Identifying Subjective Characters in Narrative 
Janyce M. Wiebe 
Department of Computer Science 
University of Toronto 
Toronto, Canada M5S IA4 
wiebe@ai.toronto.edu 
ABSTRACT 
Part of understanding fictional narrative text is 
determining for each sentence whether it takes 
some character's point of view and, if it does, 
identifying the character whose point of view is 
taken. This paper presents part of an algorithm for 
perfomling the latter. When faced with a sentence 
that takes a character's point of view, the reader 
has to decide whether that character is a previ- 
ously mentioned character or one mentioned in the 
sentence. We give particular consideration to sen- 
tences about private states, such as seeing and 
wanting, for which both possibilities exist. Our 
algorithm is based on regularities in the ways that 
texts initiate, continue, and resume a character's 
point of view, found during extensive examina- 
tions of published novels and short stories. 
i. INTRODUCTION. Part of understanding 
ficfiona~ narrative text is determining for each sen- 
tence whether it takes some character's point of 
view and, if it does, identifying the character 
whose point of view is taken. This paper 
addresses the latter. We show how structural 
regularities of third-person fictional narrative text 
can be exploited to perform this task. 
2. BACKGROUND° Our approach to point of 
view is based on Ann Banfield's (1982) theory of 
narrative sentences. Banfield characterizes the 
sentences of narration as objective or subjective. 
Objective sentences are those that objectively nar- 
rate events. Subjective sentences are those that 
present the consciousness of an experiencing char- 
acter within the story. They express characters' 
evaluations, emotions, judgments, uncertainties, 
beliefs, and other attitudes and affects. Kinds of 
subjective sentences include those that portray a 
character's thought or perception (represented 
thought or represented perception; Banfield 
1982), and those that report a character's private 
state such as seeing, wanting, or feeling ill---that 
is, some perceptual, psychological, or experiential 
state that is not open to objective observation or 
verification, l We call the character whose cons- 
ciousness is presented by such a sentence the sub- 
jective character of that sentence. 
The following passage illustrates different 
kinds of subjective sentences. 
(1) 
1"1"What are you doing in here?" 1"2Suddenly she 
\[Zoel was furious with him \[Joel. 
t.3,,Spying, of course." 
l'4"Well of all dumb things! 1"5I thought you ran 
away." l'6joe Bunch was awful. \[Oneal, War Work, p. 
130\] 
Sentence (1.2) is a private-state report: It reports 
Zoe's private state of being furious with Joe. Sen- 
tence (1.6) is a represented thought: It presents 
Zoe's thought and it expresses her evaluation of 
Joe (that he is awful). Represented perceptions 
appear in the following passage: 
z'~Certainly, Dennys thought, anything would be better 
than this horrible-smelling place full of horrible little 
pe°p~'e2There was a brief whiff of fresh air. 2"3A 
l_ 
glimpse of a night sky crusted with stars. LL'Engle, 
Many Waters, p. 25\] 
Sentence (2.1) is Dennys's represented thought 
and (2.2) and (2.3) are Dennys's represented per- 
ceptions. 
"l-bird-person narrative text is characterized 
by shifts in point of view: a single text can have 
objective sentences as well as subjective sentences 
attributed to different characters. 2 To determine 
the current point of view, readers cannot consider 
sentences in isolation. In many cases, it is only in 
t See Wiebe 1990 and the references therein for alternative 
categorization~ of subjective sentences. 
2 There are different kinds of point of view in narrative. The 
kind addressed here is psychological as opposed to, for example, spa- 
tial or temporal (Uspensky 1973). In addition, we do not consider 
texts that take the point of view of an overt narrator (Chatman 1978). 
1 401 
context that a sentence is subjective or has a par- 
ticular subjective character. Sentence (1.6) is such 
a sentence. 
We have developed an algorithm for recog- 
nizing subjective sentences and identifying their 
subjective characters in third-person fictional nar- 
rative text. The algorithm is based on regularities, 
found during our examination of passages from 
over forty novels and short stories, in the ways 
that texts initiate, continue, and resume a 
character's point of view. The rules of the algo- 
rithm were checked, by hand, on over four hun- 
dred pages from seven novels. We were able to 
categorize most exceptions according to particular 
problems that remain to be addressed, such as the 
effect of certain spatial and temporal discontinui- 
ties on the psychological point of view. These 
classes of exceptions, together with complete 
descriptions of the current algorithm and its 
implementation, can be found in Wiebe 1990. A 
preliminary version of the algorithm was 
presented by Wiebe & Rapaport 1988. 
3. IDENTIFYING THE SUBJECTIVE CHARAC- 
TER. 
3.1. Introduction. Black, Turner, & Bower (1979) 
claim the following result of their empirical inves- 
tigation of point of view in narrative: "Merely 
making a character the subject of the narrative 
statement sufficed to establish his as the dominant 
point of view" (p. 187). However, their stimulus 
materials were short, artificially constructed narra- 
tives. For extended texts (i.e., novels and short 
stories), this simple rule is inadequate. 
The subjective character of a subjective sen- 
tenee is sometimes identifiable from the sentence 
itself. This is the case if the sentence has a narra- 
tive parenthetical (such as 'Dennys thought' in 
(2.1)); it might also be the case if the sentence 
denotes a private state, but not always. If the sub- 
jective character is not identifiable from the sen- 
tence, it is often one of two previously mentioned 
characters: the subjective character of the previ- 
ous subjective sentence or the actor of an action 
denoted by a previous objective sentence. The 
availability of these characters is conditional; 
obviously, for example, the former is available 
only if a subjective sentence has previously 
appeared. The subjective character also may not 
be identifiable when the sentence appears. 
Thus, identifying the subjective character 
involves determining whether it can be identified 
from the sentence itself, or, if not, determining 
whether it can be identified from the previous con- 
text. In addition, as the text is processed, informa- 
tion that might be needed to identify later subjeco 
tive characters must be gathered. 
3.2. Identifying the Subjective Character from the 
Previous Context. A subjective character who is 
not identifiable from the sentence itself (see Sec- 
tion 3.3, below) is most often the subjective char~ 
acter of the last subjective sentence. In this case, 
the current sentence continues a character's point 
of view (if the previous sentence was also subjec- 
tive) or resumes one (if objective sentence.s have 
appeared since the last subjective sentence). Sen- 
tences (2.2) and (2.3) illustrate the former, and 
sentence (1.6) illustrates the latter; it has the same 
subjective character as (1.2), the last subjective 
sentence, and is separated from (1.2) by objective 
sentences ( 1.3)--(1.5). 
Less commonly, the subjective character is 
the actor of an action denoted by a previous objec- 
tive sentence. Since this character need not be the 
subjective character of the last subjective sen- 
tence, this is a way to initiate a new point of view. 
Following is an example: 3 
(3) 
3"IJake felt sour. 3"2He wished again that cir- 
cumstances hadn't prompted him to come back. 3"3He 
had already spent one full night on horseback, 3"4and 
now the boys were expecting him to spend another, all 
on account of a bunch of livestock he had no interest in 
in the first place. 
3"5"I don't know as I'm coming," he said. 3"6"I 
just got here. 3"7If I'd known you boys did nothing but 
chase horses around all night, I don't know that I would 
have come." 
3's"Why, Jake,3.Yl~U lazy bean," Augustus said, 
3"9and walked off. • 'Jake had a stubborn streak in 
him, 311and once it was activated even Call could sel- 
dom do much with him. \[McMurtry, Lonesome Dove, 
p. 162\] 
As this passage appears in the novel, (3.10) and 
(3.1 I) represent the point of view of Augustus, the 
3 Numbering within passages reflects the actual units of input 
given to the algorithm, which are not always sentences. E.g., the con- 
juects of compound sentences are separate input units. Even so, an 
input unit can still have multiple clauses, each of which denotes an 
event. The algorithm uses heuristics to choose the most important of 
these for tracking point of view (most often, it chooses the event 
denoted by the main clause). Thus, when we refer to a sentence that 
denotes an action, for example, we actually mean an input unit of 
which the event chosen for consideration is an action. 
2 402 
actor of an action denoted by a previous objective 
sentence, (3.9). But the last subjective character is 
~lake, so Augustus's point of view is initiated, not 
merely resumed or continued. 
In order to identify the subjective character 
in these situations, the algorithm keeps track of 
expected subjective characters, characters who are 
likely to become the subjective character. Often, 
the last subjective character is one of these. How- 
ever, drastic spatial and temporal discontinuities 
can block: the continuation or resumption of a 
character's psychological point of view. We con- 
sider one kind of discontinuity, a scene break, 
which is a break from one parallel story-line to 
another (Almeida 1987). Thus, the last subjective 
character :is an expected subjective character only 
if a subjective sentence has appeared in the current 
scene (see Wiebe 1990 for examples). 4 
The usual situation for a shift in point of 
view to an actor is one in which the action of a 
previous subjective character is focused by the 
text---specifically, when the objective sentence 
denoting the action appeared earlier in the current 
paragraph and the entire paragraph so far has been 
objective. "5 In these situations, the actor of the last 
such senl~ence, the last active character, is an 
~'.xpected subjective character. 
If the subjective character is not identifiable 
from the sentence~ then the algorithm chooses an 
expected subjective character, if them is one. If 
there isn't one, then the algorithm cannot identify 
the subjective character at this point in the text 
(see Wiebe & Rapaport 1988 and Wiebe 1990 for 
~;xamples i~x which the reader must read further in 
the text in order to identify the subjective charac° 
ter). In the case of competition between expected 
subjective ctmracters, the last active character is 
the prefenred choice, since he or she is more 
highly focused by the text than the last subjective 
character as. The last subjective character is 
chosen only if the sentence is about the last active 
character. 6 The choice of the last subjective char- 
acter is incorrect if the sentence is the last active 
character's self-reflection or self-perception; this 
n The algorithm cannot detect scene breaks; rather, it nmst be 
informed when one occurs. 
5 In addition, the action must be one that is actually performed 
hi the current scene; thus, the main verb phrase of the sentence denot- 
h~g it must pass certain linguistic tests, for example, that it not con- 
~ain a habitual adverbial such as 'usually' or a modal adverbial such 
as 'probably'. 
That is, if the last active character is the actor or experieneer 
of the event chosen for consideration. 
heuristic relies on the relative infrequency of sub- 
jective sentences about oneself. 
3.3. Identifying the Subjective Character from the 
Sentence. Most commonly in third-person narra~ 
tire text, subjective sentences continue or resume 
characters' points of view, rather than initiate new 
ones. The previous section discussed one way, 
illustrated by passage (3), to initiate a new point 
of view. In situations like those, the subjective 
character is determined from the previous context. 
A subjective sentence is also able to initiate a new 
point of view if the subjective character is 
identifiable from the sentence itself. A simple 
example is a sentence with a narrative parentheti- 
cal, such as (2.1)--the subjective character is 
always the subject of the parenthetical. 
The subjective character may also be 
identifiable from a sentence denoting a private 
state (a private-state sentence); in particular, it 
may be the experiencer of the private state 
denoted. Thus, as noted by Dole~el 1973, Cohn 
1978, and Banfield 1982, a private-state sentence 
is another way to initiate a new point of view. 7 An 
example occurs in the following: 
(4) 
4"l"Drown me?" Augustus said. 4"2"Why if' any- 
body had tried it, those girls would have clawed them 
to shreds." 4"3He knew Call was mad, 4"4but wasn't 
much inclined to humor him. 4"5It was his dinner table 
as much as Call's, 4"6and if Call didn't like the conver- 
sation he could go to bed. 
47 48 • Call knew there was no point in arguing. • That 
was what Augustus wanted: argument. "9~Ie didn t 
really care what the question was, 4"l°and it made no 
great difference to him which side he was on. n'llHe 
just plain loved to argue. \[McMurtry, Lonesome Dove, 
p. 16\] 
Sentences (4.3)-(4.6) are Augustus's subjective 
sentences and (4.7)-(4.11) are Call's. So, (4.7) 
initiates a new point of view. It is a private-state 
sentence and the subjective character, Call, is the 
experiencer of the plivate state denoted. Passage 
(4) shows, however, that the subjective character 
of a private-state sentence need not be the 
experiencer. In (4.6), for example, "Call didn't 
like the conversation" denotes a plivate state (lik° 
ing), but the subjective character of the sentence is 
7 If a subjective sentence denotes a private-state action such as 
sighing or looking, the subjective character might also be identifiable 
from the sentence itself. See Wiebe 1990 for the algorithm's treat- 
ment of these kinds of sentences. 
3 403 
Augustus, not Call. In the following subsections, 
we will consider factors that can indicate that the 
subjective character of a private-state sentence is 
not the experiencer. 
3.3.1. Textual Continuity. Point of view does not 
typically shift from one character to another 
without a paragraph break. 8 Thus, the lack of a 
paragraph break suggests that a shift has not 
occurred. In the case of a private-state sentence 
following a subjective sentence, where the 
experiencer is different from the last subjective 
character, the lack of a paragraph break between 
them suggests that the private-state sentence is not 
the experiencer's subjective sentence; if it were, 
then a shift would have occurred without a para- 
graph break. So, when a private-state sentence 
immediately follows a subjective sentence without 
a paragraph break (this is called the continuing° 
subjective text situation), the algorithm identifies 
the subjective character to be the last subjective 
character, not the experiencer of the private state. 9 
3.3.2, Subjective Elements. Another factor that 
suggests that a private-state sentence should be 
attributed to an expected subjective character 
rather than to the experieneer is the appearance of 
a subjective element. 
There are many linguistic elements (syntac- 
tic, lexical, and morphological) that can indicate 
that a sentence is subjective. Examples are 
evaluative terms such as 'the old bag' (Banfield 
1982), and evidentials such as 'evidently' and 
'apparently' (Dole~el 1973). Although some of 
these elements always indicate that a sentence is 
subjective, others only potentially do so. A sub- 
jective element is a linguistic element that actually 
indicates that a sentence is subjective in the con- 
text of use. 10 
Before we address the effect of subjective 
elements on identifying the subjective character of 
a private-state sentence, we need to consider 
interpretations of private-state sentences. There 
s This observation is predicted by Nakhimovsky's work on the 
discourse structure of narrative text, in which it is suggested that 
paragraph breaks accompany discontinuities (Nakhimovsky & Rapa- 
port 1988; Naldaimovsky 1988). 
9 Actually, the algorithm allows for a broadening or narrowing 
of point of view upon a private-state sentence. This can occur be- 
cause, as shown by Banfield 1982, a subjective sentence can be attri- 
buted to a set of characters. 
10 We borrow this term from Bantield 1982, but redefine it; 
Banfield uses it to refer to linguistic elements that always indicate 
that a sentence is subjective. 
are two subjective interpretations of private-state 
sentences: represented thoughts and private-state 
reports. 11 Interpreted to be a represented thought, 
a private-state sentence is someone's thought 
about a character's private state. For example, 
(4.6) is Augustus's represented thought that is 
about Call's private state (Call's not liking some- 
thing). A private-state sentence that is interpreted 
to be a private-state report, on the other hand, is a 
report of a character's private state; it is not 
someone's thought about it. Examples are (4.3) 
and (4.7). 12 
The subjective character of a private-state 
report is always the experiencer. Thus, the sub- 
jective character is identifiable from the sentence 
itself. On the other hand, the subjective character 
of a represented thought cannot be identified from 
the sentence itself, but depends on the context 
(unless, of course, a narrative parenthetical 
appears). This is so regardless of whether the sub- 
jective character is referred to in the sentence. For 
example, the subjective character of the following 
represented thought is the referent of 'she' (as the 
sentence appears in the short story): 
Why couldn't she? \[Mansfield, "The Garden Party"\] 
but the subjective character of the folk)wing (as 
the sentence appears in the short sto~y) is not 
referred to in the sentence: 
What nice eyes he had, small, but such a dark blue! 
\[Mansfield, "The Garden Party"\] 
Thus, if something indicates that a private-state 
sentence is a represented thought rather than a 
private-state report, then the subjective character 
should be identified to be an expected subjective 
character (as discussed in Section 3.2). 
Subjective elements do this, as in tile follow- 
ing passage; at the start of the passage, Sandy and 
Dennys are (collectively) the last subjective char- 
acter: "13 
n There is also an objective interpretation (see Wiebe 1990). 
tz Cohn 1978 does not acknowledge this ambiguity: She says 
that represented thought can be distinguished from plivate-state re- 
ports by "the absence of mental verbs" in the former (p. 104). 
13 According to the rules of the algorithm, the event of (5.1) 
chosen for consideration is the private state denoted by the subordi- 
mated clause. 
404 4 
(5) 
5'l Japheth, evidently realizing that they were no 
longer behind him, turned around 5'2and jogged back 
toward them, seemingly cool and unwinded. \[L'Engle, 
Many Waters, p. 24\] 
"l~he subjective element 'evidently' in (5.1) indi- 
cates that the sentence is not a private-state report. 
That is, (5.1) is not a report that Japheth realizes 
that they are no longer behind him. Instead, 
Sandy and Dennys (the subjective character) 
ascribe this private state to him. 
However, subordinated subjective elements, 
those within the scope of the private-state term, 
can appear in private-state reports. 14 Thus, they 
cannot 'be used to distinguish private-state reports 
from represented thoughts (and so cannot be used 
to identify the subjective character). For example: 
 61)Ugh! she \[the girl\] thoughL 6"2How could the poor thinghave married him in the first place? 
~'3johnnic Martin could not believe that he was 
seeing that old bag's black eyes sparkling with disgust 
and unsheathed contempt at him. \[CaldweU, No One 
Hears But Him, pp. 98-99\] 
Sentence (6.3) is a private-state report and the sub- 
jective ,character is the experiencer (Johnnie Mar- 
tin); this is so even though (6.3) contains the sub~ 
jeetive element 'old bag' and even though there is 
an expected subjective character (the girl) when it 
is encountered. Because 'old bag' appears within 
the scope of the private-state term 'believe', it is 
not considered in identifying the subjective char° 
acter. On the other hand, the subjective clement 
'evidently ' in (5.1) is not in the scope of 'realiz- 
ing' (i.e., it is non-subordinated), so it can be used 
to identify the subjective character. 
If a private-state sentence does not have a 
non-subordinated subjective element and does not 
appear in the continuing-subjective situation, then 
the algorithm identifies the expected subjective 
characmr to be the experiencer. 
4. SUMMARY. Following is the rule of the algo- 
rithm for identifying the subjective character of a 
subjective sentence ('SC' stands for the subjective 
character of the current sentence): 15 
14 qhis is one reason why we define private-state reports to be 
subjective. 
15 Provisions for the following have been omitted for brevity: 
broadening and narrowing of point of view, sentences denoting 
private-state actions, and private-state sentences that can be objective 
To identify SC: 
If the sentence contains a narrative parenthetical then 
SC is the subject of the parenthetical 
else if the sentence is a private-state sentence then 
if it has a non-subordinated subjective clement 
or the text situation is continuing-subjective then 
SC is identified from the previous context 
else SC is the experiencer 
end if 
else SC is identified from the previous context 
end if 
To identify SC from the previous context: 
If there are two expected subjective characters then 
if the sentence is about the last active character then 
SC is the last subjective character 
else SC is the last active character 
end if 
else if there is an expected subjective character then 
SC is the expected subjective character 
else SC is unidentified 
end if 
5. CONCLUSION. We am extending this work 
along two avenues. First, we are developing 
psychological experiments to test whether the 
regularities on which the algorithm is based 
influence the reader's recognition of subjective 
sentences and identification of subjective charac- 
ters. Second, we are extending the algorithm to 
make connections with work on focus of attention 
and discom'se structure (such as Grosz & Sidner 
1986); in particular, we are investigating how 
resolving anaphora and tracking the current point 
of view are related (Stark 1987, Hewitt 1988). An 
important direction for future research is reason- 
ing about the plausibility of a suggested interpre- 
tation, that is, whether it is plausible that the con° 
tent of a subjective sentence is a particular 
character's thought or perception. 
We lmve presented part of an algorithm for 
identifying subjective characters that is based on 
regularities in the ways that texts initiate, resume, 
and continue a character's point of view. When 
faced with a subjective sentence, the reader has to 
decide whether the subjective character can be 
identified from the sentence itself. It can be if the 
sentence is a private-state report, but not if it is a 
represented thought; we give rules for distinguish- 
ing between these two interpretations of private- 
state sentences that involve the textual situation 
and the appearance of subjective elements in the 
sentence. A subjective character who is not 
(see Wiebe 1990). 
5 405 
identifiable from the sentence itself is often one of 
two previously-mentioned characters; we identify 
those characters and specify when one of them 
should be chosen to be the subjective character. 
Our findings are based on extensive examinations 
of natural narrative. 
Acknowledgments. I wish to thank the members 
of the SUNY Buffalo Graduate Group in Cogni- 
tive Science and the SNePS Research Group for 
many discussions and ideas, and William Rapa- 
port, Graeme Hirst, and Diane Horton for helpful 
comments on earlier drafts of this paper. 

REFERENCES 

Almeida, M. J. (1987), Reasoning about the Tem- 
poral Structure of Narratives, Ph.D. dissertation, 
published as Technical Report 87-10 (Buffalo: 
SUNY Buffalo Department of Computer Science). 

Banfield, A. (1982), Unspeakable Sentences: 
Narration and Representation in the Language of 
Fiction (Boston: Routledge & Kegan Paul). 

Black, J. B.; Turner, T. J.; & Bower, G. H. (1979), 
"Point of View in Narrative Comprehension, 
Memory, and Production," Journal of Verbal 
Learning and Verbal Behavior 18: 187-198. 

Caldwell, Taylor (1966), No One Hears But Him 
(Garden City, NY: Doubleday). 

Chatrnan, Seymour (1978), Story and Discourse: 
Narrative Structure in Fiction and Film (Ithaca, 
NY: Cornell University Press). 

Cohn, D. (1978), Transparent Minds: Narrative 
Modes for Representing Consciousness in Fiction 
(Princeton: Princeton University Press). 

Dole~el, L. (1973), Narrative Modes in Czech 
Literature (Toronto: University of Toronto 
Press). 

Grosz, B. J. & Sidner, C. L. (1986), "Attention, 
Intentions, and the Structure of Discourse," Com- 
putational Linguistics, 12: 175-204. 

Hewitt, L. (1988), "Reduced Anaphor in Subjec- 
tive Contexts," unpublished manuscript, SUNY 
Buffalo Graduate Group in Cognitive Science. 

L'Engle, M. (1986), Many Waters (New York, 
Dell Publishing). 

McMurtry, L. (1985), Lonesome Dove (New York: 
Simon & Schuster). 

Nakhimovsky, A. (1988), "Aspect, Aspectual 
Class, and the Temporal Structure of Narrative," 
Computational Linguistics 14 (2): 29-43. 

Nakhimovsky, A. & Rapaport, W. J. (1988), 
"Discontinuities in Narratives," Proceedings of 
the 12th International Conference on Computa- 
tional Linguistics (COLING-88, Budapest) 
(Morristown, NJ: Assoc. for Computational 
Linguistics): 465-470. 

Oneal, Z. (1971), War Work (New York: Viking 
Press). 

Stark, H. A. (1987), "Keeping Track of Charac- 
ters in Narrative," Ph.D. dissertation, Dept. of 
Computer Science, Stanford University. 

Uspensky, B. (1973), A Poetics of Composition 
(Berkeley: University of California Press). 

Wiebe, J. M., & Rapaport, W. J. (1988), "A Com- 
putational Theory of Perspective and Reference in 
Narrative," Proceedings of the 26th Annual Meet- 
ing of the Assoc. for Computational Linguistics 
(ACL-88; SUNY at Buffalo) (Morristown, NJ: 
Assoc. for Computational Linguistics): 131-138. 

Wiebe, J. M. (1990), Recognizing Subjective Sen- 
tences: A Computational Investigation of Narra- 
tive Text, Ph.D. dissertation, published as a techni- 
cal report (Buffalo: SUNY Buffalo Dept. of Com- 
puter Science). 
