A COMPUTATIONAL THEORY OF PERSPECTIVE AND REFERENCE IN NARRATIVE 
Janyce M. Wlebe and William J. Rapaport 
Department of Computer Science 
State University of New York at Buffalo 
Buffalo, NY 14260 
wiebe~s.buffalo.edu, rapapurt~s.buffalo.edu 
ABSTRACT 
Narrative passages told from a character's perspective 
convey the character's thoughts and perceptions. We 
present a discourse process that recognizes characters' 
thoughts and perceptions in third-person narrative. An 
effect of perspective on reference In narrative is 
addressed: references in passages told from the perspec- 
tive of a character reflect the character's beliefs. An 
algorithm that uses the results of our discourse process 
to understand references with respect to an appropriate 
set of beliefs is presented. 
1. INTRODUCTION. A narrative is often told from the 
perspective of one or more of its characters; it cam also con- 
tain passages that are not told from the perspective of any 
character. We present a computational theory of how 
readers recognize the current perspective in thixd-person n~- 
rative, end of the effects of perspective on the way readers 
understand references in third-person narrative. We consider 
published novels and short stories, rather than m.ificially 
constructed narratives. 
2. BANFIELD'S THEORY. Our notion of perspective in 
narrative is based on Ann Bardield's (1982) c~t_egorization 
of the sentences of narration into subjective and objective 
sentences. Subjective sentences include those that portray a 
character's thoughts (represented thought) or present a scene 
a~ a character perceives it (represented perception). Objec- 
tive sentences present the story directly, rather than through 
the thoughts or perceptions of a character. The language 
used to convey thoughts and perceptions is replete with 
linguistic elements that make no sense unless they are inter- 
preted with respect to the thinking or perceiving character's 
consciousness. Banfield calls them subjective elements; they 
appear only in subjective sentences and cannot appear within 
objective sentences. Banfield identifies perspective in narra- 
tive with subjectivity, which is expressible via subjective 
elements. We call the thinking or perceiving character of a 
subjective sentence the subjective character. 
3. A DISCOURSE-LEVEL APPROACH. Our task of 
recognizing the currant perspective is, therefore, to recognize 
subjective sentences and the subjective characters to whom 
they are attributed. However, we cannot take a umtence- 
by-sentence approach, deciding independently for each sen- 
tence whether it is objective or subjective, and, if subjective, 
who the subjective character is. First, although thoughts and 
perceptions are often reported (as by sentences beginning 
with "He thought that ..." or "She saw ... "), end 
thoughts are often accompsnied by narrative parerttheticaLs 
(such as "he thought" or "he realized"), many thoughts 
and perceptions are not marked in these ways. Second, sub- 
jective sentences do not always explicitly indicate who the 
subjective character is. For example: 
(1) 1"1He wanted to talk to Dennys. 1"2How were they 
going to be able to get home from this strange desert 
land into which they had been cast and which was 
hcaven knew where in all the countless solar systems 
in all the countless galaxies? \[L'Engle, Many Waters, 
p. 91\] 
(2) ~lBut what \[Muhammad\] had seen in those few 
moments m~e him catch his breath in amazement. 
z~On the floor of the cave, which curved back in a 
nalmal fault in the rock, there were several large 
cylindrical objects sumdin 8 in a row. \[John Allegro, 
The Dead Sea Scrolls\] 
Sentence (1.2) is a represented thought, and (2.2) is a 
represented perception, presenting what the character sees as 
he sees it; yet neither is explicitly marked as such. Also, nei- 
ther indicates who the subjective character is. Finally, 
although a subjective element marks a santence as subjective 
(cf. Section 4.2), not all subjective sentences contain subjec- 
tive alements, and subjective elements do not in general 
indicate who the subjective character is. 
However, subjective sentences that are not marked as 
such, or that do not indicate who the subjective character is, 
usually appear in the midst of other subjective sentences 
attributed to the same subjective character. That is, once a 
clearly marked subjective sentence appears for which the 
subjective character can be deternfined" unmarked subjective 
sentences auributed to the same subjective character often 
follow. Thus, to recognize subjective smumces in general 
we need to consider subjectivity at the level of the discourse. 
For this reason, we extend the nodons of subjective and 
objective sentences to the notions of subjective and objective 
contexts, which consist of one or more subjective sentences 
attributed to the same subjective character, or one or more 
objective sentences, respectively. 
Our algorithm for recognizing the current perspective 
is a discourse process that looks for the boundaries of sub- 
jective contexts. During narrative understending, it main- 
ta~ a stack, called the current perspecdve (CP). At the 
beginning of a narrative, the CP is initialized to be the 
reader. When a new subjective context is recognized, its 
131 
subjective character is pushed onto the CP. When the ead of 
a subjective context is recognized, a character is popped 
from the CP. More precisely, since SNePS (Shapiro 1979), 
our knowledge representation system, is fully intmsional, 
only the reader's concepts of the characters are represented 
(Msida and Shapiro 1982, Shapiro and Rapaport 1987). So, 
it is actually the re~_~__--'s concepts of the char~ters that ~e 
pushed onto the CP. 
4. RECOGNIZING SUBJECTIVE CONTEXTS. To 
recognize subjective contexts, our discourse process relies 
exclusively on linguistic signals of subjective contexts. In 
this, it is incomplete: If a subjective cmltext appears in 
which these linguistic signals are not present, then the sub- 
jective context is not recognized. 
4.1. Psychological Verbs, Actions, Adjectives, and Per- 
ceptual Verbs. Reports involving psychological vorbs (e.g~ 
'think', 'wonder', 'realize', 'went', 'remember') or percep- 
tual verbs (e.g.. 'see', 'hear') signal that a subjective coatext 
will follow. So do predicate-adjective sentences with 
psychological adjectives (e.g., 'delighted', 'happy', 'jealous', 
'scared') (cf. Banfield (1982). Cohn (1978). Dole~el (1973)). 
In addition, we have identified what we call psychological 
actions--e.g., "he smiled to himself", "she gasped", "she 
winced"--which function in the same way as psychological 
verbs. 
A sentence of one of these types is a typical way of 
establishi~ a subjective context. Exemples (1) and (2), 
above, and (3), below, exhibit this l~n ~*-n: 
O) 3"1She \[Hannah\] winced as she heard them crash to 
the platform. 3"~he lovely little mirror that she had 
brought for Ellen, and the gifts for the baby\[ \[F~at~- 
chere, Hannah Herse~ p. 3\] 
In each example, the first smtence is a psycholosical or per- 
ceptual report, end the second is a represented thought or 
represented perception, respectively; the subjective character 
of the second sentence is taken to be the subject of the firsL 
In our discourse process, the subject of a porceptmd or 
psychological report, or of a In~Aicete-adjective sentenco 
with a psychological adjective, is pushed onto the CP if a 
character isn't already on the top of it. If a character is 
already on the top of the CP, then no change is made, and 
the sentence is understood to be part of the already esta- 
blished subjective ccnteXL 
4.2. Subjective Elements. Many subjective elements mark 
a sentence in third-person narrative as subjective because 
they are expressive in nalme. Some that Banfield identifies 
are ectclanmt/ons, which express emotion; que~ons, which 
express wonder;, ep/¢hets, such as 'the bastard', which 
express some qualification of the refeie~t; and certain k/n- 
ship terms, e.g., 'Daddy', 'Morn', and 'Aunt Margaret', 
which express a relationship to the referee. She also 
identifies evaluative adjec¢/ves, which express an attitude 
toward the referent, e.g., 'ghastly', 'surprising', 'poor', and 
'dasrmed', although some evaluative adjectives, such as 
'poor' and 'damned', have their evaluative meanings only 
when they occ~ in certain parts of the sentence. Intensifiers 
such as 'too', 'quite', and 'so' are also evaluative (Banfield 
1982), as in: 
(4) He could tell they were tears because his eyes were 
too shiny. Too round. \[Bddgers, All Together Now, 
p. 92\] 
So are emphasizers, such as 'really' and 'just'. An example 
is 'really' in (5.3): 
O) XtJody managed a frail smile, s2she was a little bit 
ashamed. SzShe should really try to be more cheerful 
for Aunt Margaret's sake. S~*After all, Atmt Margaret 
had Imubles of her own--she was the mother of that 
ghastly Dill. \[Gage, Miss Osborne-the-Mop, pp. 16- 
17\] 
Modal verbs of obligation, possibility, and necessity are also 
expressive. For example, 'should', in (5.3), is a modal verb 
of obligation. So are many content (or atti:udinal) disjunct~ 
which comment on the conumt of the utterance (Quirk et al. 
1985). For example, "l~ely', 'maybe', 'probably', and 
'perhaps' express some degree of doubt: 
(6) Something jingled---cer keys probably. \[Oneal, War 
Work, p. 132\] 
Conjoncts, which comment on the connection between items 
(Quirk et el. 1985), can also be expressive. For example, 
'anyhow', 'anyway', 'still', and 'after all' express conces- 
sion (Quirk et ai. 1985). An example is 'After an' in (5.4). 
Other subjective elements am sentence fragments 
(Benfieki 1982), such as (7.2), 
(7) 7"1His brain worked slowly through what he knew 
about this person. ~David's kid. \[Bridgers, All 
Together Now, p. 91\] 
and the uses of 'this', 'that', 'these', and 'those' that Robin 
Lakoff (1974) has identified as emotional dei.xb. In conver- 
sation, they are "generally linked to the speaker's emotional 
involvement in the subject-matter of his utterance" (Lakoff 
1974: 347); in thlrd-person narrative, they are linked to the 
subjective character's emotional involvement in the subject 
matter of his thoughts or perceptions. Examples are 'this' in 
(8.1) and 'That' in 0:2): 
(8) S'llbrahim could remember every time this godless pig 
had patronized lOzn ,.. \[Clancy, Red Dawn Rbing, 
p. 13\] 
132 
O) 9aAs she watched, a wave of jealousy spread through 
her. 9-2'rhat insufferable stranger who had passed 
them on the road was receiving the welcome that she 
had been dreaming of all the way from Connecticut. 
\[Franchere, Hannah Herself, p. 15\] 
In speech, the emotion, evaluation, etc, expressed by 
a subjective element is always attributed to the speaker; in 
third-persun narrative, it is auributed to a character, l Clearly, 
many types of language-understanding abilities are needed m 
understand the range of subjective elements. Our purpose 
here is to show how our discourse process uses them as 
markers of subjective contexts, and how it determin~ the 
subjective character whose thoughts or perceptions they 
mark. However, recognizing the subjective character is 
always required before a subjective element can be under- 
stood. 
When a subjective element is m~.ountered in the nar- 
rative, our discourse process uIxiates the CP according to the 
following algoriflun: 
(A1) 
If there is currently a character on the CP. 
Xthen do not change the CP 
else if there is an actor focus at the start of the current 
sentence who is a character in the scene 
2then push him or her c~to the CP 
3else create a new and indetennina~ concept and push 
it onto the CP. 
4.2,1, Discussion of branch 1. Branch I is taken when a 
subjective element continues the current subjective context. 
For example, the exclamation in O.2), which is a subjective 
element, continues the subjective context esteblished by 
(3.1). The subjective elements 'should' and *really' in (5.3) 
and 'After all', 'Aunt Margaret', 'that', and 'ghastly' in (5.4) 
continue the subjective context established in (5.1). 
4.2.2. Discussion of branch 2. The actor focus used in 
branch 2 is one of the foci that need to be maintained for the 
comprehension of definite snaphora (Sidner 1983). It is 
whoever is the agent of the current sente:r.e. (Note that 
quoted speech has its own foci. which must be maintained 
separately. In this sense, quoted speech constitutes a 
separate discourse segment (cf. Grosz and Sidne~ (1986).) 
Consider the following example: 
(10) le~XIn the kitchen she \[Jody\] set the basket down on 
the table, m'2She put the thermos and the cups in the 
sink and filled them with cool water to soak. xo-"rhen 
In stone third-pencn novels, panicularly in the 19th century, 
an overt narrator (Chamum 1978) rues mbjective elemems. We do 
not consider novels with overt narrators. 
she tiptoed upstairs to her room. l°~Perhaps Aunt 
Margaret was taking a nap. m3It wouldn't do to dis- 
nu'b her. \[Gage, Miss Osborne-the-Mop, p. 25\] 
Since Jody is the actor focus at the beginning of (10.4) (she 
is the actor focus of (10.1)-(10.3)), and she is a character in 
the scene, the mbjective element 'Perhaps' is attributed m 
her when it is encountered, and she is pushed onto the CP. 
Sidner (1983) has shown that, in anaphora 
comprehension, the current actor or discourse focus can he 
rejected as the co-apedfier of an anaphor on the basis of 
pragmatic factors. Similarly. the actor focus may be rejected 
as the subjective character to whom the subjective element is 
mributed, in favor of another character in the scene. The 
pragmatic factors involved appear to be which characters 
have been subjective characters in the past and whose 
thoughts or perceptions the sentence containing the subjec- 
tive dement is likely to be reflecting. Consider the follow- 
ing example, in which Adr~el, a seraph, has just appeared 
before Lemech and Sandy: 
(11) n'SLemech greeted him \[Adnarel\] respectfully. 
"Adnarel, we thank you." n:~rhen he said to Sandy, 
"The seraph will be able to help you. Seraphim 
know much about healing." 
n~So this was a seraph. \[L'Engle" Many Wmers, 
p. 39\] 
Lmnach is the actor focus of (11.1) and (11.2). However, it 
would be clear to someone who had read the novel up to this 
point that (11.3) is Sandy's thought. First, Sandy is a visitor 
to a slrange world, of which Lemech is an inhabitant; so it is 
Sandy, not Lemech, who is likely not to have known what a 
seraph is. Second, Irior to this passage, subjective contexts 
have been attributed to Sandy, but not to Lemech. We are 
investigating the reasoning required by the re~d_~ in reject- 
ins the cm'rem actor in favor of another character. 
4.2.3. Discussion of branch 3. Branch 3 is taken when the 
reader cannot identify the subjective character and must read 
further in the text in order to do so. In this case, an indemr- 
minate, inta~ional concept is pushed onto the CP. When 
the re~_d~ finally identifies the subjective character, the 
information that this character and the indeterminate one are 
co-extensional is built (that is, it is asserted that they are 
concepts of the same individual; cf. Malda and Shapiro 
(1982), Shapiro and Rapaport (1987)). Naicong Li (1986) 
uses this approach in her pronoun resolution algorithm if the 
infonna~n needed to resolve a lmmoun is supplied after the 
pronoun is encountered. 
Often, the subjective charact~ is identified by a nar- 
rative parenthetical, as in the following example: 
(12) 12JWhat was holding Washington up? the colonel 
asked himself. Iz2Ail he needed was a simple yes or 
no. \[clancy, Red Storm Rising, p. 170\] 
133 
Sentence (12.1) begins with a question, which is a subjective 
element. It occurs just after a shift in scene, so it does not 
continue a current subjective context, and there is no actor 
focus. When 'What' is encountered in (12.1), branch 3 
pushes a new concept onto the CP: At this point, the reader 
has recognized someone's thought' but does not yet know 
whose. When the reader encounters the parenthetical, she 
identifies the subjective character as the colonel, and builds 
a proposition that the colonel and the new concept are co- 
extemsional (that is, she comes to believe that the question 
was the colonel's thought). 
4.2.4. Comparison of evaluative and psychological 
adjectives. Before leaving oor discussion of subjective ele- 
ments, it will be useful to concast the ways that predicate- 
adjective sentences with psychological and with evaluative 
adjectives are treated by our discotwse process. Compare 
the following sentences: 
(A) Jody was delighted. 
(B) Jody was ghastly. 
Sentence (A) contains the psychological adjective 
'delighted', and (B) contains the evaluative adjective 
'ghastly'. In (A) (assuming no previous subjective context), 
Jody is pushed onto the CP, and (A) establishes a subjective 
context atlributed to Jody. In (B)' algorithm (A1) determines 
whose atti_-_~a_e toward Jody is being expressed, and it does 
not choose the subject of the sentence. Thus, psycbological 
adjectives can establish the perspective of the subject, 
whereas evaluative adjectives express an altitude toward the 
subject. 
$. RECOGNIZING ENDING BOUNDARIES OF SUB- 
JECTIVE CONTEXTS. Recognizing the ending boun- 
d&ies of subjective contexts is a more difficult problem than 
recognizing the beginning boundaries. While not all subjec- 
tive contexts are signaled in the ways discussed in Section 4, 
it is very common that they are. However, we have not 
found equally reliable or common signals for the ending 
boundaries. It appears that the reader often has to reason 
about the content of the current sentence and confirm that it 
can continue the subjective context; if it cannot, then the 
ending boundary has been found. 
Nevertheless, we have identified two reliable ways of 
recognizing the ending boundaries of subjective contexts. 
One way subjective contexts are ended is by a shift in scene, 
as in the following example: 
(13) He \[Sandy\] wanted to talk to Dennys. How were 
they going to be able m get home from this storage 
desert land into which they had been cast and which 
was heaven knew where in all the countless solar sys- 
tears in aLl the countless galaxies? 
<Chapter Break> 
Dennys was sleeping fitfully when he heard the tent 
flap move. \[L'Engle, Many Waters, pp. 91-92\] 
Dannys and Sandy are not at the same place. The shift in 
scene at the chapter break ends the subjective context attri- 
buted to Sandy. 
A second way is by a negated perceptual report 
whose subject is the subjective character and whose object is 
something in the scene. For example, 
(14) x4"xSbe \[Yalith\] was not sure why she was hesitant. 
x4~'he breathed in the strange odor of his wings, 
smelling of stone, of the cold, dark winds which came 
during the few brief weeks of winter. 
X~Envelopod in Eblis's wings, she did not hear the 
rhydmtic thud as a great lion galloped toward them 
across the desert, roaring as it neared them. 14~rhan 
both Yalith and Eblis turned and saw the lion rising 
to its hind legs... \[L'Engle, Many Wooers, p. 47\] 
The subjective context in the first paragraph is ended by the 
negated perceptual verb in (14.3). Sentence (14.4) then 
establishes a new subjective context attributed to Yalith and 
Eblis. In a similar way, subjective contexts can be ended by 
negated factive verbs, as in "He did not realize that... ". 
6. BELIlC~F AND SUBJECTIVE CONTEXTS. Since 
subjective contexts portray thoughts and perceptions, the 
reader understands that the information they convey reflects 
the subjective character's beliefs (cf. Fillmore (1974)' 
Banfield (1982), Usponsky (1973)). Whatever else the 
reader may infer that the characters believe, she has to attri- 
bute the informafi.__on in subjective contexts to the subjective 
ch&acter. Brian Reiser (1981) showed that one of the 
effects of perspective on a read_~'s understanding is that it 
focuses inec.essing. In particular, he showed that the read& 
primarily infers the goals and plans of the ch&acter whose 
perspective the narrative is taking. In a similar way, per- 
spective focuses the reader's atm'bution of beliefs to the 
characters. 
References in subjective contexts reflect just what the 
subjective character believes. (CL Clark and Marshall 
(1981), Cohen, Perranlt, and Allen (1982), and Wilks and 
Bien (1983) for discussions of belief and reference in 
conversation.) The subjective charecter might be mistaken, 
o~ know less about the referent than the reader or the other 
characters know, or know more than the other characters. 
The remainder of this paper addresses the auribution of 
beliefs to ch&acte~s in order to understand references in sub- 
jective contexts. 
6.1. An Algorithm for Understandin 8 References Using 
the CP. Ore- belief representation, described in Rapalx~ 
and Shapiro (1984), Rapaport (1986)' and Wiebe and Rapa- 
port (1986)' is based on the notion of belief spaces. A be//ef 
space is accessed by a stack of individuals, and consists of 
what the bottom member of the stack believes that.., the 
top member believes. The re~__aer_ is always the bottom 
member of the stack, and the belief space corresponding to a 
stack consisting only of the re~_d~ contains the set of 
134 
propositions that the reader believes are true. All proposi- 
tions in the knowledge base appear in at least one belief 
space, and a single proposition can appear in more than one 
belief space. This occurs, for example, if the reader believes 
a proposition and believes that a character believes it, too. 
The CP determines the current belief space with 
respect to which references are understood. So far, our 
analysis extends only to non-anaphoric, specific ref~er~ces. 
The following is our algorithm for understanding a non- 
anaphoric, specific reference 'X' in third-person mmmive 
(there may actually be more then one proposition found or 
built in order to understand 'X', for example, if 'X' is plural 
or a possessive): 
(A2) 
If 'X' is art indefinite noun phrase of the form 'a Y', 
tthen create a new concept, N; build in the CP's 
belief space the proposition that N is a Y; 
return N 
else if 'X' is a definite noun phrase or proper name, 
then if a proposition that N is X can be found in 
the CP's belief space, 
2then return N 
else if a proposition that N is X can be found in a 
belief space other than the CP's, 
3then add the found proposition to the CP's 
belief space; remm N 
4else create a new concept, N; build in the CP's 
belief space the proposition that N is X; 
retm'n N. 
6.1.1. Discussion of branch 1. Indefinite references intro- 
duce new individuals into the CP's belief space. We discuss 
this in Section 6.3, below. 
6.1.2. Discussion of branches 2 and 3. The search for the 
referent of a non-enaphoric definite noun phrase or proper 
name starts in the CP's belief space. Branch 2 is taken if 
the referent can be found there. 
If the test in branch 2 falls, then the rest of the 
knowledge base must be searched. To see why this is so, 
suppose that the reference is 'Ellen' and that it occurs in a 
subjective context. It is possible that Ellen has been referred 
to previously in the narrative but not under any cir- 
cumstences that would have required the reader to explicitly 
amibute the belief that she is named 'Ellen' to the subjective 
character. Perhaps she has only been referred to in objective 
contexts, for example. So, to find the referent, other belief 
spaces than the CP's must be searched. 
Branch 3 is taken if the search is successful, and, 
before the referent is returned, the proposition that the 
referent is X is A_dded to the CP's belief space. In the case 
just discussed, the fact that the reference occurs in a subjec- 
tive context indicates that the belief that the referent is 
named 'Ellen' should now be attributed to the subjective 
character. 
6.1.3. Discussion of branch 4. Branch 4 is taken in order 
to understand definite noun phrases and proper names that 
refer to individuals who have not been previously introduced 
into the narrative. Fur example, in (3.2), above, neither the 
mirror, the gifts, the baby, nor Ellen have been mentioned 
before in the novel. A new referent for each is introduced 
into the CP's belief space; that is, by virtue of understanding 
the references in (3.2), the reader comes to believe that Han- 
nah. the subjective character, believes that there is a mirror, 
some gifts, a baby, and a person named 'Ellen'. 
6.1A. An example. We now illustrate our algorithms on a 
passage that reflects a character's mistaken belief. The pas- 
sage is f~om a novel in which a character, Dwayne, mistak- 
enly believes that another character, Casey. is a boy: 
05) L~'IHis \[Dwayne's\] brain worked slowly through what 
he knew about this person \[Casey\]. tS'2David's kid. 
~The name stumbled into place. Is~nis was 
David's boy. ~S-SDavid was in the war, and here was 
his kid in the arc~_de scared of something. \[Bridgers, 
All Together Now, p. 91\] 
Note that (15.1) end (15.3) are psychological reports that 
employ metaphor, rather than psychological verbs, to report 
the character's psychological experience ((9.1), above, 
employs metaphor in a similar way). Memphur is beyond 
the scope of this work. so, before applying our algorithm to 
this passage, we paraphrase it as follows: 
(15a) L~He \[Dwayne\] thought of what he knew about this 
person \[Casey\]. tX2David's kid. XS~He remembered 
the name. ls~*\]'his was David's boy. lkSDavid was 
in the war, and here was his kid in the arcade scared 
of something. 
Fn'st, consider the operation of our discourse process. 
Sentence (15a.1) is a psychological report, and so Dwayne, 
its subject, is pushed onto the CP; this establishes a subjec- 
five context, auributad to Dwayne, which is continued 
throughout the passage. Note that when the sentence frag- 
ment (which is a subjective element) is encotmtered, no 
change is made to the CP because there is already a charac- 
ter on the top of it. Similarly, no change is made to the CP 
when (15a.3), a psychological report, and the second con- 
jun~ of (15a.5), a predicate-adjective umtew.e with a 
psychological adjective, are encountered, since there is 
already a character on the top of the CP. 
Now, consider the refe~e4-c.e to David in (15a.2). The 
reader knows that David is Casey's father. If, before read- 
ing (15L2), the reader didn't explicitly believe that Dwayne 
knew about David too, then branch 3 of algorithm (A2) 
would be taken to understand this reference; the result is that 
the reader now explicitly believes that Dwayne knows about 
David. 
135 
'David's boy' in (15a.4) reflects Dwayne's mistaken 
belief about Casey, and branch 2 of algorithm (A2) is taken 
in order to understand it. To illustrate that information in 
subjective contexts is atu'ibuted to the subjective character, 
suppose that (15a.4) were "This was David's girl"; in that 
case. the reader would have to infer that Dwayne had 
somehow found out that Casey is a girl. 
6.1.5. Further discussion of algorithm (A2)..Note that if 
a reference is a subjective elemertt, such as 'the bastard', it 
may be a non-classifw.atory noun (Banfield 1982); that is, it 
carmot be understood entirely propositionally, since it 
expresses subjectivity. How it should be undemtood 
depends on the particular subjective elmnmt. Thus, specific 
algorithms for nouns that are subjective elemants must 
supersede algorithm (A2). 
As mentioned above, algorithm (A2) is enable to 
understand anaphoric references. However, mmphor 
comprehension can he affected by perspective. Consider the 
following passage: 
(16) mtThe man had mined, xt'~He started to walk away 
quickly in the direction of the public library. 
le°"O.K.," said Joe, "get Rosie." 
x6~Zoe crept back to the blinker, xt~She felt hollow 
in her smmuch. 1~She'd never really expected to see 
the Enemy again. \[Ones\], War Work, p. 64\] 
In (16.6), 'the Enemy' is m ana~oric reference that occurs 
in a subjective context (established by (16.5), which is a 
psychological report); it co-specifies 'the man' in (16.1) and 
'He' in (16.2). It reflects Zoe's belief that the raan is an 
enemy spy, although it is not at all clear to the reader, at 
point' that he is. 
Personal pronouns can also reflect the beliefs of a 
character. The following passage is a continuation of pas- 
sage (15) (italics ours): 
(17) He \[Dwayne\] wasn't sure of what. What in the 
arcade could scare a boy like that? He rubbed his 
head under his baseball cap. He could see tears in 
Casey's eyes. He could tell they were tears 
h/s eyes were too shiny. Too round. WelL it was all 
right to cry. He'd cried when they took him to that 
place a few years back. Now Casey was in a new 
place, too, feeling maybe the same as him. If he just 
knew what to do about it. 
"let's don't play that game anymore," he said. "I 
don't like that one." 
Casey wiped her face on her sleeve... \[Bridgers, 
All Together Now, p. 92\] 
Both italicized pronouns refer m Cesey; the first occurs in a 
subjective context attributed to Dwayne, and the second 
occurs after the subjective context has ended (in this pas- 
sage, the subjective context is ended by direct speech). 
6.2, Assertive Indefinite Pronouns. Assertive indefinite 
pronouns--e.g., 'someone', 'something', 'somebody'----are 
specific, though unspecified (Quirk et al. 1985); that is, they 
generally refer to particular people, things, etc., without 
identifying them. When referring to a particular referent, a 
speaker typically uses an assertive indefinite pronoun if (1) 
she doesn't know the identity of the referent, (2) she doesn't 
want the addressee to know the identity of the reread-a, or 
(3) she doesn't believe that the identity of the referent is 
relevant to the conversation. A cheract~'s thoughts end 
perceptions are not directed toward an addressee, and so the 
first of these uses is the predominant one in subjective con- 
texts. Used in this way, they express a lack of knowledge, 
and so are subjective elements. When one of them appears 
in a subjective context, the r~d~ onderstands that the sub- 
jective character does not know who or what the refermt is. 
Often. the pronoun is the only source of this information. 
Consider the following example: 
(18) maSuddenly she \[Zoe\] gasped, n~She had touched 
somebody! \[Oneal, War Work., p. 129\] 
The~ is no explicit statement in the novel that Zoe does not 
know whom she touched; this has to be inferred from the 
use of 'somebody'. Sentences (6) and (15.5) provide fm'ther 
examples. 
6.3. Indefinite Refereneas. In conversation, definite refer- 
ances are used only if the speaker believes that the addressee 
has enough inform~on to intexpret them. As mentioned 
above, thoughts and perceptions are not directed toward m 
addressee, and so the use of definite references in subjective 
contexts is not subject to this conslraint; as illustrated by 
(3.2), they me used to refer to referents familiar to the sub- 
jective character, whether or not the r-,~ler has been told 
about them before. So, when a specific indefinite reference 
appears in a subjective context, the reader understands that 
the referont is unfamiliar to the subjective character;, other- 
wise, a defmlte reference would have appeared (Fdlmore 
1974). 
However, the referent may not be unknown to the 
reader or to the other characters. For example, 
(19) There they \[the King and his men\] saw close beside 
them a great rubbleheap; and suddenly they were 
aware of two small figures lying on it at their ease, 
grey-cled, hardly to be seen among the stones. \[Tol- 
kien, The Two Towers, p. 206\] 
The reader knows that the King and his men have come 
upon two hobbits, Menv end Pippin. The King and his men 
do not know the hobbits, but other characters also present in 
the scene do know them. When the King and his man are 
on the top of the CP (after 'saw' and continued by 'were 
aware of'), the hobbits are not referred to by name, but as 
'two small figures'. Branch 1 of algorirlun (A2) crea~ new 
referents and, in the belief space of the King and his men, 
136 
builds propositions that they are small figures. The new 
referants can be asserted to be co-exteusional with the con- 
cepts who the reader and other characters believe are named 
'Merry' and 'Pippin'. 
Indefinite references can sometimes indicate that the 
subjective character doesn't evan know what the refermt is. 
This occurs when the head noun is a SUlZ~ordinate, rather 
than a basic-level, term (Rosch and Lloyd 1978). The basic 
level is the preferred level at which people identify things. 
If a superordinate, rather than a basic-level, term appears in 
an indefinite refarance in a subjective context, the reader 
understands that the subjective character can't even idantify 
the referant at the basic level. In example (19), the hobbits 
are referred to as 'two small figures', because the King and 
his man have never seen hobbits before. Here is an example 
that is not from a fantasy novel: 
(20) Slowly Hannah raised her head and blinked her eyes. 
Small dots of purple coveted the ground around h~ 
and she reached out to explore. Violets l \[Fx~.-~che~, 
Hannah Herse~, p. 25\] 
Whan she first sees the violets, Hannah can only identify 
them as 'small dots of purple'. Another occurs in (2.2): 
The fact that the reference 'several large cylindrical objects' 
includes the ~dinate term 'objects' indicates that 
Muhammad doesn't know what the referents are. Another 
example is (21): 
(21) He felt firm reslralnts of some sort holding him in 
place. \[Wu, Cyborg, p. 141\] 
Peters and Shapiro (1987ab) describe a SNePS 
represantation for natural category systems in which superor- 
dinate categories can be distinguished from basic-level and 
subordinate categories. After an indefinite reference with a 
superordinate term in a subjective context has been parsed, 
the fact that the subjective character was able to identify the 
referent only at a supemrdinate level is represented in the 
knowledge base by using their represcntatio~ 
7. CONCLUSIONS AND FUTURE RESEARCH. Many 
problems remain to be solved. Our discourse process cannot 
recognize subjective contexts that are not established by the 
linguistic signals it relies on, and general principles are 
needed to explain how readers recognize the ending boun- 
daries of subjective contexts. We are investigating how 
tense, deictic terms (cf. Bruder et aL (1986), Bantield 
(1982)), the characters' goals (cf. Wilansky (1983)), and the 
argumant structure (of. Cohen (1987)) often exhibited by 
thoughts might be used to recognize the boundaries of sub- 
jectlve contexts. Branches 2 and 3 of algorithm (A1) need to 
be expanded to determine who the subjective character is if 
the actor focus isn't a reasonable candidate and no 
paranthetical appea~. We are investigating how focus of 
attention (cf. Grosz (1981), Sidner (1983)) can be incor- 
porated into algorithm (A2) in such a way that anaphoric 
references reflecting the beliefs of a character can be under- 
stood. Finally, there is the general problem of revision. Our 
algorithms assume that signals occur at the beginning of 
subjective contexts. However, there are cases when a sub- 
jective context cannot be recognized until some of it has 
already been parsed. A difficult case is inustrated by the fol- 
lowing: "CaJody was rich and famous, c2why wasn't she 
happy? Bin wondered." Only after reading (C2) can the 
reader recognize that (C1) is a represented thought. 
We have erguad that a discourse-level approach must 
be taken to the problem of recognizing character's thoughts 
and perceptions in third-person narrative. Our discourse 
process, which is implemented in an ATN grammar inter- 
faced to SNePS, recognizes subjective contexts that are 
linguistically signaled in ways frequantly employed in 
naturally-occurring narratives. By using the results of the 
discourse process to determine the belief context needed m 
onderstand refermces, our reference algorithm demonstrates 
how perspective affects referance in third-person narrative. 
8, ACKNOWLEDGMENTS. We are indebted to Mary 
Galbraith, David Zubin, Sandra Peters, Stuart Shapiro, and 
the other members of the SUNY Buffalo Graduate Group in 
Cognitive Science and the SNePS Research Group for many 
discussions end ideas. This research was supported in part 
by NSF grauts IST-8504713, IRI-8610517. 
REFERENCES 
Banfield, Arm (1982), Unspeakable Sentences: Narratwn 
and Representation in the Language of Fiction (Boston: 
Roufledge & Kegan Paul). 
Bruder, Gail A.; Duchan, Judy F.; R~papmt, W'dliam J.; 
Sagal, Erwin M.; Shapiro, Smart C.; and Zubin, David A. 
(1986), "Deictic Canters in Narrative: An Interdisciplinary 
Cognitive-Sciance Project," Technical Report 86.20 
(Buffalo: SUNY Buffalo Dept. of Computer Science). 
Chamm~ Seymour (1978), Story and Discourse: Narrative 
Structure in Fiction and Film (Ithaca, NY: Comell Univer- 
sity Press). 
Clark, Herbert H. end Marshall, Catharine R. (1981), 
"Definite Ref~,w,ce and Mutual Knowledge," in A. Joshi, 
B. Webber, and L Sag (eds.), Elements of Discourse Under- 
~and/ng (Cambridge: Cambridge University Press): 10-63. 
Cohen, Philip R.; Perrault, C. Raymond; and Allen, James F. 
(1982), "Beyond Question Answering," in W. Lchnert and 
M. Ringle (eds.), Strategies for Natural Language Process- 
/ng (Hillsdale, NJ: Lawrence Erlbamn): 245-274. 
Cohen, Robin (1987), "Analyzing the Structure of 
Argumentative Discourse," Computational Linguistics 13: 
11-24. 
Cohn, Dorrit (1978), Transparent Minds: Narrative Modes 
for Representing Consciousness in Fiction (Princeton: 
Princeton University Press). 
137 
Dole~el, Lubomir (1973)' Narrative Modes in Czech l..itera. 
tare (Toronto: University of Toronto Press). 
Fillmore, Charles (1974)' "pragmatics and the Description 
of Discourse," in C. Fillmore, G. Lakoff, and R. Lakoff 
(eds.), Berkeley Studiea in Syntax and Semantics I (Berkeley: 
University of California Dept. of Linguistics and Institute of 
Human I.e.aming): VI-V21. 
Grosz, Barbara J. (1981)' "Focusing and Description in 
Natural Language Dialogues," in A. Joshi, B. Webber, and 
L Sag (eds.), Elements of Discourse Understanding (Cam- 
bridge: Cambridge University Press): 84-105. 
Grosz, Barbara J. and Sidner, Candace L. (1986), "Amm- 
tion. Intantions, and the Slructure of Disoom'se," Computa- 
tional Linguistics, 12: 175-204. 
Lakoff, Robin (1974)' "Remarks on 'this' and 'that'," 
Papers from the Tenth Regional Meeting of the Chicago 
Linguistic Society (Chicago: Chicago Linguistic Society): 
345-356. 
Li, Nalcong (1986). "Pronoun Resolution in SNePS." 
SNe.RG Technical Note 18 (Buffalo: SUNY Buffalo Dept. of 
Computer Science). 
Maida, Anthony S. and Shapiro, Smart C. (1982)' "Inten- 
sional Concepts in Propositional Semantic Networks," Cog- 
hi:ire Science 6: 291-330. 
Quirk. Randolph; Greenbaum. Sidney;, Le_ec.h, Geoffxe~ end 
Svartvik. Jan (1985)' A Comprehensive Grammar of the 
English Language (New York: Longuum). 
Peters, S~adra L. and Shapiro, Stuart C. (1987a), "A 
Relxesentation for Natural Category Systems," Proceedings 
of the 9th Annual Conferotce of the Cognitive Science 
Society (Seattle) (Hillsdale, NJ: Lawrence Erlhaum): 379- 
390. 
Peters, Sandra L. and Shapiro, Smart C. (19871)), "A 
Representation for Natural Category Systems," Proc~d'mgs 
of the lOth International Joint Conference on Artificial Intel. 
ligence (1JCAI-87; Milan) (Los Altos, CA: Morgan Kanf- 
mann): 140-146. 
Rapaport, William J. end Shapiro, Stuart C. (1984)' 
"Quasi-Indexical Reference in Propositional Semantic Net- 
works," Proceedings of the lOth International Conference 
on Computational Linguistics (COLING-84 ; Stanford Univ.) 
(Monistown, NJ: Assoc. for Computational L,ingulstics): 
65-70. 
Rapaport, William L (1986)' "Logical Foundations for 
Belief Representation,'" Cognitiv e Science 10: 371-422. 
Reise~, Brian J. (1981), "Character Tracking and the Under- 
standing of Narrative," Proceedings of the 7th International 
Joint Conference on Artificial Intelligence (IJCAI-81; Van. 
couver) (Los Altos, CA: Morgan Kanfmarm): 209-211. 
Rosch, Eleanor and Lloyd, B3. (1978)' Cognition and 
Categorization (Hillsdale, NJ: Lawrence Erlbaum Associ- 
ates). 
Shapiro, Smart C. (1979)' "The SNePS Semantic Network 
Processing System," in N.V. Findler (ed.), Associative Net. 
worka (New York: Acedomic): 179-203. 
Shapiro, Smart C. and Rapapo~ William J. (1987), "SNePS 
Comidered as a Fully Intensional Propositional Semantic 
Network," in N. Cerc~m and G. McCalla (eds.)' The 
Knowledge Frontier (New York: Springer-Verlag): 262- 
315. 
Sidner, Cmdaze L. (1983)' "Focusing in the Comprehension 
of Definite Anaphora," in M. Brliy md R. Berwick (eds.), 
Computational Models of Discourse (Cambridge, MA: The 
Mrr Press): 267-330. 
Uspemky. Boris (1973), A Poetics of Composition (Berke- 
le~ University of California Press). 
Wiebe, Janyce M. and Rapalx~ W'dliam J, (1986). 
"Representing De Re and De Dicto Belief Reports in 
Discourse and Nan'afive'" Proc. IEEE 74: 1405-1413. 
Wilemky, Robert (1983)' Planning and Understanding 
(Reading, MA: Addison-Wesley). 
Willm, Yorick and Bien, Janusz (1983), "Befiefs, Points of 
View, and Multiple Environments," Cognitive Science 7: 
95-119. 
CITED TEXTS 
Allegro, John (1977), The Dead Sea Scrolls (Harmon&- 
worth, Eng.: Panguin). 
Bridgers, Sue Ellen (1979)'.AII Together Now (New York: 
Knopf). 
Clarity, Tom (1986), Red Storm Rising (New York: G.P. 
Pumam's Sons). 
Franchere" Ruth (1964)' Hannah Herself (New York: Tho- 
rues Y. Crowell). 
Gage, Wilson (1963)' Miss Osborne-the-Mop (Cleveland: 
World Publishing). 
L'Engle. Madeleine (1986)' Many Waters (New York, Dell 
PubUsh~). 
Oneal. Zibby (1971)' War Work (New York. Viking Press). 
Tolkian. LR.R. (1965)' The Two Towers (New York: Bal- 
lantine Books). 
Wu, William (1987), Cyborg (New York: Ace Books). 
138 
