<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-2004">
  <Title>Tracking Point of View in Narrative</Title>
  <Section position="4" start_page="237" end_page="238" type="metho">
    <SectionTitle>
4. Approach
</SectionTitle>
    <Paragraph position="0"> Reasoning about whether a sentence is objective or the subjective sentence of this or that character is certainly part of tracking POV (Fillmore 1974). But it is reasonable to hypothesize that, in the face of all of the inferential possibilities, discourse expectations too are involved in tracking POV; that is, in the absence of an explicit indication of POV, readers are intended to assume that POV is being manipulated in one of the usual ways, and to try to interpret the sentence accordingly. (Similar suggestions have been made by Carberry \[1989\] with respect to resolving intersentential ellipsis and by Sidner \[1983\] with respect to pronoun resolution.) We can view the text as composed of maximal blocks of objective sentences (objective contexts) and maximal blocks of subjective sentences that have the same subjective character (subjective contexts). Fur1 My wording in this paper often attributes agency to sentences. I might say, for example, that a sentence states something, communicates something, initiates a new POV, or refers to someone. Sidner (1983) and Webber (1983) object specifically to using a noun phrase as the agent of the verb 'refer,' since it is the writer who is doing the referring, not the noun phrase. I do not disagree---my wording is for convenience only.</Paragraph>
    <Paragraph position="1">  Janyce M. Wiebe Tracking Point of View in Narrative ther, we can view the process of tracking POV as recognizing the following discourse operations: .</Paragraph>
    <Paragraph position="2"> .</Paragraph>
    <Paragraph position="3"> .</Paragraph>
    <Paragraph position="4"> Sentence s continues the current POV: s and the previous sentence are either both objective or both subjective sentences of the same character. Sentence s resumes y's POV: s is a subjective sentence of y and is preceded by an objective context, which is in turn preceded by a subjective context of the same character y.</Paragraph>
    <Paragraph position="5"> Sentence s initiates y's POV: s is a subjective sentence of y and either s is the first subjective sentence of a scene, or the SC of the previous subjective sentence is a different character z.</Paragraph>
    <Paragraph position="6"> The approach taken in this work is to seek, by extensive examination of naturally occurring narratives, regularities in the ways that authors initiate, resume, and continue a character's point of view, and to develop an algorithm that tracks point of view on the basis of the regularities found. Given certain combinations of sentence features (e.g., tense, aspect, lexical items that potentially express subjectivity, the types of states of affairs that the sentence is about, and the identities of the actors or experiencers of those states of affairs), and of the current context (e.g., whether the previous sentence was subjective or objective, whether a paragraph break separates the current and previous sentences, and the identity of the SC of the previous subjective sentence, if there was one), particular point-of-view operations can be expected. A simple example: a sentence that (1) is about an action, (2) is in the past progressive, and (3) follows, without a paragraph break, the subjective sentence of a character c, usually continues c's POV.</Paragraph>
    <Paragraph position="7"> The examination of texts mentioned above was not a formal empirical investigation, so the algorithm should be viewed as a hypothesis that can be subjected to such tests. However, I examined passages from over forty novels and short stories to identify the regularities upon which the algorithm is based, and strictly adhered to the practice of considering only naturally occurring examples in developing the algorithm. Further, some preliminary empirical tests of the algorithm have been performed (see Section 12), among them psychological experiments of specific aspects of the algorithm. The results of these experiments are positive. I describe them as &amp;quot;preliminary&amp;quot; so as not to suggest that the entire algorithm has already been subjected to psychological experimentation.</Paragraph>
  </Section>
  <Section position="5" start_page="238" end_page="239" type="metho">
    <SectionTitle>
5. Previous Work
</SectionTitle>
    <Paragraph position="0"> There are literary theorists and linguists who investigate linguistic aspects of subjective sentences. The present work greatly benefited from their investigations, most directly from Dole~el (1973), Uspensky (1973), Kuroda (1973, 1976), Fillmore (1974), Cohn (1978), and especially Banfield (1982). However, the relevant work in the above fields is descriptive only; it describes characteristics of subjective sentences, but does not address the problem of tracking POV. An exception is work on POV and aspect that shows that aspect is only a context-sensitive marker of subjectivity (Ehrlich \[1987, 1990\] and Caenepeel \[1989\]; see Section 9).</Paragraph>
    <Paragraph position="1"> In AI, Nakhimovsky (1988) suggests a discourse-processing approach to tracking POV, but does not develop it in any depth. Also, Reiser (1981) simply suggests that POV may be established by syntactic clues, and by including &amp;quot;more episodes and internal information&amp;quot; about a character (p. 209).</Paragraph>
    <Paragraph position="2">  Computational Linguistics Volume 20, Number 2 con text ~ l{ }, { }, { }, presubjective-nonactive / loop if ~SENTENCE(ITEM( text, i) ) then context ~- NEW-CONTEXT' (ITEM(text, i), context) else interpretation ~ POV (FEATURES(ITEM( text, i) ),context) context ~ NEw-CONTEXT(interpretation, context)</Paragraph>
    <Paragraph position="4"> Figure 1 Overview. A great deal of work in AI has involved inferring speaker and hearer attitudes in conversation (such as work in plan recognition in conversation--see, e.g., Allen and Perrault \[1980\]; in user modeling--see, e.g., the papers in the special issue of Computational Linguistics on user modeling 1988; and in dynamically constructing nested-belief environments during understanding--see, e.g., Wilks and Bien \[1983\]) and character's attitudes in stories (see, e.g., Wilensky \[1983\] and Dyer \[1983\]). Such beliefs might be about what agents who are mentioned in the discourse believe, but the question is not addressed as to whether an utterance itself presents what is actually the object of an agent's attitude, where that agent is not mentioned in the sentence.</Paragraph>
    <Paragraph position="5"> In summary, there is no previous detailed investigation of the problem of tracking the psychological POV.</Paragraph>
  </Section>
  <Section position="6" start_page="239" end_page="249" type="metho">
    <SectionTitle>
6. Overview of the Algorithm
</SectionTitle>
    <Paragraph position="0"> The algorithm will be developed in detail in subsequent sections. This section provides an overview, giving the input and output of the basic components. While this involves using some terms before they are defined, it will provide the reader with a framework in which to understand the material that follows. As well, it will enable me to clarify the focus of the work and to state exactly what has been implemented.</Paragraph>
    <Paragraph position="1"> Figure I gives the algorithm at the highest level and a corresponding flow-diagram.</Paragraph>
    <Paragraph position="2"> Of the functions shown in Figure 1, it is functions POV, NEW-CONTEXT, and NEW-CON-- null Janyce M. Wiebe Tracking Point of View in Narrative TEXT' that are addressed in this work. Functions ITEM and FEATURES, preprocessing functions, represent a subset of the other tasks performed by an overall NLU system. The remainder of this section gives the input/output mappings of all five functions.</Paragraph>
    <Section position="1" start_page="240" end_page="241" type="sub_section">
      <SectionTitle>
6.1 The Preprocessing Functions
</SectionTitle>
      <Paragraph position="0"> * Function ITEM maps the text and the current position in the text into the input item at that position, i.e., the paragraph break, scene break (see Section 8.2.1), or sentence (see Section 6.3.2) that is at the current position: ITEM : Text x Position --+ Inputltem.</Paragraph>
      <Paragraph position="1"> This function is not implemented in the system, so a facility is provided that enables the user to input the current input item.</Paragraph>
      <Paragraph position="3"> A FeatureSet consists of the following (some of the features will be expanded upon in later sections, as indicated; for extensive detail, see Wiebe \[1990\]):</Paragraph>
      <Paragraph position="5"> The potential subjective elements in the sentence, if any. The potential subjective elements form a large class that includes lexical items used with particular meanings, tense, aspect, and certain syntactic properties (see Section 9.2).</Paragraph>
      <Paragraph position="6"> The type of state of affairs that each clause is about. For example, the main clause of the following is about a private state and the subordinated clause is about an action: John wondered whether Mary opened the door.</Paragraph>
      <Paragraph position="7"> Note that each of the following is about a private state: Mary was afraid of the dark.</Paragraph>
      <Paragraph position="8"> The darkness made Mary afraid.</Paragraph>
      <Paragraph position="9"> Wiebe (1990) gives a list of private-state terms, noting which syntactic roles the experiencer fills in various clause structures (this material is drawn from Quirk et al. \[1985\]). Section 6.3.1 gives the categorization of states of affairs used in this work.</Paragraph>
      <Paragraph position="10"> An indication of whether or not the head noun of the subject of the main clause is a private-state noun (e.g., 'pain' and 'astonishment') and, if it is, the state of affairs that that noun is about. Examples of sentences with such private-state nouns are:  (10) The feeling went away.</Paragraph>
      <Paragraph position="11"> (11) The pain increased.</Paragraph>
      <Paragraph position="12"> (12) His astonishment grew.</Paragraph>
      <Paragraph position="13"> 4. The experiencers and actors (i.e., particular case fillers \[Fillmore 1968\]) of  the states of affairs of items (2) and (3). Because a single action might have more than one actor, and a single state might have more than one experiencer, each of these is a set. If an experiencer or actor is not mentioned in the sentence (as is the case for the subject nouns in (10) and (11)), then that experiencer or actor is the empty set.  Computational Linguistics Volume 20, Number 2 .</Paragraph>
      <Paragraph position="14"> .</Paragraph>
      <Paragraph position="15"> An indication of whether or not the sentence contains a narrative parenthetical, and, if it does, the identity (or identities) of the individual(s) referred to by the subject of the parenthetical. Some other syntactic information not included above.</Paragraph>
      <Paragraph position="16"> Function FEATURES involves the resolution of syntactic, semantic, and discourse/pragmatic ambiguities that are outside the scope of this work (see Section 14 for a discussion of interactions between point of view and discourse processing). While the implemented system demonstrated in Appendix B takes actual sentences as input, it does not truly implement FEATURES, but is successful in computing the FeatureSet of a sentence only for sentences that fall within its limited coverage. There is another version of the system that queries the user for the information returned by function FEATURES, to enable the algorithm to be tested on unlimited text, without concern for problems not addressed in this work. It is this system that is used in the test of the algorithm presented in Appendix C and to support the psychological experiments mentioned in Sections 8.3.1, 9.3, and 12.</Paragraph>
    </Section>
    <Section position="2" start_page="241" end_page="246" type="sub_section">
      <SectionTitle>
6.2 The Central Functions
</SectionTitle>
      <Paragraph position="0"> * Function POV, which is implemented, maps a FeatureSet and a Context into an Interpretation: null POV : FeatureSet x Context ~ Interpretation.</Paragraph>
      <Paragraph position="1"> A Context and an Interpretation are as follows. A Context consists of (1) the identity of the SC of the last subjective sentence that appeared in the text, if there was one, (2) the identity of the last active character (defined in Section 8.2.2), if there was one, (3) the identities of any characters whose points of view were taken earlier in the text, and (4) the current text situation (defined just below in Section 6.3.4):</Paragraph>
      <Paragraph position="3"> An Interpretation is either that the sentence is the subjective sentence of a particular character, or that the sentence is objective and has a particular active character (ActiveCharacter is the empty set for objective sentences without active characters): Interpretation E {{objective, ActiveCharacter} \] ActiveCharacter C Characters} U {{subjective, SC} i SC c Characters}.</Paragraph>
      <Paragraph position="4"> * Functions NEW-CONTEXT and NEW-CONTEXT' return the Context of the next input item, the former for the case in which the current item is a sentence; the latter for the case in which the current item is a scene or paragraph break:</Paragraph>
      <Paragraph position="6"> These functions are also implemented; algorithms for them follow trivially from the definitions of an interpretation and of a context and its components.</Paragraph>
      <Paragraph position="7">  Janyce M. Wiebe Tracking Point of View in Narrative The Context of the i th input item in text t, ci, is</Paragraph>
      <Paragraph position="9"> where &amp;quot;presubjective-nonactive&amp;quot; is the text situation of the first item in a text, and SENTENCE(X) is true iff x is a sentence. To refer to the components of a context, we shall use the functions LAST-SC-OF, LAST-ACTIVE-CHARACTER-OF, PREVIOUS-SCS-OF, and TEXT-SITUATION-OF, which map a context C to the LastSC, LastActiveCharacter, PreviousSCs, and TextSituation of C, respectively.</Paragraph>
      <Paragraph position="10"> In the above definitions, LastSC, LastActiveCharacter, SC, and ActiveCharacter are sets, and PreviousSCs is a set of sets. This is because a subjective sentence can represent the shared psychological POV of more than one physical character (Banfield 1982). For example: (13) 13&amp;quot;1Leaning out of the window side by side the two women watched the man'...132Now he threw away his cigarette. 133They watched him. 134What would he do next? \[Woolf, The Years, p. 103; cited by Banfield 1982, p. 96\] Sentence (13.4) is a represented thought whose SC is more than one physical character.  6.3 States of Affairs, Sentences, and Contexts: Some Further Details 6.3.1 States of Affairs. The following are the range of state-of-affairs types that can be included in a FeatureSet. For more refined categorizations for the purpose of analyzing tense and aspect, see, for example, Reichenbach (1947) and the papers in the special issue of Computational Linguistics on tense and aspect in 1988.</Paragraph>
      <Paragraph position="11"> .</Paragraph>
      <Paragraph position="12"> 2.</Paragraph>
      <Paragraph position="13"> 3.</Paragraph>
      <Paragraph position="14"> 4.</Paragraph>
      <Paragraph position="15"> private-state actions, such as looking and sighing (see Section 11); other kinds of actions; private states; and  nonprivate states, such as being six feet tall.</Paragraph>
      <Paragraph position="16"> The rough part of this categorization is the boundary between the second and fourth items. We shall assume that function FEATURES classifies only clear instances of nonprivate states as such, and that states of affairs that fall between the categories of action and nonprivate state, such as processes, are classified as actions. In quoted speech, there are states of affairs on two levels: those spoken of and the action of speaking itself. The regularities on which the algorithm is based involve the latter. Thus, the contents of the quoted string are not considered, and, even if a discourse parenthetical does not actually appear, the sentence is viewed as one whose main clause is about a communicative action (a subtype of (2)above); that is, as one whose main verb phrase contains a communicative verb with a quoted string as object. We return to this point in Section 9.5.</Paragraph>
      <Paragraph position="17">  Computational Linguistics Volume 20, Number 2  &amp;quot;sentence,&amp;quot; although such an Inputltem is sometimes smaller than an actual sentence, sometimes larger. The former occurs for compound sentences and the latter occurs for sentences of quoted speech.</Paragraph>
      <Paragraph position="18"> Suppose that a compound sentence with conjuncts c1,..., Cn starts at position i in text t. Then, ITEM(t,/) = cl, ..., ITEM(t,i+n-1) = Cn. We shall call each ci a &amp;quot;sentence.&amp;quot; In the case of quoted speech, everything enclosed within a single pair of quotes, together with the accompanying discourse parenthetical, if there is one, counts as a single Inputltem that we shall call a &amp;quot;sentence.&amp;quot; For example, (4.4) and (4.5) together compose a single Inputltem (a single &amp;quot;sentence&amp;quot;). Since the algorithm does not consider the contents of quoted strings (see Section 6.3.1), there is no reason for each constituent sentence of a quoted string to be a separate Inputltem.</Paragraph>
      <Paragraph position="19"> From this point forward, numbering within cited passages is as follows: each Inputltem that is not a paragraph or scene break is given a separate number. Now that the units of input have been specified, the following can be noted. In this work, if any part of a sentential Inputltem s is subjective, then s as a whole is considered to be subjective. This makes the algorithm easier to understand. Enabling it to report which part of s is subjective, if relevant, would involve straightforward refinements of the text situations, of interpretations, and of some conditional steps of the algorithm into subcases.</Paragraph>
      <Paragraph position="20"> 6.3.3 Choosing a State of Affairs. Out of the states of affairs included in a FeatureSet for a sentential input item s, the algorithm chooses just one to consider. Specifically, the algorithm for POV uses the following function:</Paragraph>
      <Paragraph position="22"> As we shall see in Section 8, private states are particularly important to consider, because if a sentence that is about a private state is interpreted to be a private-state report, then the SC of the sentence is the experiencer of the private state. A subordinated clause can report a character's private state; an example is &amp;quot;thinking ...&amp;quot; in the following sentence (as it appears in the novel): When he \[Call\] got within fifteen miles of Lonesome Dove he cut west, thinking they would be holding the herd in that direction. \[McMurtry, Lonesome Dove, p. 181\] Following is the specification of CHOSEN-STATE-OF-AFFAIRS. Let Cmain be the main clause of s; let Cl,C2,... ,Cn be the other n clauses of s; and let hn be the head noun of the subject of Cmain. Further, let SOamain be the state of affairs that Cmain is about; let SOahn be the state of affairs that hn is about (but if hn is not about a state of affairs--for example, if it is a proper noun--then let sOahn be nil); and let soai (1 K i &lt; n) be the state of affairs that ci is about. Then, the result of CHOSEN-STATE-OF-AFFAIRS(FEATURES(S)) is as follows: If SOamain is a private state then the result is sOamain else if SOahn is a private state then the result is soahn else if 3 soai (1 &lt; i &lt; n) such that soai is a private state, and ci is not subordinated to another clause Ck such that SOak is a private state then  Janyce M. Wiebe Tracking Point of View in Narrative Table 1 Choosing a state of affairs: examples.</Paragraph>
      <Paragraph position="23"> s: &amp;quot;Japheth turned the book over in a puzzled manner.&amp;quot; Chosen: SOamain S: &amp;quot;The pain increased.&amp;quot; Chosen : SOahn S: &amp;quot;The pain angered him.&amp;quot; Chosen: SOamain S: &amp;quot;When he got within fifteen miles of Lonesome Dove he cut west, thinking they would be holding the herd in that direction.&amp;quot; Chosen: soai, where ci = &amp;quot;thinking they would be holding the herd in that direction&amp;quot; s: &amp;quot; 'Rosie's just saying that. She doesn't really care,' Zoe said.&amp;quot; (sentence (15.3)). Chosen: SOamainn, which is the action of Zoe saying the quoted string. s: &amp;quot;What are you doing in here?&amp;quot; (sentence (4.1)).</Paragraph>
      <Paragraph position="24"> Chosen: SOamain, which, as s appears in the novel, is Zoe saying the quoted string. That Zoe is the speaker is in FEATURES(s) (see items (2) and (4) of section 6.1 and the end of section 6.3.1).</Paragraph>
      <Paragraph position="25"> the result is $oai (if there is more than one such soai, one is randomly chosen) else the result is SOamain.</Paragraph>
      <Paragraph position="26"> Table 1 gives some examples.</Paragraph>
      <Paragraph position="27"> Note that private-state terms appearing in certain types of constituents cannot be used to report private states. An example is a manner adverbial (Quirk et al. 1985), such as the italicized portion of &amp;quot;Japheth turned the book over in a puzzled manner.&amp;quot; (There are others addressed in Wiebe \[1990\], but because of space limitations, we shall ignore them in this paper.) The phrase &amp;quot;in a puzzled manner&amp;quot; does not report a private state, but rather describes the manner in which something is done. The state of affairs chosen for this sentence is simply the main-clause action.</Paragraph>
      <Paragraph position="28"> To facilitate discussion, a sentential Inputltem of which the state of affairs chosen for consideration is of type X will be called an &amp;quot;X sentence,&amp;quot; for example, &amp;quot;action sentence&amp;quot; and &amp;quot;private-state sentence.&amp;quot; 6.3.4 The Text Situations. Following are the text situations (recall that a text situation is part of the context, defined in Section 6.2). 2 To make it easier to understand them, definitions are given in both English and diagrams. Each diagram shows what appears between the start of the current scene, represented by &amp;quot;start-of-scene,&amp;quot; and the current position, represented by a diamond (c~). The start of the current scene is either a scene break or the very beginning of the text. The symbol &amp;quot;P&amp;quot; represents a paragraph break; &amp;quot;objective-sentence&amp;quot; and &amp;quot;subjective-sentence&amp;quot; represent objective and subjective sentences, respectively; &amp;quot;sentence&amp;quot; alone represents either a subjective or objective sentence; the symbol &amp;quot;*&amp;quot; means 0 or more occurrences, &amp;quot;/&amp;quot; means 1 or more occurrences, and &amp;quot;...&amp;quot; represents any number of paragraph breaks and sentences (but not scene breaks, since only what has appeared since the start of the current scene is shown). A scene is assumed to always begin with a paragraph break.</Paragraph>
      <Paragraph position="29"> 2 1 am indebted to Stuart C. Shapiro for suggesting the names.</Paragraph>
      <Paragraph position="30">  Computational Linguistics Volume 20, Number 2</Paragraph>
      <Paragraph position="32"> presubjective-nonactive: a subjective sentence has not appeared so far in the current scene, and a sentence with an active character (defined in Section 8.2.2) has not appeared so far in the current paragraph.</Paragraph>
      <Paragraph position="33"> start-of-scene P (objective-sentence + P)&amp;quot; objective-sentence ~ o None has an active character presubjective-active: a subjective sentence has not appeared so far in the current scene, but a sentence with an active character has appeared earlier in the current paragraph.</Paragraph>
      <Paragraph position="34"> start-of-scene P (objective-sentence + P)~ objective-sentence +, * At least one has an active character continuing-subjective: the current sentence follows a subjective sentence without a paragraph break.</Paragraph>
      <Paragraph position="35"> start-of-scene ... P sentence ~ subjective-sentence o broken-subjective: the current sentence follows a subjective sentence, but after a paragraph break.</Paragraph>
      <Paragraph position="36"> start-of-scene ... subjective-sentence P o interrupted-subjective: the current sentence follows an objective sentence, but an earlier sentence in the current paragraph is subjective. start-of-scene . .. P sentence&amp;quot; subjective-sentence sentence ~  objective-sentence o postsubjective-nonactive: a subjective sentence has appeared in the current scene, and an objective sentence and a paragraph break have appeared since the last subjective sentence. However, a sentence with an active character does not appear earlier in the current paragraph. start-of-scene ... subjective-sentence ... objective-sentence P o  -orstart-of-scene ... subjective-sentence ... P objective-sentence +, * Y None has an active character postsubjective-active: like the postsubjective-nonactive situation, except that a sentence with an active character does appear earlier in the current paragraph.</Paragraph>
      <Paragraph position="37"> start-of-scene ... subjective-sentence ... P objective-sentence +, *  At least one has an active character 7. Focus of This Work Algorithms for the preprocessing functions are not given in this paper, and algorithms for NEW-CONTEXT and NEW-CONTEXT' are given in Appendix A. The remainder of  Janyce M. Wiebe Tracking Point of View in Narrative this paper develops an algorithm for function POV, which at the highest level is the following:</Paragraph>
      <Paragraph position="39"> Active characters are discussed in the section about identifying the SC (Section 8), because the raison d'etre of the active-character component of an interpretation is that the active character of an objective sentence may become the LastActiveCharacter of a later context, contexti, and as such, may become the SC of a subjective sentence that is processed in contexti.</Paragraph>
      <Paragraph position="40"> 8. Identifying the Subjective Character</Paragraph>
    </Section>
    <Section position="3" start_page="246" end_page="249" type="sub_section">
      <SectionTitle>
8.1 Introduction
</SectionTitle>
      <Paragraph position="0"> The SC of a subjective sentence can always be identified from the sentence itself if the sentence has a narrative parenthetical, such as 'Dennys thought' in (2), and can sometimes be so identified if the sentence is about a private state. When the SC is not identifiable from the sentence, she is often a previously mentioned character. Thus, as the text is processed, any algorithm for tracking POV must keep track of characters who are likely to become the SC of a later subjective sentence. I call such characters expected subjective characters. The algorithm presented in this paper considers two possibilities: the last subjective character and the last active character of the current context. Each is an expected subjective character only in certain text situations. The idea of keeping track of entities evoked in the text in order to interpret later sentences is attributable to work on anaphora resolution (e.g., Sidner \[1983\] and Grosz, Joshi, and Weinstein \[1983\]). Active characters are based specifically on Sidner's actor focus, but while Sidner's actor focus is &amp;quot;whoever is currently the agent in the sentence&amp;quot; (p. 282), many sentences with agents do not have active characters, as we shall see in Section 8.2.2.</Paragraph>
      <Paragraph position="1"> In some cases, the SC of a subjective sentence is not identifiable when the sentence appears, but may be identifiable after later sentences are processed. The algorithm, as presented in Wiebe (1990), handles one such case. Because of space limitations, however, this aspect of the algorithm is not presented in this paper, but is only briefly described (in Section 8.2.4).</Paragraph>
      <Paragraph position="2"> Here is a high-level algorithm for identifying the SC (comments are preceded by</Paragraph>
      <Paragraph position="4"> The following subsections refine and illustrate the above algorithm. We first consider identifying the SC based on expected subjective characters (in Section 8.2), and then identifying the SC from the sentence (in Section 8.3).</Paragraph>
      <Paragraph position="5"> 8.2 Identifying the SC from the Context In this section, we consider cases in which the SC of a subjective sentence cannot be identified from the sentence itself. We first consider the last subjective character (Section 8.2.1), then the last active character (Section 8.2.2), and then the cases in  which both (Section 8.2.3) or neither (Section 8.2.4) are expected subjective characters. 8.2.1 The Last Subjective Character. An SC who is not identifiable from the sentence itself is most often the last subjective character. In this case, the current sentence is continuing a character's POV, if the previous sentence was also subjective, or resuming one, if objective sentences have appeared since the last subjective sentence. Sentence (1.2) above illustrates the former, and passage (15) illustrates the latter: (15)  ls'nThere was no use arguing. \[Oneal, War Work, p. 40; italics in original\] Sentences (15.1), (15.2), and (15.11) are Zoe's subjective sentences. Sentence (15.11) expresses Zoe's judgment that there is no use in arguing, resuming Zoe's point of view: it has the same SC as the last subjective sentence, (15.2), and is separated from (15.2) by objective sentences (15.3)-(15.10).</Paragraph>
      <Paragraph position="6"> If there has not been a subjective sentence so far in the text, then the last subjective character, which is the empty set in this case, clearly should not be an expected subjective character. Moreover, drastic spatial and temporal discontinuities can block the continuation or resumption of a character's psychological POV. This paper considers one such kind of discontinuity, a scene break. The condition under which the last subjective character of contexti is an expected subjective character is when a subjective sentence has appeared in the current scene. That is, LAST-SC-IS-AN-EXPECTED-SC (contexti) is true iff TEXT-SITUATION-OF(contexti) ~ {presubjective-nonactive, presubjective-active}. A scene break is a break from one parallel story-line to another. Almeida (1987) analyzes parallel story-lines as forming separate narrative-lines, which are stretches of narrative that are controlled by single now-points. The following passage illustrates  Janyce M. Wiebe Tracking Point of View in Narrative the situation in which a scene break blocks the resumption of a character's point of view:  fort out of imagining Lanette's advice. \[Gibson, Mona Lisa Overdrive, pp. 275-276; ellipsis in original\] As passage (16) appears in the novel, the SC of all of the subjective sentences before the chapter break is the character Slick, and a scene break occurs at the chapter break. While the sentence following the break is subjective, the SC of that sentence should not be identified to be the last subjective character, Slick.</Paragraph>
      <Paragraph position="7">  may also be the actor of an action that a previous objective sentence is about (but less commonly than the last subjective character). Since this character need not be the SC of the last subjective sentence, this is a way for the author to initiate a new point of view. Following is an example (by this point in the novel, both Jake and Augustus have been the SC of previous subjective sentences):  (17) 17.1 Jake felt sour. 17&amp;quot;2He wished again that circumstances hadn't prompted him to come back. 173He had already spent one full night on horseback, 174and 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.</Paragraph>
      <Paragraph position="8"> 17&amp;quot;5&amp;quot;I don't know as I'm coming,&amp;quot; he said. 17'6&amp;quot;I just got here. If I'd known you boys did nothing but chase horses around all night, I don't know that I would have come.&amp;quot; 177&amp;quot;Why, Jake, you lazy bean,&amp;quot; Augustus said, 17&amp;quot;Sand walked off. 179jake had a stubborn streak in him, 171degand once it was activated even Call could seldom do much with him. \[McMurtry, Lonesome Dove, p. 162\]  As this passage appears in the novel, (17.1)-(17.4) are the subjective sentences of Jake, and (17.9)-(17.10) are the subjective sentences of Augustus, the actor of an action that a previous objective sentence was about (sentence (17.8)). However, it is Jake who is the last subjective character, so Augustus's point of view is being initiated, not merely resumed or continued.</Paragraph>
      <Paragraph position="9"> The situation I observed in which POV shifts to an actor (who is not also the last subjective character) is one in which the actor was the SC of some previous subjective  Computational Linguistics Volume 20, Number 2 sentence in the text, and the sentence about his or her action is focused by the text. The precise situation is captured by the following specifications of what an active character is, what the last active character is, and of the text situations in which the last active character is expected.</Paragraph>
      <Paragraph position="10"> Suppose that os is an objective sentence that is Inputltemj. Saying that os has an active character means that POV(FEATURES(0S), contextj) = /objective, ac I, where ac is not the empty set. This is the case iff: os is about an action that is actually performed in the current scene by ac (more precisely, when the state of affairs chosen for consideration is such an action), and ii. acc PREVIOUS-SCS-OF (contextj). That is, ac is in the set of characters who have been the SC of some previous subjective sentence in the text (possibly before the current scene).</Paragraph>
      <Paragraph position="11"> The algorithm for POV determines whether an action meets the conditions in (i) by looking at such things as the tense, aspect, and mood of os (the features the algorithm considers are in FEATURES(0s)). First, to guarantee that the action is not performed earlier or later than the current moment in the story, the main verb phrase of os must be in the simple past. Also, to be about a specific action, os cannot be habitual. So, the main verb phrase cannot be accompanied by an adverbial such as at times, usually, rarely, or on weekends. 3 Finally, to be about an action that actually occurs, the main clause of the sentence must not contain modal auxiliary verbs such as could, going to, had better, have to, might, should, or must, modal adverbs such as likely, maybe, perhaps, or possibly, and it must not be negated. If the action is quoted speech, then these restrictions apply to the discourse parenthetical.</Paragraph>
      <Paragraph position="12"> Now we turn to the last active character. In contexti, LAST-ACT1VE-CHARACTER-OF(contexti) is the empty set if no sentences with active characters have appeared; otherwise, it is the active character of the last sentence that had one.</Paragraph>
      <Paragraph position="13"> The last active character is an expected subjective character only when a subjective sentence has not appeared earlier in the current paragraph, and there is an earlier sentence in the current paragraph that has an active character. That is, LAST-ACTIVE-CHARACTER-IS-AN-EXPECTED-SC (contexti) is true iff TEXT-SITUATION-OF(contexti) E {presubjective-active, postsubjective-active }.</Paragraph>
      <Paragraph position="14"> 8.2.3 When There Are Two Expected Subjective Characters. When the last active character and the last subjective character are both expected subjective characters (which, as the reader may have noticed, is when the current text situation is postsubjective-active), the algorithm chooses the last active character in most cases, since he or she is more highly focused by the text. In fact, there is only one case in which the algorithm chooses the last subjective character: when the sentence is about the last active character (specifically, when the last active character is the experiencer or actor of the state of affairs chosen for consideration). Following is the algorithm for function CHOOSE-AN-EXPECTED-SC introduced in Section 8.1 above. The only new function is EXPERIENCER-OR-ACTOR-OF, which maps a state of affairs soa and a feature set f into</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="249" end_page="266" type="metho">
    <SectionTitle>
3 The fact that a sentence r is habitual is part of FEATURES(r). The system demonstrated in Appendix B
</SectionTitle>
    <Paragraph position="0"> relies on adverbials to decide whether or not a sentence is habitual. However, an adverbial is not necessary for a simple-past narrative sentence to be habitual.</Paragraph>
    <Paragraph position="1">  Janyce M. Wiebe Tracking Point of View in Narrative the actor or experiencer of soa (if soa or its actor or experiencer is not in f, the result is the empty set).</Paragraph>
    <Paragraph position="2"> CHOOSE-AN-EXPECTED-SC(featureSet, context) if EXPERIENCER-OR-ACTOR-OF (CHOSEN-STATE-OF-AFFAIRS ~eatureSet),featureSet) = LAST-ACTIVE-CHARACTER-OF (context) then return LAST-SC-OF (context) else return LAST-ACTIVE-CHARACTER-OF (context)</Paragraph>
    <Paragraph position="4"> The criterion for choosing the last subjective character is correct for the situation in which the last subjective character's attention is directed toward the last active character, and the sentence represents the last subjective character's reflection about or observation of the other. It is incorrect, however, if the sentence is the last active character's self-reflection or self-perception; this heuristic relies on the relative infrequency of subjective sentences about oneself.</Paragraph>
    <Paragraph position="5"> Consider subjective sentence (15.11). When this sentence is encountered, only the last subjective character (Zoe) is expected, because the sentence is at the beginning of a new paragraph. The algorithm correctly identifies her to be the subjective character.</Paragraph>
    <Paragraph position="6"> Now consider (17.9), which is also subjective. When it is encountered, not only is the last subjective character expected (Jake), but so is the last active character (Augustus): Augustus is the active character of (17.7)-(17.8), because he has been the SC of previous subjective sentences and (17.7)-(17.8) are objective sentences about his current actions; and, when (17.9) is encountered, the last active character is expected, since (17.8) (but no subjective sentences) appeared earlier in the current paragraph. The algorithm correctly identifies the SC to be Augustus (the last active character), rather than Jake (the last subjective character), because the criterion for choosing the last subjective character is not satisfied: the sentence is not about the last active character. Competition also arises in the following passage, but this time it is the last subjective character who should be chosen. When the passage is encountered, Lorena is the last subjective character: (18) ls'l&amp;quot;I never tolt on you, Lorie,&amp;quot; he \[Lippy\] said. 18'2He looked like he might cry too. 183you'll just have to cry, she \[Lorena\] thought.</Paragraph>
    <Paragraph position="7"> \[McMurtry, Lonesome Dove, p. 218\] By this point in the novel, Lippy has been a subjective character. Thus, since (18.1) is about his current action, he is the active character of (18.1). After (18.1), Lippy, as the last active character, is expected, because (18.1) is an objective sentence that begins a new paragraph. Sentence (18.2) is subjective because the evidential 'looked like' appears. Competition is (correctly) resolved in favor of the last subjective character (Lorena), because the sentence is about the last active character (Lippy).</Paragraph>
    <Paragraph position="8"> 8.2.4 When There Are No Expected Subjective Characters. If no character is expected, then the algorithm fails to identify the SC at this point in the text. This eventuality is rare, relative to others: it can arise only upon the first subjective sentence of a scene (otherwise, the last subjective character would be expected) and only in the absence of one of the things that are usually used to initiate a character's point of view (such as  Computational Linguistics Volume 20, Number 2 a narrative parenthetical or private-state report, discussed in Section 8.3). An example of this is (16.6), and another is (19.1), which is the beginning of a novel: (19) 19&amp;quot;1Captain Scalawag's treasure! 19&amp;quot;2It was the first thing Pete thought of when he woke up. \[Lorimer, The Mystery of the Missing Treasure, p. 1\] In a case such as (19.1), it is not possible to identify the SC without reading further in the text. In a case such as (16.6), however, it might be possible to do so. That is, a reader might be able to infer who the SC is from clues in the sentence, such as indications of place, or facts that only a certain character knows. This process is not addressed in this work. However, the author could have made identifying the SC easier by using, for example, a narrative parenthetical or private-state report (see Section 8.3); by not including one of these, the author is deliberately demanding some extra work from the reader.</Paragraph>
    <Paragraph position="9"> The algorithm as presented in Wiebe (1990) can identify the SC after later sentences are processed in the case where a later sentence contains a narrative parenthetical or is a private-state report. As illustrations, (16.7) is a private-state report whose SC, Mona, is the SC of (16.6), and (19.2) is a private-state report whose SC, Pete, is the SC of (19.1).</Paragraph>
    <Paragraph position="10"> 8.3 Identifying the SC from the Sentence We now turn to cases in which the SC is identifiable from the sentence itself (that is, cases in which IDENTIFY-SC-FROM-THE-SENTENCE(featureSet, context) ~ {}; see Section 8.1). In these cases, the SC of sentence s is chosen from among certain characters in FEATURES(s). Such a character need not be the last subjective character; when she is not, s initiates her POV. Thus, the cases discussed in this section--i.e., uses of sentences with certain features in particular contexts--are ways to initiate a character's POV. The straightforward case is when s contains a narrative parenthetical, such as sentence (2). The SC is always the subject of the parenthetical.</Paragraph>
    <Paragraph position="11"> The less straightforward case is when s is a private-state sentence. Dole~el (1973), Cohn (1978), and Banfield (1982) all note that a private-state sentence is a way to initiate a character's POV. In the framework presented in this paper, the SC may be the experiencer of the private state, even if she is not the last subjective character. An example occurs in (20): (20) 2dega&amp;quot;Drown me?&amp;quot; Augustus said. 2deg'2&amp;quot;Why if anybody had tried it, those girls would have clawed them to shreds.&amp;quot; 2deg'3He knew Call was mad, 2deg'4but wasn't much inclined to humor him. 2deg'5It was his dinner table as much as Call's, 2deg'6and if Call didn't like the conversation he could go to bed. 2deg'7Call knew there was no point in arguing. 2deg'8That was what Augustus wanted: argument. 2deg9He didn't really care what the question was, 2deg'1degand it made no great difference to him which side he was on. 2deg'nHe just plain loved to argue. \[McMurtry, Lonesome Dove, p. 16\] Sentences (20.3)-(20.6) are Augustus's subjective sentences and (20.7)-(20.11) are Call's. Thus, (20.7) initiates a new POV. It is a private-state sentence and the SC, Call, is the experiencer of the private state. But passage (20) also shows that the SC of a private- null Janyce M. Wiebe Tracking Point of View in Narrative state sentence need not be the experiencer. In (20.6), for example, &amp;quot;Call didn't like the conversation&amp;quot; is about a private state (Call not liking the conversation), but the SC of the sentence is Augustus, not Call. In the following subsections, we will consider factors that can indicate that it is not the experiencer who is the SC of a private-state sentence.</Paragraph>
    <Paragraph position="12"> 8.3.1 Textual Continuity. POV does not typically shift from one character to another without a paragraph break. Thus, the absence of a paragraph break suggests that a shift has not occurred. Consider the following schema, in which a subjective sentence S, whose SC is X, is followed, without a paragraph break, by a private-state sentence P whose experiencer is a different character Y:</Paragraph>
    <Paragraph position="14"> of P, then a shift would have occurred, from X's POV to Y's POV, without a paragraph break. The fact that no paragraph break appears--that is, that P is in the continuing-subjective situation--suggests that P continues X's POV rather than initiating Y's.</Paragraph>
    <Paragraph position="15"> When a private-state sentence appears in the continuing-subjective situation, therefore, the algorithm identifies the SC to be the last SC rather than the experiencer of the private state.</Paragraph>
    <Paragraph position="16"> The question of whether there is a psychological link between paragraph breaks and tracking POV has not been previously investigated. Stark (1987, 1988) performed psychological experiments that showed that there is a significant correlation between paragraph breaks and discourse discontinuities, but the sorts of discontinuities she investigated did not include changes in POV. Nakhimovsky and Rapaport (1988) suggest that in narrative, paragraph breaks accompany changes in POV, but they did not investigate this hypothesis experimentally. In fact, we have performed psychological experiments (Bruder and Wiebe 1990 and in press) that did establish such a link. Specifically, through manipulation of paragraph breaks in naturally occurring passages, the experiments showed that readers' interpretations of private-state sentences are influenced by paragraph breaks as we predicted on the basis of the algorithm.</Paragraph>
    <Paragraph position="17">  whether they are private-state reports or represented thoughts. 4 Consider sentence (21): (21) John knew Mary had the key.</Paragraph>
    <Paragraph position="18"> Sentence (21) is about a private state: &amp;quot;John knew Mary had the key.&amp;quot; ps(pl,experiencerl, attitude1, objectl) 4 A private-state sentence may also be objective. An example is a simple-past sentence with a negated factive term and a propositional object, such as &amp;quot;John did not know that Mary was in the next room.&amp;quot; This sentence cannot be John's subjective sentence; it is either objective or the subjective sentence of someone else. Because of space limitations, how the algorithm recognizes and processes such sentences is not discussed in this paper; see Wiebe and Rapaport (1988) and Wiebe (1990).  Computational Linguistics Volume 20, Number 2 Under a private-state report interpretation of (21), pl is not itself the object of some other private state. But under a represented thought interpretation of (21), pl is the object of some other private state P2, the experiencer and attitude of which are implicit: ps (P2, experiencer2, attitude2, &amp;quot;John knew Mary had the key.&amp;quot;</Paragraph>
    <Paragraph position="20"> To my knowledge, this ambiguity in the interpretation of private-state sentences and its importance in tracking POV have not been previously discussed in linguistics or literary theory. For example, Cohn (1978) says that represented thoughts can be distinguished from private-state reports by &amp;quot;the absence of mental verbs&amp;quot; in the former (p. 104).</Paragraph>
    <Paragraph position="21"> The SC of a private-state report is always the experiencer of the private state. So, if some oracle were to inform you that (21) is a private-state report, you would then know that the SC is the experiencer of the private state (John). On the other hand, if the oracle were to inform you that (21) is a represented thought, you could not then identify the SC just by looking at the sentence alone. In fact, it is true of any represented thought without a narrative parenthetical, private-state sentence or otherwise, that you cannot identify the SC from the sentence itself. This is so regardless of whether or not the SC happens to be referred to in the sentence. Consider the following two sentences, which are represented thoughts from different pages of a short story (&amp;quot;The Garden Party&amp;quot; by Katharine Mansfield):  (22) Why couldn't she? (23) What nice eyes he had, small, but such a dark blue!  As these sentences appear in the story, the SC of (22) happens to be the referent of &amp;quot;she&amp;quot; (corresponding to the conversational utterance, &amp;quot;Why can't I?'), but the SC of (23) is not mentioned in the sentence at all. Even though you know that (22) and (23) are represented thoughts, you need to consider the context to identify their SCs. Thus, if a private-state sentence s contains some indication that s is a represented thought, then the SC cannot be identified from s itself, and, as discussed in Section 8.2, the expected subjective characters should be considered.</Paragraph>
    <Paragraph position="22"> Subjective elements indicate that a sentence is a represented thought (this statement is qualified later in this section and in Section 10). Subjective elements are linguistic elements that express emotions, uncertainty, evaluations, and other kinds of subjectivity (Banfield 1982) (they are discussed in detail in Section 9). Examples are evaluative terms such as 'the old bag' (Banfield 1982) and evidentials such as 'evidently' and 'apparently' (Dole~el 1973).</Paragraph>
    <Paragraph position="23"> In the following passage, a subjective element indicates that the SC of a private-state sentence is not the experiencer of the private state. At the start of the passage, Sandy and Dennys are (collectively) the last subjective character: (24) 24&amp;quot;1Japheth, evidently realizing that they were no longer behind him, turned around 242and jogged back toward them, seemingly cool and unwinded. \[UEngle, Many Waters, p. 24\] The subjective element 'evidently' in (24.1) indicates that the sentence is not a private-state report. That is, (24.1) is not a report that Japheth realizes that they are no longer  Janyce M. Wiebe Tracking Point of View in Narrative behind him. Rather, Sandy and Dennys (the collective SC) ascribe this private state to him.</Paragraph>
    <Paragraph position="24"> However, subordinated subjective elements, those within the scope of the private-state term, can appear in private-state reports. (This is one reason why I define private-state reports to be subjective.) Thus, they cannot be used to distinguish private-state reports from represented thoughts, and so cannot be used as evidence that the SC of a private-state sentence is not the experiencer. For example: (25) 25'1Ugh! she \[the girl\] thought. 25&amp;quot;2How could the poor thing have married him in the first place? 25&amp;quot;3johnnie Martin could not believe that he was seeing that old bag's black eyes sparkling with disgust and unsheathed contempt at him. \[Caldwell, No One Hears But Him, pp. 98-99\] Sentence (25.3) is a private-state report and the experiencer is the SC (Johnnie Martin); this is so even though (25.3) contains the subjective 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 SC. On the other hand, the subjective element 'evidently' in (24.1) is not in the scope of 'realizing' (i.e., it is nonsubordinated); thus, it can be used as evidence that the SC is not the experiencer of the private state.</Paragraph>
    <Paragraph position="25"> If a private-state sentence does not have a nonsubordinated subjective element and does not appear in the continuing-subjective situation, then the algorithm identifies the SC to be the experiencer.</Paragraph>
    <Paragraph position="26"> 8.3.3 Broadening and Narrowing of POV. Recall that an experiencer, actor, SC, or expected subjective character may be more than one physical character (recall that these entities are represented as sets). A broadening of point of view occurs when a new subjective character is a superset of the old subjective character, and a narrowing occurs when a new one is a subset of the old one. One situation in which such changes occur is when the experiencer of a private-state report is such a subset or superset. One addition to the algorithm as described so far is needed to allow for this situation: for a private-state sentence in the continuing-subjective situation (without subjective elements that can be considered), if the experiencer is a superset or subset of the last subjective character, then it is the experiencer who the algorithm chooses to be the SC, rather than the last subjective character. For example: (26) 26&amp;quot;1In the clear late afternoon light they \[Call and Augustus\] could see all the way back to Lonesome Dove and the river and Mexico.</Paragraph>
    <Paragraph position="27"> 26&amp;quot;2Augustus regretted not tying a jug to his saddle---~63he would have liked to sit on the little hill and drink for an hour. \[McMurtry, Lonesome Dove, p. 241\] As this passage appears in the novel, (26.1) is the subjective sentence of both Call and Augustus, but (26.2) is the subjective sentence only of Augustus; thus, point of view narrows upon (26.2). The above rule precludes the reading of a sentence such as (26.2) as the represented thought of the last subjective character (i.e., Call and Augustus in passage (26)). However, no examples of such a reading were found in the texts examined; note that the rule applies only if there are no subjective elements that can be considered in the sentence.</Paragraph>
    <Paragraph position="29"> is about (see Section 6.3.3). Both the experiencer of ps and the SC of (27.2) happen to be the same character (Sandy). However, the SC of (27.2) cannot be identified from the sentence itself, because the experiencer of ps is not mentioned in the sentence.</Paragraph>
    <Paragraph position="30"> (I call such experiencers unspecified experiencers; note that EXPERIENCER-OR-ACTOR-OF(ps,FEATURES(27.2)) = {}.) Thus, for a private-state sentence with an unspecified experiencer, IDENTIFY-SC-FROM-THE-SENTENCE returns the empty set, and the algorithm goes on to consider expected subjective characters. In (27), for example, the algorithm correctly identifies the SC of (27.2) to be Sandy, the last subjective character.</Paragraph>
    <Paragraph position="31"> 8.3.5 Summary. Following is the algorithm for function IDENTIFY-SC-FROM-THE-SENTENCE (introduced in Section 8.1). One of the arguments is a feature set, featureSet; we will use s to refer to the sentence thatfeatureSet is a feature set of (i.e., FEATURES(S) = featureSet). With the exception of calls to previously mentioned functions, statements are given in English. Further, two cases discussed later (in Sections 10 and 11) are not included (the complete version is given in Appendix A).</Paragraph>
    <Paragraph position="33"> if s contains a narrative parenthetical then return the subject of the narrative parenthetical</Paragraph>
    <Paragraph position="35"/>
    <Section position="1" start_page="256" end_page="256" type="sub_section">
      <SectionTitle>
9.1 Introduction
</SectionTitle>
      <Paragraph position="0"> We now turn to deciding whether or not a sentence is subjective in the first place. Authors could unambiguously mark each subjective sentence as subjective, by including a narrative parenthetical in each, for example. But suppose that a sentence S that the author intends to be subjective appears in the continuing-subjective situation: (i) P sentence&amp;quot; subjective-sentence sentence-S sentence * P</Paragraph>
      <Paragraph position="2"> A character's POV very often continues at least until the end of the paragraph. So, in schema (i), S and any sentences after S until the paragraph break will very often be subjective sentences of X. Thus, the reader has a strong expectation that X's POV will continue, so a weak hint that S is subjective is sufficient for the reader to recognize  that it is.</Paragraph>
      <Paragraph position="3"> Now consider a text situation in which there has been a subjective sentence in the scene, but objective sentences and paragraph breaks have appeared since then: (ii) start-of-scene ... subjective-sentence (objective-sentence + P)+ sentence-S</Paragraph>
      <Paragraph position="5"> In S's context in schema f/i), X is an expected subjective character. The reader expects X's POV to be resumed, but not as strongly as the reader expects X's POV to be continued in schema (i), since the local context of S in (ii) is not subjective as it is in (i). An unambiguous indication that S is subjective is not necessary, but a stronger hint should be included than is sufficient in (i).</Paragraph>
      <Paragraph position="6"> The main sorts of &amp;quot;hints&amp;quot; of subjectivity that the algorithm considers are linguistic elements that potentially express subjectivity (potential subjective elements). Some of these are weaker hints than others, and many are usually subjective only in certain text situations. The algorithm, which, recall, tracks point of view on the basis of regularities, uses the text situation to decide whether an instance of one is indeed subjective.</Paragraph>
      <Paragraph position="7"> &amp;quot;Subjective element&amp;quot; is the term I use for an instance of a potential subjective element that actually is subjective in the context of use. This term is borrowed from Banfield (1982), but redefined; Banfield uses it to refer only to linguistic elements that are always subjective.</Paragraph>
      <Paragraph position="8"> Section 9.2 identifies a number of potential subjective elements, and Section 9.3 specifies how the algorithm uses them and other information to recognize subjective sentences.</Paragraph>
    </Section>
    <Section position="2" start_page="256" end_page="257" type="sub_section">
      <SectionTitle>
9.2 Potential Subjective Elements
</SectionTitle>
      <Paragraph position="0"> Previous work in linguistics and literary theory noted the presence in subjective sentences of many (but not all) of the potential subjective elements considered by the algorithm. However, with the exception of two of the elements, the perfective and progressive aspects (see below), previous work did not address the problem that many of the elements are only potentially subjective, and can also appear in objective sentences. Further, Wiebe (1990) contains an extensive, detailed catalogue of the potential subjective elements (specified mainly in terms of syntactic and semantic categories presented in Quirk et al. \[1985\]); such a catalog did not previously exist.</Paragraph>
      <Paragraph position="1"> Most of the potential subjective elements are lexical. But it is not words and phrases themselves that are potential subjective elements, but rather words and phrases used</Paragraph>
    </Section>
    <Section position="3" start_page="257" end_page="257" type="sub_section">
      <SectionTitle>
3.1 Subordinators such as 'whoever' and 'whatever', when used in reference to
</SectionTitle>
      <Paragraph position="0"> particular individuals, as in &amp;quot;Whatever it was, it had flown by quickly&amp;quot;</Paragraph>
    </Section>
    <Section position="4" start_page="257" end_page="257" type="sub_section">
      <SectionTitle>
3.2 Adjectival phrases such as 'some kind of', when used in reference to particular
</SectionTitle>
      <Paragraph position="0"> individuals, as in &amp;quot;The object in her hand was some kind of weapon&amp;quot; with particular meanings. For example, 'poor' is a potential subjective element only with its evaluative meaning, as in &amp;quot;Poor John was sick,&amp;quot; but not with its nonevaluative meaning, as in &amp;quot;John was poor&amp;quot; (Banfield 1982; the evaluative meaning of 'poor' is one of the elements that Banfield argues is always subjective).</Paragraph>
      <Paragraph position="1"> Tables 2 and 3 list some potential subjective elements, giving very brief characterizations. For further details, see Wiebe (1990). All of the citations in Tables 2 and 3 are with respect to the linguistic categories of the elements. Some of those who discuss the appearance of the elements in subjective sentences are as follows: Banfield (1982) discusses (1), (2.1)-(2.3), (4), (5), and (12); Dole~el (1973) discusses (1), (2.1)-(2.3), (5), (6.1), (6.2), and (8); Brinton (1980) discusses (9) (she shows that a simile can be a marker of represented perception) and (12); Ehrlich (1990) discusses (11), and Ehrlich (1987) discusses (12).</Paragraph>
      <Paragraph position="2"> The past perfective is potentially subjective simply because a character can reflect on what occurred (or might have occurred) in the past. However, as discussed by Ehrlich (1990) and many others, the narrative past may be expressed by the simple-past tense in the midst of a subjective context; detecting simple-past references to the past is not addressed in this work.</Paragraph>
    </Section>
    <Section position="5" start_page="257" end_page="259" type="sub_section">
      <SectionTitle>
9.3 Recognizing Subjective Sentences
</SectionTitle>
      <Paragraph position="0"> Subjective elements are important for recognizing represented thoughts and perceptions, not private-state reports (recall, in fact, that a nonsubordinated subjective element is evidence that a private-state sentence is not a report). That is, they are important for recognizing subjective sentences whose subjective characters are to be identified from the context, rather than from the sentence itself.</Paragraph>
      <Paragraph position="1"> My examination of novels and short stories suggests the following (we are currently performing psychological experiments investigating the aspects of the algorithm discussed in this section): (1) Two potential subjective elements, the past perfective and  Some potential subjective elements (continued).</Paragraph>
      <Paragraph position="2"> 4 Sentence fragments, such as (30.6) 5 Kinship terms, such as 'Dad' and 'Aunt Margaret' 6 Evidentials, which, in the broadest sense, qualify the information conveyed by a statement (Chafe 1986) 6.1 Evidentials that express certainty or uncertainty, such as 'surely' and 'might' 6.2 Evidentials that express certainty or uncertainty and also that one's knowledge is based partly on evidence. Examples are 'evidently', 'seemingly', 'must have', 'appear to be', 'as if', 'as though', and 'look', as in &amp;quot;He looked like he might cry&amp;quot; 6.3 Hedges, e.g., adverbs such as 'more or less' and 'sort of' when used as modifiers of adjectives and adverbs, as in &amp;quot;It was more or less green&amp;quot;, or as adverbials (Quirk et al. 1985), as in &amp;quot;The man more or less held a large stretch of the border&amp;quot; 6.4 Evidentials that address expectations 6.4.1 Signal that expectations have been met, such as 'of course' (when used as an emphasizer subjunct (Quirk et al. 1985)) as in &amp;quot;John of course sat down&amp;quot; 6.4.2 Signal that expectations have not been met. Examples are adverbs such as 'just', 'merely', and 'only' (when used as attitude diminishers (Quirk et al. 1985)), as in &amp;quot;He just sat and drank&amp;quot; (it was expected that he would do something &amp;quot;more&amp;quot; than sit and drink) 7 Adverbials that are conjuncts, which connect units of discourse (Quirk et al. 1985) (i.e., cue phrases; Reichman 1985, Grosz and Sidner 1986, Cohen 1987). Examples are 'first', 'in addition', 'for instance', 'on the other hand', 'after all', 'anyway', and 'yet' as in &amp;quot;Yet, they were the pride of the family&amp;quot; 8 Conditional clauses 9 Comparative 'like', as in &amp;quot;They followed her like acolytes behind a goddess&amp;quot; 10 Habitual sentences, such as &amp;quot;Gus himself often joked about it&amp;quot; 11 The past perfective, but only in the main verb phrase 12 The progressive, but only in the main verb phrase  the progressive, can typically serve only to continue a character's POV and only within a paragraph (see Ehrlich \[1987\] for an analysis of why this is so for the progressive); (2) stronger ones can continue a character's POV after a paragraph break, or resume a character's POV within a paragraph; (3) still stronger ones, such as evidentials and sentence fragments, can resume the last subjective character's POV or initiate the last active character's just as long as they are expected subjective characters; and (4) the strongest subjective elements, such as exclamations and questions, are always subjective, even when there is not an expected subjective character to whom to attribute the  Computational Linguistics Volume 20, Number 2 sentence. The sets of text situations corresponding to (1)-(4) are:</Paragraph>
      <Paragraph position="4"> Expectations for a subjective sentence are strongest in situation (lts) and weakest in situation (4ts), so the algorithm takes even the weakest potential subjective elements to be subjective in (lts), but only the strongest ones to be subjective in (4ts). In general, each potential subjective element pse is associated with a set of text situations t such that the algorithm interprets pse to be subjective iff the current text situation is in t.</Paragraph>
      <Paragraph position="5"> There is an its (1 &lt; i &lt; 4) such that t contains the situations in lts through its but not those in its + 1 through 4ts. We shall say that pse is associated at the highest level with the situations in its.</Paragraph>
      <Paragraph position="6"> In addition to potential subjective elements, there is another source of information the algorithm considers: the type of state of affairs the sentence is about. First, private-state action sentences can be subjective; see Section 11. Second, private-state sentences are usually subjective (we are not considering objective private-state sentences in this paper; see Wiebe \[1990\] and footnote 4). Third, a nonprivate-state sentence in the continuing-subjective situation usually continues the subjective context. For example: (28) 28.1Lorena didn't like it that Gus acted like Jake wasn't much. 282He had a reputation for being a cool man in a fight. \[McMurtry, Lonesome Dove, p. 190\] Sentence (28.1) is Lorena's subjective sentence, and (28.2), a nonprivate-state sentence, continues her subjective context. In the continuing-subjective situation, therefore, the algorithm interprets a nonprivate-state sentence to be subjective)</Paragraph>
    </Section>
    <Section position="6" start_page="259" end_page="262" type="sub_section">
      <SectionTitle>
9.4 Examples and Discussion
</SectionTitle>
      <Paragraph position="0"> Consider the following passage: (29) 29&amp;quot;1Call had heard from someone that she had been raised rich, in the East, with servants to comb her hair and help her into her shoes when she got up. 29'2It might just have been a story--293it was hard for him to imagine a grownup who would need to be helped into their own shoes--39&amp;quot;4but if even part of it was true she had come a long way down. 29&amp;quot;5Ned Spettle had never got around to putting a floor in the 5 Notice that there are three things that are taken as evidence that a sentence is subjective only in the continuing-subjective situation: being a nonprivate-state sentence, being in the perfective, and being in the progressive. Caenepeel (1989), in work done simultaneously, analyzed aspect and perspective linguistically, and reached similar conclusions with respect to these three types of evidence. Caenepeel suggests that states appearing in what I call the continuing-subjective situation continue the current POV. Her notion of state includes sentences in the perfective and progressive aspects, regardless of the type of state of affairs that the sentence is about. It might be desirable to revise this aspect of the algorithm on the basis of her work, which focused on aspect, to arrive at a more general treatment of these three kinds of evidence in this situation.</Paragraph>
      <Paragraph position="1">  Janyce M. Wiebe Tracking Point of View in Narrative shack of a house he built. 29&amp;quot;6His wife was rearing eight children on the bare dirt. 29'7He had heard it said that Ned had never got over the war, which might have explained it. \[McMurtry, Lonesome Dove, p. 176\] All of these sentences are Call's subjective sentences. Thus, the text situation is continuing-subjective after each of them. Sentences (29.1), (29.3), and (29.7) are Call's private-state reports, and (29.2) and (29.4) contain potential subjective elements that are associated with other situations--and not merely the continuing-subjective one ('might' in (29.2) and 'if even' in (29.4)). The interesting sentences are (29.5) and (29.6), since they contain potential subjective elements that are associated only with the continuing-subjective situation (the past perfective in (29.5) and the progressive in (29.6)). (These sentences express Call's reasons for his belief, expressed in (29.4), that &amp;quot;she had come a long way down.&amp;quot;) In the following passage, a subjective element appears in a situation other than continuing-subjective. The situation is continuing-subjective at the beginning, and Sandy is the last subjective character.</Paragraph>
      <Paragraph position="2">  Many Waters, p. 39\] Sentence (30.1) continues Sandy's subjective context, because it contains the subjective elements 'incredibly' and comparative 'like.' Sentences (30.2)-(30.4) are objective, and a paragraph break appears before (30.5), so the situation at the beginning of (30.5) is postsubjective-nonactive (one of the situations in (3ts)) and the last subjective character, Sandy, is an expected subjective character. The algorithm is able to recognize that (30.5) is Sandy's subjective sentence, because it contains 'so' used as a conjunct, which is subjective as long as there is an expected subjective character.</Paragraph>
      <Paragraph position="3"> Consider passage (19) (reprinted here); recall that it is the beginning of a novel, so there isn't an expected subjective character when it is encountered.</Paragraph>
      <Paragraph position="4"> (19) 19&amp;quot;1Captain Scalawag's treasure! 192It was the first thing Pete thought of when he woke up. \[Lorimer, The Mystery of the Missing Treasure, p. 1\] Since an exclamation is subjective in any situation, the algorithm is able to recognize that (19.1) is subjective.</Paragraph>
      <Paragraph position="5"> In the following passage, in contrast, potential subjective elements that are not necessarily subjective appear when there isn't an expected subjective character. This passage is of the type that Banfield has characterized as having an empty center (Banfield 1987); 6 it describes perceptions and impressions that one would have if observing 6 Banfield (1987) extends the definition of subjective sentences given in Banfield (1982) to include these kinds of sentences; our interest is to recognize character's subjective sentences (for the reasons given in Section 3), so we adopt the earlier definition.</Paragraph>
      <Paragraph position="6">  Computational Linguistics Volume 20, Number 2 the scene, but no character is present to whom to attribute them. There is an expected subjective character at the beginning of the passage, but a scene break appears after the third sentence. (The blank lines after (31.3) and (31.12) appear in the original. The sentences following '--' are a kind of unuttered quoted speech.) (31) 31&amp;quot;l&amp;quot;We're coming,&amp;quot; Oholibamah said. 31&amp;quot;2arid they hurried toward the central section of the oasis, where Noah's vineyards were, and his grazing grounds, and his tents. 31'3And where Dennys was waiting for them.</Paragraph>
      <Paragraph position="7"> 314The moon set, its path whiter than the desert sands dwindling into shadow. 31SThe stars moved in their joyous dance across the sky.</Paragraph>
      <Paragraph position="8"> 31&amp;quot;6The horizon was dark with that deep darkness which comes just before the dawn.</Paragraph>
      <Paragraph position="9"> 31&amp;quot;7A vulture flew down, seemingly out of nowhere, stretching its naked neck, settling its dark features.</Paragraph>
      <Paragraph position="10"> 31'B--Vultures are underestimated. Without us, disease would wipe out all life. We clean up garbage, feces, dead bodies of man and beast.</Paragraph>
      <Paragraph position="11"> We are not appreciated.</Paragraph>
      <Paragraph position="12">  walk into the desert. \[UEngle, Many Waters, pp. 118-119\] There are scene breaks after (31.3) and after (31.12). Between the breaks, there are no private-state sentences or narrative parentheticals, and, since none of the creatures in the scene has been the SC, there are no sentences with active characters (this is true for the elided sentences as well). Thus, if none of the potential subjective elements are subjective, then there are no expected subjective characters from (31.4) to (31.12). There are strong potential subjective elements that would be subjective if there were an expected subjective character--'seemingly' in (31.7), 'sound' in (31.9), 'yet' used as a conjunct and 'seemed' in (31.10), 'oddly' in (31.11), and a sentence fragment in (31.12). However, since these are associated at the highest level with the situations in (3ts), and not with the presubjective-nonactive situation, the algorithm correctly does not interpret them to be subjective.</Paragraph>
      <Paragraph position="13"> There are three things to note about passages such as (31). First, the potential for an overt narrator to appear is strong when there isn't an expected subjective character and a strong element such as 'seemingly' appears; my restriction to texts that do not have overt narrators allows the algorithm to exclude this possibility from consideration. Second, the algorithm does not revise its decision as to whether a sentence is subjective in light of later sentences. However, one could imagine a sentence within passage (31) that might cause the reader to decide that earlier sentences were actually subjective. For example, the following sentence, inserted after (31.10), would suggest this: &amp;quot;Dennys was mystified by the spectacle.&amp;quot; The algorithm would interpret this sentence to be  Janyce M. Wiebe Tracking Point of View in Narrative subjective; this would affect its interpretation of the remainder of the passage--the algorithm would interpret the potential subjective elements in (31.11) and (31.12) to be subjective--but it would not affect its interpretation of earlier sentences. Third, to be conservative, only Banfield's emotive and evaluative subjective elements, which must be understood to express someone's emotions or evaluations, are associated with the presubjective-nonactive situation (4ts). As mentioned in Section 8.2.4, the number of sentences that appear in this situation is relatively small, and, of those that do, many are private-state reports or have narrative parentheticals. It may be that some of the potential subjective elements associated at the highest level with the situations in (3ts) should also be associated with the one in (4ts); the appearance of the relevant kind of subjective sentence in this situation was too rare in the texts examined to decide this. Because there are 34 classes of potential subjective elements, the majority of which have multiple members (Wiebe 1990), a significant number of each potential subjective element in each situation was not found. The association of elements with text situations is based on the examples that were found, and, for the ones that did not appear very often in the texts examined, on my judgments as to which of the ones that did appear often are closest to them in strength. Psychological experiments (with Gail Bruder) of this aspect of the algorithm are underway. We plan to revise the association of potential subjective elements with text situations as needed in light of the results.</Paragraph>
    </Section>
    <Section position="7" start_page="262" end_page="266" type="sub_section">
      <SectionTitle>
9.5 Quoted Speech
</SectionTitle>
      <Paragraph position="0"> In quoted speech, there are two points of view: the point of view taken by the quoted string--the speaker's--and the one taken by the discourse parenthetical, which may be objective or a character's. 7 It is the point of view taken by the discourse parenthetical that concerns us here: the speaker's point of view is not directly presented by a quoted string, as the subjective character's is by a subjective sentence, but is conveyed indirectly through a communicative act. Quoted speech is a major way of communicating the beliefs, intentions, etc., of characters who are not to become the SC; merely the fact that what a character says expresses her point of view should not lead the reader to anticipate later subjective sentences of that character, as the reader does after a subjective sentence.</Paragraph>
      <Paragraph position="1"> Therefore, the algorithm considers subjective elements that appear in discourse parentheticals, but not those within quoted strings. For example: (32) &amp;quot;I'll talk to Amy,&amp;quot; Daddy said, &amp;quot;and make sure she behaves herself.&amp;quot; \[Sachs, Amy and Laura, p. 100\] The subjective element 'Daddy' in the discourse parenthetical is attributed to an expected subjective character, Laura. In contrast, the algorithm interprets the following sentence from passage (20) to be objective, even though there is a question in the  quoted string: &amp;quot;Drown me?&amp;quot; Augustus said.</Paragraph>
      <Paragraph position="2"> 7 See Banfield (1982) for a principled account of the relationship between the points of view of a quoted string and its discourse parenthetical.  Computational Linguistics Volume 20, Number 2 10. A Return to Private-State Sentences In contrast to what was implied in Section 8.3.2, there are some subjective elements that do not suggest that a private-state sentence is a represented thought, even when they are nonsubordinated. First, Cohn (1978) shows that private-state reports 8 do not always report private states experienced specifically at the current moment in the story, but instead have &amp;quot;almost unlimited temporal flexibility&amp;quot; (p. 34). One consequence of this is that private-state reports can be habitual. Second, Cohn (1978) also shows that private-state reports can employ simile. 9 Thus, comparative 'like' can appear nonsubordinated in private-state reports. For example: (33) His \[Sandy's\] head began to swell, to be filled with hot air like a balloon, so that he was afraid he was going to float off into the sky.</Paragraph>
      <Paragraph position="3"> \[UEngle, Many Waters, p. 27\] Finally, some intensifier adverbs, when nonsubordinated in a private-state report, simply indicate the degree to which the private state is experienced. An example is 'hardly,' as in: (34) Sandy, his flannel shift still draped over his head, was hardly aware that he was supporting his brother. \[UEngle, Many Waters, p. 27\] Sentence (34), as it appears in the novel, is Sandy's private-state report. Given these observations, we need to revise the algorithm as presented so far: the algorithm does not consider the above types of subjective elements when it decides who the SC of a private-state sentence is, even when they are nonsubordinated to the private-state term.</Paragraph>
      <Paragraph position="4"> 11. Private-State-Action Sentences A private-state action is an action from which a private state can be inferred. Examples are looking, glancing, sighing, frowning, smiling, and shivering. In contrast to a private-state report such as &amp;quot;She was unhappy,&amp;quot; the sentence &amp;quot;She frowned&amp;quot; narrates a private-state action from which unhappiness or displeasure can be inferred, but does not directly report the character's private state. 1deg In a given context, the private state that can be inferred may or may not be significant for tracking point of view. It is significant in the following passage:  (35) 35&amp;quot;lZoe looked at the notebook. 352On the first page Joe had written WAR WORK in large block letters in red and blue crayon. 35&amp;quot;3On the next page he had written the date 35&amp;quot;4and under it all about seeing Miss Lavatier's boyfriend. \[Oneal, War Work, p. 47\] 8 Cohn's term for a private-state report is psycho-narration.</Paragraph>
      <Paragraph position="5"> 9 Cohn's term for such reports is psycho-analogy.</Paragraph>
      <Paragraph position="6"> 10 Brinton (1980) notes that perceptions may be reported from an &amp;quot;outer perspective&amp;quot; with terms such as 'gaze' and 'look' or from an &amp;quot;inner perspective&amp;quot; with terms such as 'see' and 'hear&amp;quot; (p. 370-371), but she does not address the significance of this difference for tracking point of view.  Janyce M. Wiebe Tracking Point of View in Narrative As this passage appears in the novel, (35.2)-(35.4) continue the subjective context established by (35.1)--they are subjective sentences presenting what Zoe sees. Interpreting (35.1) to be subjective, the algorithm is able to recognize that (35.2)-(35.4) are also subjective, because the past perfective is a subjective element in the continuing-subjective situation, and it is able to determine that the SC is Zoe, because she is the last subjective character. On the other hand, a private-state action sentence might not be the subjective sentence of the actor; an example is (36.1): (36) 36'1japheth looked at them. 36&amp;quot;2&amp;quot;You are flushed. And wet.&amp;quot; 36&amp;quot;3He himself did not seem to feel the intense heat. \[UEngle, Many Waters, p. 20\] As this passage appears in the novel, (36.3) is the subjective sentence of the last subjective character, Sandy and Dennys. If (36.1) were Japheth's subjective sentence, then it would be to Japheth that the subjective element 'seem' in (36.3) would be attributed, rather than to Sandy and Dennys.</Paragraph>
      <Paragraph position="7"> Like quoted speech, a private-state-action sentence is a way to communicate something about the consciousness of a character who is not to become the SC. The reader infers from (36.1) that Japheth sees the people he is looking at; however, there are no subsequent subjective sentences about what he sees, such as the sentences in (35) that show what Zoe sees.</Paragraph>
      <Paragraph position="8"> The first decision to be made when one encounters a private-state-action sentence is whether it should be treated as a private-state sentence or as an action sentence. Whether the sentence is subjective and, if so, who the SC is then depend upon the factors already presented.</Paragraph>
      <Paragraph position="9"> The algorithm's treatment of these sorts of sentences is based on the observation that a more direct appeal to a character's consciousness, such as a private-state report or a narrative parenthetical, is usually used to establish a character as an SC for the first time. Thus, the actor of a private-state-action sentence that is the actor's subjective sentence has usually been the SC before. While consistency in the interpretation of a passage with this sort of sentence must be supported by other factors, this regularity is a strong one in the texts examined.</Paragraph>
      <Paragraph position="10"> Thus, if the actor has been a SC, then the algorithm treats a private-state-action sentence s as it would have treated a private-state sentence in the same context; otherwise, it treats s as it would have treated an action sentence in the same context (see Wiebe \[1990\] for illustrations of the various consequences of treating s one way or the other).</Paragraph>
      <Paragraph position="11"> 12. Tests of the Algorithm This section summarizes the tests of the algorithm that have been performed. First, the algorithm was hand-simulated on over 700 pages (roughly 17,500 sentential input items) from 7 novels that represent a range in the number of different subjective characters they contain. Given the large amount of data and the preprocessing requirements of running the algorithm, the purpose of this test was not to compile statistical measures, but rather to find out what kinds of exceptions occur. Generally, point of view is manipulated in these novels as expected by the algorithm. Of the exceptions, many can be attributed to problems that are specifically not addressed in this work, such as how the spatial and temporal points of view affect the psychological one (discussed in the next section), how point of view is manipulated in  Computational Linguistics Volume 20, Number 2 relatively rare situations, such as the very beginning of a novel, and what constitutes a &amp;quot;significant&amp;quot; subjective context for the purpose of interpreting private-state-action sentences (discussed in Appendix C). (The classes of errors and examples of them are given in Wiebe \[1990\].) In order to obtain more specific numeric results, the algorithm was tested, and results tabulated, for a more modest amount of text (32 pages with 900 sentential input items). The results are positive, and are given in Appendix C. Fully testing the algorithm would require a much larger corpus, in which a significant number of each of the possibilities arises. Such testing would be a major endeavor in itself. Finally, as mentioned in Section 8.3.1, we have performed psychological experiments that showed that readers' interpretations of private-state sentences are affected by paragraph breaks as we predicted on the basis of the algorithm (Bruder and Wiebe 1990 and in press). We are continuing this line of research with psychological experiments of other aspects of the algorithm.</Paragraph>
      <Paragraph position="12"> 13. Relation to Anaphora Resolution A question likely to arise in the reader's mind is how tracking POV and anaphora resolution are related. Anaphora resolution is necessary for tracking POV (the actors and experiencers of states of affairs must be known; see item (4) of Section 6.1). But it is certainly not sufficient, and an additional mechanism is needed. A case that clearly illustrates this is a private-state sentence in which the experiencer is referred to with a pronoun, such as: (37) He hated holidays.</Paragraph>
      <Paragraph position="13"> Resolving the pronoun is not sufficient for identifying the SC of (37), since the SC may or may not be the referent of 'he' (see Section 8.3.2).</Paragraph>
      <Paragraph position="14"> However, the pipeline architecture of the algorithm is not realistic. Almost certainly, POV affects anaphora resolution, and also recognizing the discourse structure of the text. Specific issues I am exploring are discussed in the next section. 14. Directions for Future Research and Conclusion One direction for future research is investigating the interactions of different points of view. A large class of exceptions to the algorithm can be attributed to interactions between the spatial and temporal points of view and the psychological one. For example, there are spatial and temporal discontinuities other than scene breaks after which a character should no longer be expected: the current &amp;quot;here&amp;quot; and &amp;quot;now&amp;quot; in the story may shift away from the character, or the character may leave the scene while the &amp;quot;here&amp;quot; and &amp;quot;now&amp;quot; remain unchanged. These situations must be distinguished from the situation, occurring with represented thoughts, in which there is a &amp;quot;projected here&amp;quot; and a &amp;quot;projected now&amp;quot; of events that are being thought about, and from the situation in represented perception in which there is a &amp;quot;projected here&amp;quot; of events being perceived (Bruder et al. 1986; Rapaport et al. 1989a, 1989b).</Paragraph>
      <Paragraph position="15"> An important area of future research is investigating interactions among tracking POV, recognizing discourse structure, and anaphora resolution. I am currently focusing on discourse structure within subjective contexts. We can view a subjective context as presenting a sequence of private states, where the experiencers and attitudes of some are only implicit (in the case of represented thoughts). In addition to discourse relations among sentences as wholes, there can be discourse relations among objects of  Janyce M. Wiebe Tracking Point of View in Narrative private states, even a hierarchical structure involving several such objects. As a concrete example, the contrast signaled by the cue-phrase 'but' appearing in a represented thought may be between a represented thought and the object of a previous private-state report. One example of a potential effect on pronoun resolution: if a sentence s is to be interpreted as a represented thought, and is to be incorporated into a discoursesegment structure among previous private-state objects, then pronouns in s can be resolved against the focus space(s) corresponding to the private-state-object segments (as in Grosz and Sidner's \[1986\] theory). Of course, determining whether or not s is to be so interpreted is a difficult problem. But such interactions among POV, discourse structure, and anaphora resolution might be usefully cast as constraints, with evidence regarding interpretation with respect to one limiting the options to be pursued for the others.</Paragraph>
      <Paragraph position="16"> The psychological POV is related to other pragmatic and discourse phenomena.</Paragraph>
      <Paragraph position="17"> Subjective contexts are paragraph-level analogs of opaque contexts in belief reports (Wiebe \[1991\] specifically addresses this issue). In addition, the discourse phenomena identified by Fauconnier (1985) are similar to the psychological point of view. Just as a private-state report can begin a discourse segment in which subsequent sentences are understood to continue a character's point of view, an adverbial such as &amp;quot;in 1969&amp;quot; can begin a discourse segment in which subsequent sentences are understood to refer to events that occurred in 1969, even though the date is not subsequently mentioned. Or, consider discussing someone else's work, say Smith's, in a research paper or textbook. After an initial reference to Smith's work, you may go on to describe his or her theory without explicitly saying in each sentence that you are doing so (with a locution such as &amp;quot;Smith shows that,&amp;quot; &amp;quot;In Smith's theory, .... In Smith's algorithm,&amp;quot; or &amp;quot;According to Smith&amp;quot;) (William J. Rapaport, personal communication). Along with subjective contexts, an NLU system must be able to recognize such discourse phenomena in order to recover information implicitly communicated in the discourse. A detailed investigation of one of them suggests directions for investigating the others. The one investigated here, subjective contexts in particular kinds of texts, is a good one to investigate because it is possible to constrain the problem, because there are so many prototypical instances in those kinds of texts, and because there is a great deal of previous work in linguistics and literary theory to build upon.</Paragraph>
      <Paragraph position="18"> In conclusion, this paper presented an algorithm for tracking the psychological point of view in third-person fictional narrative text. The algorithm is based on regularities, found by extensive examination of naturally occurring narratives, in the ways that authors manipulate point of view. The algorithm has been implemented, preliminary empirical studies have been performed, which support the algorithm, and psychological experimentation is continuing. This is the first detailed computational approach to the problem of tracking the psychological point of view.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="266" end_page="266" type="metho">
    <SectionTitle>
Acknowledgments
</SectionTitle>
    <Paragraph position="0"> I wish to thank the members of the SUNY Buffalo Graduate Group in Cognitive Science and the SNePS Research Group for many discussions and ideas, and William J.</Paragraph>
    <Paragraph position="1"> Rapaport, Graeme Hirst, Peter Heeman, Russ Greiner, the anonymous reviewers, and especially Diane Horton for helpful comments on earlier drafts of this paper.</Paragraph>
    <Paragraph position="2"> This research was supported in part by</Paragraph>
    <Section position="1" start_page="266" end_page="266" type="sub_section">
      <SectionTitle>
National Science Foundation grants
</SectionTitle>
      <Paragraph position="0"> IST-8504713 and IRI-8610517, and the preparation of this paper was supported in part by the Natural Sciences and Engineering Research Council of Canada.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>