<?xml version="1.0" standalone="yes"?>
<Paper uid="P86-1021">
  <Title>THE INTONATIONAL STRUCTURING OF DISCOURSE</Title>
  <Section position="4" start_page="0" end_page="136" type="metho">
    <SectionTitle>
3. FO Synthesis
</SectionTitle>
    <Paragraph position="0"> To synthesize the fundamental frequency (f0) contours for the TNT script, we used the intonation synthesis program  2. That is, the tutorial is organized around a series of data processing tasks, which the subject is guided through. See \[18\] for discussion of the characteristics of task-oriented domain discourse.</Paragraph>
    <Paragraph position="1"> 3. The process by which speakers signal that they have (temporarily) finished speaking and by which hearers interpret such signals \[21\].</Paragraph>
    <Paragraph position="2">  described in \[22\] in \[23, 24\]. It permits explicit control over the different dimensions of variation in the intonation system. The dimensions we will discuss here are phrasing, pitch range, accent location, and tune. We illustrate each in our synthesis of the introduction to TNT: i. Hello.</Paragraph>
    <Paragraph position="3"> TISO F.96 H* L L~  2. Welcome to word processing.</Paragraph>
    <Paragraph position="4"> TISO F.g6 H* H* L L~ 3. That's using a computer to write letters and T136 F.90 H* H* H* H* H* reports.</Paragraph>
    <Paragraph position="5"> H* L L~ 4. Word processing makes typing easy.</Paragraph>
    <Paragraph position="6"> T136 F.96 H* H* H* H~ L L~ 5. Make a typo? T125 L* H* H H~ S. No problem.</Paragraph>
    <Paragraph position="7"> TII5 F.96 H* H* L L~ 7. Just back up, type over the mistake, TI15 F.96 H* H* L H~ H* H* L H~ and it's gone.</Paragraph>
    <Paragraph position="8">  In Figure 1 and in all figures below, 'T' indicates the top of the pitch range in Hz, 'F' indicates amount of compression of the pitch range at the end of declarative phrases, 'H' and 'L' indicate high and low tones, '*' indicates a tone's alignment with a stressed syllable, and '%' indicates a phrase boundary tone. We discuss these phenomena and our notational system in more detail below.</Paragraph>
    <Section position="1" start_page="136" end_page="136" type="sub_section">
      <SectionTitle>
3.1 Phrasing
</SectionTitle>
      <Paragraph position="0"> The first dimension of variation, phrasing, may be indicated by a pause, by a lengthening of the phrase-final syllable, and by the occurrence&amp;quot; of extra melodic elements on the end of the phrase. Variation in phrasing is illustrated in Figures 2 and 3. 4 In Figure 2, line 8 is produced as a single phrase, whereas in Figure 3, And is set off as a separate phrase.</Paragraph>
      <Paragraph position="1"> One consequence of this strategy is that And becomes more prominent in the second version. Phrasing variation will not be of central concern here. Because of the syntactic simplicity of TNT, there were only a few cases where the phrasing could be varied in interesting ways.</Paragraph>
      <Paragraph position="2">  4. Note that phonetic transcriptions given in these and subsequent figures represent the somewhat eccentric output ofthe TTS system.</Paragraph>
      <Paragraph position="4"/>
    </Section>
  </Section>
  <Section position="5" start_page="136" end_page="138" type="metho">
    <SectionTitle>
AND, IT ELIMINATES RETYPING
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="136" end_page="136" type="sub_section">
      <SectionTitle>
3.2 Pitch Range
</SectionTitle>
      <Paragraph position="0"> When a speaker raises his/her voice, his/her overall pitch range - the distance betweer~ the highest point in the f0 contour and the speaker's baseline (defined by the lowest point a speaker realizes over all utterances) -- is expanded. Thus, the highest points in the contour become higher and other aspects are proportionately affected. Figure 4 shows an f0 contour for line 1 in the scriot above in the default pitch range used by TTS.</Paragraph>
      <Paragraph position="2"> The shape of the actual contour is the same as in Figure 4 but its scaling is different. Changes in pitch range appear to reflect the overall structure of the discourse, with major topic shifts marked by marked increases in pitch range.</Paragraph>
      <Paragraph position="3"> In addition to variations in overall pitch range, the intonation system exploits a local time-dependent type of pitch range variation, called final lowering. In the experiments reported in \[24\], it was found that the pitch range in declaratives is lowered and compressed in .anticipation of the end of the utterance. Final lowering begins about half a second before the end and gradually increases, reaching its greatest strength right at the end of the utterance. This phenomenon appears to reflect the degree of 'finality' of an utterance; the more final lowering, the more the sense that an utterance 'completes' a topic is conveyed. Contrast Figures 6 and 7.</Paragraph>
      <Paragraph position="5"> In the notational system employed here, T represents the topline, of a phrase -- the maximal value for the f0 contour in the phrase. F expresses the amount of final lowering in terms of the ratio of the lowered pitch range to the starting pitch range. The default value assumed below for T is 115 Hz and for F is 0.87.</Paragraph>
    </Section>
    <Section position="2" start_page="136" end_page="138" type="sub_section">
      <SectionTitle>
3.3 Accent
</SectionTitle>
      <Paragraph position="0"> Pitch accents, which fall on the stressed syllable of lexical items, mark those items as intonationally prominent. In line 16, for example, right has no pitch accent. If right were to be especially emphasized, it would have an accent. (In our notation, the absence of a specified accent indicates that a word is not accented; where we wish to highlight this point, we will employ '-' to mark a deaccented word.) The contrasting outcomes are shown in Figures 8 and</Paragraph>
      <Paragraph position="2"> In the first case, the last f0 peak occurs on work and there is a fall to a low pitch on right, then a rise at the end of the phrase. In the second case, the entire peak-fall-rise configuration occurs on the word right.</Paragraph>
      <Paragraph position="3"> There are six types of pitch accent in English \[23\], two simple tones -- high and low -- and four complex ones. The most frequently used accent, the simple high tone, comes out as a peak on the accented syllable (as, on right in Figure 9) and will be represented below as H*. The 'H' indicates a high tone, and the '*' that the tone is aligned with a stressed syllable. In some cases, we have used a L* accent, which occurs much lower in the pitch range than H* and is phonetically realized as a local f0 minimum. The accent on make in Figure 13 below is a L*. The other English accents have two tones. Figure 10 shows a version of the sentence in Figures 2 and 3 with a L+tt* accent substituted for both H* accents in the second phrase.</Paragraph>
      <Paragraph position="5"/>
    </Section>
  </Section>
  <Section position="6" start_page="138" end_page="139" type="metho">
    <SectionTitle>
AND, IT ELIMINATES
</SectionTitle>
    <Paragraph position="0"/>
    <Paragraph position="2"> Note that there are still peaks on the stressed syllables, but now a striking valley occurs just before each peak.</Paragraph>
    <Paragraph position="3"> In our synthesis of the TNT script, we have made extensive use of the type of accent transcribed in \[23\] as H*+L. This accent, like other bitonal accents, triggers a rule which compresses the pitch range on following material in the phrase, a phenomenon known as downstep or catathesls. For example, a simple contrast between H* H* and H*+L H*+L is illustrated in Figures 11 and 12 in two versions of the tutorial command to hit the 'remind'</Paragraph>
    <Paragraph position="5"> Figure 12. H*+L H*+L L L~o We have made particular use of downstepped contours such as this -- i.e., sequences of H*4-L tones -- which we will term H*+L sequence in the discussion below. (See Section 4.3.) The way a speaker is structuring a text helps to determine where pitch accents will fall, as a speaker indicates how referents of accented or deaccented items are related to other items in the utterance or in some larger context.</Paragraph>
    <Paragraph position="6"> In addition to pitch accents, each intonational phrase has a phrase accent and a boundary tone. These two extra tones may be either L or H. The boundary tone (indicated by '~o') falls exactly at the phrase boundary, while the phrase accent (indicated by an unadorned H or L) spreads over the material between the last pitch accent and the boundary tone. Each intonational phrase contains one or more pitch accents, a phrase accent, and a boundary tone.</Paragraph>
    <Section position="1" start_page="138" end_page="139" type="sub_section">
      <SectionTitle>
3.4 Tune
</SectionTitle>
      <Paragraph position="0"> A phrase's tune or melody is defined by its particular sequence of pitch accents, phrase accent, and boundary tone.</Paragraph>
      <Paragraph position="1"> Thus, H* L L~o represents a tune with a H* pitch accent, a L phrase accent, and a L~:~ boundary tone. This is an ordinary declarative pattern with a final fall. A interrogative contour is represented by L* H H~o. The contrast between these two melodies is illustrated in Figures 13 and 14. Figure 13 shows the actual</Paragraph>
      <Paragraph position="3"> With the declarative intonation characteristic of imperatives, 5 would probably convey that the hearer was being ordered to produce a typo. Roughly speaking, the tune appears to convey information about speaker attitudes and intentions (as, the speech act the speaker intends to perform) and about the relationship between utterances in a discourse.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="139" end_page="141" type="metho">
    <SectionTitle>
4. Intonational and Discourse Phenomena
</SectionTitle>
    <Paragraph position="0"> The major questions underlying our research are: First, what is the relationship between particular'-,intonational phenomena and particular discourse phenomena?~ For example, what discourse phenomena are associated with ehPS~iges in pitch range? With the accenting or deaccenting of particular lexical items? With choice of tune? More generally, we also characterize the contributions of these intonational phenomena in terms of the theory.of, discourse structure developed in \[15\], by relating intouationM .contributions to aspects of intentional and attentional discourse structure. Second, how do int0na~ionul features such as these interact with one another? Does an expansion of pitch range affect the interpretation of a rise-fall-rise contour \[25\], for example, and if so how? Third, when several discourse features predict conflicting intonational strategies, how is a decision made? When the information represented by a single referring expression, for example, is both 'given' and 'contrastive' -- and thus both deaccentable and accentable -- how is the choice to be made?</Paragraph>
    <Section position="1" start_page="139" end_page="140" type="sub_section">
      <SectionTitle>
4.1 Pitch Range Manipulation
</SectionTitle>
      <Paragraph position="0"> Students of discourse commonly observe that discourses often exhibit a hierarchical structure - into major topics, their subtopics, sub-subtopics, and so on. In task-oriented domains, it has been claimed that this structure reflects the hierarchical structure of a task and its subtasks \[18\]. So, for example, the TNT introduction above might be segmented as follows (where utterances are labeled by line number): 5  This bracketing schema defines a discourse segment as any node together with all the nodes it dominates; for example, lines 1-11 form a segment, as do lines 14-16, and so on. An alternative depiction of the hierarchy above would be \[{0}\[1/2 3 \[4 \[5 6 7\] \[8 9 10 11\]\]\] \[12 \[13\] \[14 15 16\]\]\]fi Evidence for such hierarchical segmentation in general is found in instances of pronominal reference to referents linearly distant in the discourse; in such cases, a notion of hierarchical proximity appears plausible.</Paragraph>
      <Paragraph position="1"> Previous research \[12,14\] has observed that 'topic jump' can be signalled by raised pitch, as well as increased a~plitude and markers of self-editing, hesitation, and discontinuity -.and that pauses and changes in rate characterize segment boundaries. In our work with the TNT script, we found that a hierarchical segmentation of discourse can be marked by systematic variation in pitch range, which can signal movement betweeen levels in the segment hierarchy. In addition, by varying the amount of final rais'ihg or lowering at the end of phrases, we can indicate the degree of conceptual continuity between one phrase and the next. We have  developed algorithms for assigning pitch range and fiual raising/lowering in terms of the discourse segmentation.</Paragraph>
      <Paragraph position="2"> 5. We do not claim this is the only possible segmentation, only that it is a plausible one to convey.</Paragraph>
      <Paragraph position="3"> 6. Note that I and 2 are treated as a unit here, although they are synthesized as  separate phrases, since it seemed semantically correct.</Paragraph>
      <Paragraph position="4"> To illustrate the algorithms, we relate the TNT introduction presented in Figure 1 to its segmentation in Table 1. When the introduction is synthesized using the TTS'default pitch range of 75-11~ Hz, the topline for each utterance will remain around 115 Hz. However, the hierarchical relationship schematized above among the various segments may be signalled more clearly if the pitch range is varied. In our version of the script, each segment boundary is marked by a variation in pitch range which correlates with the segment's position in the overall discourse. So, major boundaries are denoted by the largest increases, with smaller increases marking subsegment boundaries, and so on. The segment beginning at 1, for example, is marked by raising the f0 topline to 150 Hz; that beginning at 14, by raising the topline to 136 Hz; and that beginning at 15, by raising the topline to 125 Hz. 7 Human speakers do seem to employ a wider spectrum of pitch range variation than we have been able to use in synthesis, however.</Paragraph>
      <Paragraph position="5"> We would claim that the appropriateness of changes in pitch range is a function of the segmentation hierarchy -- and is not inherent in the utterance in isolation. Our algorithm for pitch range assignment can in fact enforce one segmentation of a given discourse over another and, in so doing, can disambiguate among potentially ambiguous reference resolutions. For example, It in line 7 of Figure 1 coindexes mistake, while it in line 8 coindexes word processing. A simple linear approach to reference resolution (such as \[26\] ) would have the second coindexical with the previous noun-phrase (np), mistake, but a hierarchical approach to discourse structure holds out the possibility that a referent in a segment dominating the current segment may also provide a referent \[18\], as, in fact, is the case here. While a little thought will make the appropriate referent clear, it is clearer when line 8 is produced with a larger pitch range to signal the beginning of a new subsegment of the segment headed by 4. By so doing, we lessen the possibility that a referent for this it will be sought in lines.:5-7..The most likely candidate, found in 4, is now both intonationally and conceptually 8's superordinate discourse segment.</Paragraph>
      <Paragraph position="6"> While an increase in the pitch range indicates segment boundaries, a decrease in the final lowering effects can indicate the absence of such boundaries, and thus indicate that a given utterance and one which follows it are part of the same segment. So, manipulation of final lowering can also serve to indicate discourse structure, by identifying the internal structure of segments. For example, at one point in the TNT script, the following utterance constitutes an entire discourse segment, so it has ,the default final lowering (F=0.87); in consequence, the L~ tone at the end of had will be only 87% as high as it would have been if final lowering had not applied.</Paragraph>
      <Paragraph position="7">  Here, the same utterance is synthesized with less final lowering -the L~ tone at the end of had, in particular, will attain 93% of its target height. In this segment, the first line does not end the segment. We further propose that the degree of final lowering may correlate with the utterance's position in the discourse hierarchy. Specifically, we suggest that minimal final lowering may indicate a 'push' onto the segment stack and greater degrees of final lowering Our choice of ranges was determined in part by the TTS synthesizer, which tends to sound best when its topline ranges between 115-150 Hz. Preliminary investigation of pitch range changes in human speech indicates that, for male speakers, these choices are reasonable. Note also that it is the relationship among different range levels, not the actual values in Hz, which is important here.</Paragraph>
      <Paragraph position="8">  may be associated with 'pops' of this stack. In our synthesis of the TNT text, we have varied degree of final lowering for such 'pops' based upon the level of the segment which this utterance 'completes' (or, equivalently, the level of the segment the next utterance begins). So, to determine the amount of final lowering to assign when synthesizing line 7 in the TNT introduction, we first determine whether it completes a segment (representing a pop) or not (representing a push). If the former, we may note either that it completes the segment begun at line 5 (with a topline of 136 Hz), or that the subsequent segment is begun (by line 8) with a topline of 136 Hz. We assign final lowering of 0.90 when synthesizing line 7 based on either observation; this rather large amount (close to the synthesizer's default maximum of 0,87) conveys a relatively important change of subtopic within the larger discourse segment by indicating rather more disjunction than we would want, for example, between lines 9 and 10.</Paragraph>
      <Paragraph position="9"> We are currently testing the associations between pitch range/ final lowering variation and discourse structure proposed above in several ways: by pitch-tracking a large corpus of natural speech, 8 by recording and analyzing subjects reading structured texts, and by asking subjects to perform tasks such as reference resolution from texts synthesized with varying pitch ranges.</Paragraph>
    </Section>
    <Section position="2" start_page="140" end_page="141" type="sub_section">
      <SectionTitle>
4.2 Accent Placement
</SectionTitle>
      <Paragraph position="0"> Accent placement, too, can convey information about the structure of a discourse. Traditionally, it has been noted that stress, or accent, can convey information about the focus of an utterance, about given or new information in the discourse, o about parallelism, or about contrastiveness. In more general terms, one might say that accent placement appears to be associated with Grosz and Sidner's \[15\] attentlonal structure -- the salience of discourse entities, properties, relations, and intentions at any point in the discourse. We have particularly noted that the decision to accent or deaccent some item is sensitive to the position of that item in the discourse structure - that is, just as salience is always determined relative to some particular context, accent placement must be determined with respect to the segment in which the accentable item appears. We take the position that it is the signaling of salience relative to the discourse segment that produces the secondary effects of given-new distinction, topic-hood or contrastiveness, and the favoring of one reference resolution over another. One of the more common observations about the role of accent placement and the structuring of discourse is that accent can mark some item in the discourse as in focus - i.e., as 'what is being talked about' \[28,29} -- particularly when syntactic or thematic information might predict otherwise. For example, in the following instructions, erase is accented in line 2 to indicate that the action of 'erasing' is the focus of the current task.</Paragraph>
      <Paragraph position="1">  1. Type hello.</Paragraph>
      <Paragraph position="2"> H* H* L L~ 2. Next, let's erase hello.</Paragraph>
      <Paragraph position="3"> H* L H~ H* H* - L L~ 3. Hit hint.</Paragraph>
      <Paragraph position="5"> For similar reasons, we accent hello in line 1 and deaccent it in line 2.</Paragraph>
      <Paragraph position="6"> While focus considerations clearly influence accent placement, determining accent placement solely on the basis of utterance-level focus (as proposed in Gussenhoven \[29} and Cull-From interviews collected by A. Kroch and G. Ward and from recordings made of a radio financial advise program by J. Hirschberg and M. Pollack. Prince \[271 notes that the 'given/new' distinction has been variously defined as predictable/unpredictable, salient/not salient, shared/not shared knowledge, and proposes a more complex taxonomy of 'assumed familiarity' classifying discourse entities as new, inferrable, or evoked (either textually or situationally). This is closely related -- and often confused with -- the notion of utterance topic/focus.</Paragraph>
      <Paragraph position="7"> cover and Rochemont \[30\] ) is insufficient. Considerations such as the given/new distinction play an important role.</Paragraph>
      <Paragraph position="8"> Speakers typically deaeeent given information and accent new information, as when the 'new' information typing is accented and the 'old' word processing is not in line 3 below: Welcome to word processing.</Paragraph>
      <Paragraph position="9"> B* H* L L~ That's using a computer to write letters H* H* H* H* H* and reports.</Paragraph>
      <Paragraph position="10"> H* L L~ Word processing makes typing easy.</Paragraph>
      <Paragraph position="11"> H* H* H* L L~ Note that these items are marked as 'given' and 'new' within the current segment -- although they may have other status within the larger discourse. Furthermore, items appear 'given' or 'new' not simply because of prior mention (or lack thereof) in a context but via 'physical co-presence', where speaker, hearer, and referents are physically and openly present together; \[31\] shared world knowledge; or conceptual proximity \[11. For example, the tutor can treat m as given in the following text because the student has just (incorrectly) typed mary; the character 'm', the student, and the tutor, are thus physically copresent.</Paragraph>
      <Paragraph position="12"> Oops. capital m.</Paragraph>
      <Paragraph position="13"> H* L H~ H* - L L~ The new information is that 'm' is to be capitalized. Thus cap~al is accented. Similarly, in the introduction to the tutor presented in  We also examine how pronominalization interacts with accent placement. Since the ability to pronominalize is itself a standard test of givenness, prowords, like other given items, are commonly deaecented. If they are accented, the hearer may draw very different conclusions from an utterance. The following utterance, for example, may well convey an instruction to type the word something or even a reprimand for not typing anything yet: Let's hegln by typing something.</Paragraph>
      <Paragraph position="14"> H* H* H* Since the TNT script employes little pronominalization, we often use deaccenting to 'intonationally pronominalize' repetitions of lexical items.</Paragraph>
      <Paragraph position="15"> Accent can also signal that a discourse referent other than that which would be 'most likely' without special accentuation should be sought, as in: I.</Paragraph>
      <Paragraph position="16"> 2.</Paragraph>
      <Paragraph position="17"> We can't answer questions, if you are confused.</Paragraph>
      <Paragraph position="18"> H* H* H* L H~ H* L L~ We have to le~ the computer do all the teaching.</Paragraph>
      <Paragraph position="19"> H* H* H* H* L LS Here (and in particular at line 1), we is intended to refer to the humans supervising the testing of the tutor, although these humans have not previously been mentioned in the script. However, this reference might easily be interpreted as referring to the  tutorial system itself. Since pronouns -- as 'given' information -are commonly deaccented, we accent this one to indicate that an 'unusual' referent should be sought, 10 So, both accent placement and manipulation of pitch range can be used to reorder the list of potential referents for a given referring expression.</Paragraph>
      <Paragraph position="20"> FinMly, eontrastiveness or parallelism may also be communicated via accent. For example, second is accented in 3, although it is certainly given in this segment (via mention of second draft in 1):  I. Nned a second draft? L* L* L* H H~ 2. No problem.</Paragraph>
      <Paragraph position="21"> H* H* L L~ 3. Just change the first, and you've got the second.</Paragraph>
      <Paragraph position="22">  H* H* L H~ H* H* L L~ Note that, while second may be 'given' at the discourse segment level, the decision to accent it is based on contrast within a smaller context, 3. Furthermore, if this function of accent is ignored, contrastiveness may be inferred incorrectly. If we accent we in the last line of the tutorial introduction WE will help you out, for example, the student would be entitled to infer that others will not be helpful.</Paragraph>
      <Paragraph position="23"> We are currently developing Mgorithms for determining accent placement, based upon the interaction of focus, given/new, parallelism, contrastiveness, and pronominal reference within segment and phrase.</Paragraph>
    </Section>
    <Section position="3" start_page="141" end_page="141" type="sub_section">
      <SectionTitle>
4.3 Choice of Tune
</SectionTitle>
      <Paragraph position="0"> It is now widely accepted that the overall melody a speaker employs in an utterance can communicate some semantic or pragmatic information. However, since there are few particular tune types for which we can specify with any confidence just what the meaning might be, it is difficult to generalize about what type of information tunes in generM can convey. From those tunes whose 'meaning' seems fairly well understood -- namely, declarative, yes-no question \[23\], surprlse/redundancy \[10\], contradiction contour \[33\] rise-fall-rlse \[25\], and continuation rise \[34,35\] contours -- we propose that tunes convey two sorts of information about discourse.</Paragraph>
      <Paragraph position="1"> First, we believe that contours can convey propositional attltudes n the speaker wishes to associate with the propositional content of an utterance. For example, the speaker may wish to convey that s/he knows x, or that s/he believes x, or that s/he is uncertain about x, or that s/he is ignorant of x. In the case of H*+L sequences, it appears that a speaker may convey his/her (propositional) attitudes about a hearer's (propositional) attitudes toward an utterances. This tune seems to indicate the speaker's belief that the speech act s/he is performing is superfluous. For example, a speaker may employ it to convey that the propositional content of his/her utterance is already known or would be obvious to the hearer (who, of course, may or may not be attending to it).</Paragraph>
      <Paragraph position="2"> Note that the speaker may or may not believe that this information is known, in order to wish to convey this meaning. Particularly in pedagogical texts, this contour seems appropriate to introduce straightforward material, as in the following instruction to hit the remind key.</Paragraph>
      <Paragraph position="3"> Remind, tells you again what to do If you forget.</Paragraph>
      <Paragraph position="4"> R* L H~ tt* H* H* H* L L~ Hit re~ind.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="141" end_page="141" type="metho">
    <SectionTitle>
H*+L H*+L L L~
</SectionTitle>
    <Paragraph position="0"> However, an H*+L sequence is not appropriate in the following similar exchange: I0. The standard example of accentuation influencing pronominal reference resolu- tion in this way is'john hit Bill and then HE hit HIM 1321. 11. Propositional attitudes include knowing, believing, intending, uncertainty, and ignorance.</Paragraph>
    <Paragraph position="1"> Next, let's erase hello.</Paragraph>
    <Paragraph position="2"> H* L H~ H* H* L L~ Hit hint.</Paragraph>
    <Paragraph position="3"> H* H* L L~ In general, such contours do not seem felicitous when the utterance conveys information which the speaker believes will be unexpected for the hearer. Here tune choice may reflect attentional as well as intentional aspects of the discourse structure. Like the deaccenting of references to given items, tt*+L sequence contours seem to convey 'givenness' at a more general level.</Paragraph>
    <Paragraph position="4"> Second, we believe that tune can convey the speaker's commitment to some semantico-pragmatic structural relationship holding between the propositional content of utterances (as, that one 'completes' another or is subordinate to another). Many such relations have been proposed in textual analysis \[36,37,15\]. In the phonological literature, continuation rise has been commonly associated with some sense of 'continuation' or 'more to come' \[34\]. We have found, howe-rer, that this contour can be characterized more precisely as conveying a subordination relationship between the phrase uttered with continuation rise and other utterances in the discourse segment. For example, if the second phrase of line 1 is uttered with continuation rise, then this utterance appears to be subordinated to 2.</Paragraph>
    <Paragraph position="5"> I. We can't answer questions, if you are confused.</Paragraph>
    <Paragraph position="6"> H* H* H* L H~ H* L H~ 2. We have to let the computer do all the teachlng.</Paragraph>
    <Paragraph position="7"> H* H* H* H* L L~ 3. But if the computer is not working right, we will help H* H* L H~ H* H* you out.</Paragraph>
    <Paragraph position="8"> H* L L% That is, 2 'completes' I. Without continuation rise on 1, all three utterances will appear to have equal status in th'esegfiaent. Furthermore, continuation rise is not felicitous in ~.ll ~ontexts in which the simple sense that 'there is more to come' clearly should be appropriate; for example, continuation rise over 3 -- at the end of the tutorial introduction -- seems quite odd, even though more will clearly follow.</Paragraph>
    <Paragraph position="9"> In synthesizing the TNT script, we have employed only a small subset of possible English tunes. Analysis of the 'meaning' of additional tunes is part of our future research. More generally, we must examine how structural relationships conveyed by tunes such as H*+L sequence are ~.ssociated with those conveyed by pitch range.</Paragraph>
    <Paragraph position="10"> We have described certain mappings between intonational features and discourse phenomena, associating pitch range variation with the identification of discourse segments and with their internal coherence; accent with types of information status such as topic (focus) and the given/new distinction, with reference resolution and with contrastiveness; and tune choice with~ ~he, relationships among propositions in the discourse as well as w~l~b.,~ome propositional attitude the speaker wishes to associate with :those propositions. It appears that pitch range and accent placement are most closely associated with a diseourse's attentional structure, while tune choice is more closely associated with its intentional structure. However, clearly this picture is too simple. SeverM intonational features may be used together to create some discourse effect; moreover, in some cases two distinct intonational phenomena seem to produce discourse effects that seem intuitively to be closely related. And sometimes several discourse phenomena may indicate conflicting intonational strategies. These problems are the subject of our future research.</Paragraph>
  </Section>
class="xml-element"></Paper>