with knowledge from an underlying conceptual
01800	structure containing a world-model and an ability to draw inferences.
01900	How to achieve this combination efficiently represents a huge task for
02000	both theory and implementation.
02100	     Since the behavior  being simulated by  our paranoid model is the
02200	linguistic(non-nonverbal) behavior of paranoid patients in a psychiatric
02300	interview, the model must have some  ability to process and respond to
02400	natural language input in a manner indicating the underlying pathological
02500	beliefs characteristic of the paranoid mode.  This our purpose was
02600	       to develop a method for understanding everyday English sufficient
02700	for the model to behave conversationally in a paranoid way in a
02800	circumscribed situation.  What is said in this situation is far
02900	richer than what is said in conversations with a block-stacking
03000	
03100	robot but its requirements for constructing an interpretation
03200	of an input are not as complex as trying to understand anything
03300	said in English BY anybody in any dialogue situation.
03400	We took a pragmatic approach which considered "understanding"
03500	to represent "getting the message" of an utterance by
03600	gleaning some {not all} of the relations between them.
03700	this straightforward approach to a complex problem has its
03800	drawbacks, as will be shown, but we were striving for a
03900	sufficiency to demonstrate paranoia rather than complete
04000	comprehension of English.
04100	     Linguistic approaches cite traditional problems with
04200	ambiguity, as illustrated in the following example from
04300	Wilks { }.  Suppose I walked up to you, a stranger, on
04400	the street on Sunday morning and said
04500	     {1} He fell while getting to the ball.'
04600	Admittedly this is a strange scene and in this situation
04700	you would think me to be crazy, hungover and maybe still
04800	drunk, but the example is no more weird than the isolated
04900	examples discussed in the linguistics literature.  Suppose
05000	further that in your personal `dictionary' the word 'ball'
05100	has at least two senses, {A} a spherical physical object
05200	used in a game, and {B} a formal dance.  {It probably has
05300	also a third sense as a verb but we will ignore this more
05400	or less recent example of semantic shift}.  Having no
05500	further information in this situation and attempting to
05600	construct an interpretation of my utterance, you would be
05700	puzzled as to whether I was referring to a ball game or a
05800	dance.  If we then continued on our respective ways, saying
05900	nothing else, your puzzlement would continue and even increase
06000	--I don't know what he was referring to nor why he even said
06100	that to me.
06200	
06300	     The ambiguity arises because of the two word senses for
06400	ball, each of which would give the utterance a meaningful
06500	interpretation.  But the example is extremely forced and
06600	artificial.  Such isolated utterances cannot be disambiguated
06700	{uniqueated is a better term} but this is no handicap for
06800	ordinary human convversations in which ambiguities hardly arise
06900	at all.  Besides the utterance itself, extra information is
07000	usually available in the form of contextual and situational
07100	knowledge. Even better, one can always ask. If I had said
07200	only utterance {1} to you, you could simply ask:
07300	
07400	            {2} `What do you mean?'
07500	
07600	and my reply would indicate something about a game or a
07700	dance or who `he' was.
07800	     Utterances occur in conversations which take place in
07900	sociopsychological situations. The communicants have roles
08000	and intentions towards one another.  If the situation is that
08100	of a medical or psychiatric interview between doctor and
08200	patient and the doctor asks:
08300	
08400	     {3} `How much do you drink?'
08410	
08500	We know from the nature of the situation that drink means
08600	`drink alcohol' and does not refer to a total fluid intake.
08700	
08800	     Dialogues represent connected discourse in which all
08900	the utterances, except perhaps for opening greetings, are
09000	connectable to previous utterances, contexts and sub
09100	contexts, topics and subtopics surround any given utterance
09200	and activate relevant word senses such that alternative
09300	senses do not arise in the comprehension process.  In
09400	spoken dialogues intonations and word emphases are further
09500	means for avoiding ambbiguities, but the connected discourse
09600	of dialogues brings problems of its own to the algorithmist
09700	whose program must keep track of what is going on and what
09800		as been said before.  Foremost is the problem of anaphoric
09900	reference.
10000	
10050	ANAPHORA
10100	     An anaphoric reference is a word or phrase which refers
10200	backwards {usually--i.e. there are some rare forward references}
10300	to something in a previous utterance.  A common example in
10400	intervview dialogues is that of pronouns.
10500	
10600	     {4} Patient - `My father was an alcoholic.'
10700	
10800	     {5} Doctor - `Were you very close to him?'
10900	
11000	Where the term `him' in {4} obviously refers to `father' in
11100	{3}, it is not too difficult a probblem for a program to make
11200	the correct assignments in personal pronoun anaphora.  Of
11300	greater compexity are utterances containing the words 'it'
11400	or `this'.  For example, suppose the interview continued after
11500	{5} as follows:
11600	
11700	     {6} Patient - `Yes I was, even though his drinking
11800	         upset me.'
11900	
12000	     {7} Doctor - `How did it upset you?'
12100	
12200	Here the comprehension algorithm must grasp that the `it' of
12300	{7} refers to the topic of the father's drinking in {6}.
12400	Further, if the dialogue continued:
12500	
12600	     {8}Patient - 'It embarassed me when my friends saw
12700	       him drunk.
12800	
12900	     {9} Doctor - `Do you think he sensed this.'
13000	
13100	then it must bbe understood that `this' refers to the patient's
13200	embarassment.  How this is implemented in the program will be
13300	described in xxx.
13400	
13500	FRAGMENTS
13600	
13700	     Another major probblem for algorithms which attempt to understand
13800	discourse consists of the fact that many of the input expressions
13900	are not well-formed.  All sorts of fragments and ellipses appear
14000	which must somehow be connected to conceptualizations under discussion.
14100	For example, consider the following exchange:
14200	
14300	   
14400	     {10} Dr. - `How do you like the hospital?'
14500	
14600	     {11} Pt. - `I shouldn't be here.'
14700	
14800	     {12} Dr. - `Why not?'
14900	The question {12} is an elliptical expression for the full concept
15000	ualization
15100	
15200	     `Why should you not be in the hospital?'
15300	
15400	     Junk words {`well now'} {`tell me more'} and go ahead signals
15500	must be responded to by continuation of a topic.
15600	
15700	For example:
15800	
15900	     {13} Pt.- `I went to the track last week.'
16000	
16100	     {14} Dr. - `Really?'
16200	
16300	Such expressions as {14} stand in a meta-relation to the topic and
16400	serve to keep the conversation going.
16500	
16600	REJOINDERS
16700	
16800	     Sometimes the input expression from the interviewer is a rejoinder
16900	, a reply to a reply by the patient.  For instance:
17000	
17100	     {15} Dr. - `Who are you afraid of?'
17200	
17300	     {16} Pt. - `The Mafia is out to get me.'
17400	
17500	     {17} Dr. - `I would bbe afraid of them also.'
17600	
17700	in which {17} is a rejoinder.  Such expressions are not requests for
17800	information but provide information for the patient's model of the
17900	interviewer.
18000	
18100	INTERVIEWER-INTERVIEWEE RELATIONS
18200	
18300	     It is characteristic of psychiatric interviewing that the
18400	participants from time to time do not simply talk about the
18500	patient.  Two situations exist concurrently in an interview,
18600	one being talked about and one the participants are in.  At
18700	times the second situation becomes the first.  When the partici
18800	pants discuss one another and their relation, the dialogue
18900	expressions contain intentional verbs which in English fit the
19000	pattern `I X you' or `you X me'.  The comprehension process must
19100	distinguish clearly between subbject and object in the case of some
19200	of these verbs.  For example in
19300	
19400	     {18} `I like you'
19500	
19600	the speaker 'I' experiences the liking but in
19700	
19800	     {19} `Do I please you?'
19900	
20000	the `you' experiences the pleasure as a consequence of something
20100	`I' does.