<?xml version="1.0" standalone="yes"?>
<Paper uid="W90-0122">
  <Title>The Computer Generation of Speech with Dlscoursally and Semantically Motivated Intonation</Title>
  <Section position="3" start_page="164" end_page="167" type="metho">
    <SectionTitle>
3. Modelling Intonation
</SectionTitle>
    <Paragraph position="0"> 3.1. ,am Overview of the Generation of Intonation Let as imagine that Ivy (the 'person' of whose ufind GENESYS models a part) is about to generate a sentence, Let us suppose that she is being consulted by the Personnel Officer of a large institution, who draws regularly on her specialist knowledge and advice, and that he has just asked her Wghere does Peter Piper live?&amp;quot;.</Paragraph>
    <Paragraph position="1"> (We shall come later to how intonation is represented.) Like most human users of language, Ivy makes reasonable assumptions about (loosely, she 'knows') where she is in any current transaction (e.g. at the start, in the middle or at the end), and where she is in the current exchange. This affects the pitch level of what she says. She needs to choose a tone (the change in pitch marked by a stepping or a slide on the tonic syllable) which will express the MOOD of the final matrix clause of her sentence. ('Matrix' here means 'at the top layer of structure'.) She nee&amp; to locate that tone on an item which will be thereby marked as new Information. She needs to decide if it is to be presented simply as 'neW, or as 'contrastively new' (in the terms used here).</Paragraph>
    <Paragraph position="2"> And she needs to deride on the Information status of any chunks of information that are to be presented as separate from the main Information ear of the clause. (The information that guides these choices comes from various aspects of the higher belief system, which there is unfortunately no space to discuss here.) As we shaft see, these various components of the semantic level of intonation account, in a different way from the usual approach in British intonation studies, for Halliday% well-known triad of TONE, TONALITY and TONICITY. While it is perfectly possible m talk about the contrasts in intonational form to which these three rder as 'systems', I suggest that it is more insightful to take, as the level of contrasts to be modelIed in system networks, the meanings that lie behind (or, ia the SFG metaphor, above them). These semantic features are then realized in the purely intonational contrasts of TONE, TONICITY and TONA.LITY.</Paragraph>
    <Paragraph position="3"> The accounts of the various aspects of intonation in what follows will inevitably be introductory, and may to the specialist appear simplistic. A somewhat fuller treatment is given in Teach and Fawcett 1988, and a very fur treatment is given in Tenth 1987, which includes summaries of relevant work by other intonation speC/islists.</Paragraph>
    <Paragraph position="4"> (I omit here, for reasons of space, a specification of how one might model the way in which the position in a transaction and an exchange affects intonation.) Let us assume, then, that Ivy is preparing a response to the Personnel Officer's question, using the information that, while Mr Peter Piper's address is currently 11 Romilly Crescent, Canton, Cardiff, he is moving from there after one month.</Paragraph>
    <Paragraph position="5"> In discourse planni.g terms, she chooses that her move will be a 'support' for a 'solicit information', gad that the act at the head of the move is a 'give new content ) (see Fawcett, van der Mije and van Wissen 1988). As wc shall see, these choices pro.</Paragraph>
    <Paragraph position="6"> select in the MOOD network of the lexieegrammar the features \[information\] and \[giver\]. But first there is a more basic system to consider.</Paragraph>
    <Paragraph position="7"> 32. The MODE System The inidal rule of the semantics of the lexico-grammar to be considered here is: situation -&gt; &amp;quot;MODE'&amp; 'TENOR'&amp;'CONGRUENCZ~SIT'.</Paragraph>
    <Paragraph position="8"> Thus means that, for any 'situation' ( roughly = 'proposition') that you are generating, you must make choices in all three of the systems named.</Paragraph>
    <Paragraph position="9"> (Notice, then, that 'parallelism' lies at the heart of the grammar.) Here we shall be concerned only with the MODE system. (It is from CONGR-UENCE SIT that the main part of the network is entered, to generate configurations of participant roles, such as Agent and Affected, and choices in MOOD, such as 'information seeker', and very many others.) The MODE system is very simple: 'MODE' -&gt; 70% spoken / 30% written.</Paragraph>
    <Paragraph position="10"> This means: 'In the MODE system, you must choose between generating a spoken output (for which under random generation there is a 70% probability) and generating a written output (which carries a 30% probability). Clearly, sinc~  Ivy is in a spoken interaction, she will be strongly disposed to select \[spoken\]. but in principle she need not We shall not discuss here the interesting reasons for and ~in~t introducing this system to the lexiC/o-grammar itself, except to point to two 91~,niEcant advantages that it brings. Nor, unfortunately, is there space to discuss the roles of the probabilities and the ways in which they are assigned (sometimes simply a guess at the overall pattern for central types of text; somet;mes based on textual studies).</Paragraph>
    <Paragraph position="11"> (The ne,~ few lines presuppose some familiarity with systemic grammars; for those without this knowledge it may be advisable to re-read this section after seeing the working of the examples.) What is the role of this system? Fh'st, it enables the grammat-builder to refer, at any point on this initial pass through the system network, to whichOver feature in this system has been chosen as an entry condition to a later system. In other words, where there is greater richness of choice in me aning in the spoken mode (as is typically the case with m~nings real|Ted in intonation, as gain.st those reAI;~,~d in punctuation), we can ensure that those systems are only entered when the feature \[spoken\] has been chosen. We shall shortly see the great value of this. Second, the 'MODE' system enables us to refer, at any point in a realization rule, on this or any subsequent pass through the network, to this feature as a conditional feature for the realization of some othex feature, In other words, we can ensure that if \[spoken\] has been chosen the realization will take the form of intonation, and if \[wrkten\] has been chosen it is expressed in punctuation. Both of these faclh'fies contribute greatly to the elegant operation of the lexieogramm~tr as a whole, both in me~nin~ rC/91h'?d ill intonation and in many other ways.</Paragraph>
    <Section position="1" start_page="165" end_page="165" type="sub_section">
      <SectionTitle>
3.3. The Sentence Generator= MOOD
</SectionTitle>
      <Paragraph position="0"> The unmarked choice in the 'CONGRUENCE SIT' system is, unsurpri~ngly, \[congruent_sill. This is the choice that opens up the whole array of me~-;,~gs a.~sociated with realization in a clause, and many parallel systems follow. Among these is the MOOD network. This is a fairly large and complex network of meanings, and these are re~l;-ed partly in syntax, partly in items (such as &amp;quot;please'), and partly in tone (= variation in pitch). The network is too large and complex to present here, but we shall trace a route through it that shows why it is central to an understan,4;-g of intonation. The first options in the current GENESYS network are shown below:</Paragraph>
      <Paragraph position="2"> 70% giver (1.2) / 30% seeker (16).</Paragraph>
      <Paragraph position="3"> The second line reads:' In the MOOD(A) system you must choose between the feature \[information\], which overall has a 90% probability of being selected, and \[directive\], which has only a 10% probability. As so often, the choice of a single feature leads to further parallel systems, one of which continues the MOOD network itself.</Paragraph>
      <Paragraph position="4"> The last line in the above rules exemplifies the use of numbers in brackets after the features; it is the number of the realization rule for the feature concerned. What will this look like? Here is a slightly simplified version of the realization rule for \[giver\]:</Paragraph>
    </Section>
    <Section position="2" start_page="165" end_page="167" type="sub_section">
      <SectionTitle>
1.2 : giver :
</SectionTitle>
      <Paragraph position="0"> if falls 'Z' and (simplex sit or final co ordinated situation) ten on r st3,ass itten then 'E' &lt; &amp;quot;!')&amp;quot;'. if on,~first_pass spoken then 'E' &lt; The. effect of this rule is on the 'Ender' (i.e. 'E', the last element in the structure of the clause). If \[written\] is chosen in the 'MODE' system it is expounded by a full stop (Br. E. for 'period'), but if the choice is \[spoken\] it is expounded by a final intonation unit boundary, i.e. \[. However, says the rule, neither realization will occur unle~ the clause (1) directly fdls the element 'Sentence' (represented by'Z' a~ an approximation to sigma) and (2) is 'simplex ~, i.e. is not co~ordinated with one or more other clauses or, if it is, is the final clause.</Paragraph>
      <Paragraph position="1"> This may seem a surpr/~gly complex rule to those in NIP used to working with minigrammars. But this is typical of the working level of complexity in a natural language, and those who ate used to working with the problems of building broad coverage grammars will appreciate that this is not a particularly complex rule. In the case of our example the effect is to give to Ivy's output a final intonation unit boundary.</Paragraph>
      <Paragraph position="2"> We come next to an example of the value of being able to use of the feature \[spoken\] as an entry condition to a system. This is necessary because the MOOD network also builds in variables in 'key' (in the sense of HaRiday 1970); i.e. finer choices w/t h;. the MOOD options. These correspond to what Tench treats separately as variations in attitude. While accepting the 'view that these more delicate choices can be seen as  serving a separate ftmction from the function of the basic tone, the fact is that in any systemic computational implementation the way in which they enter the choice systom is simply us more delicate choices that are directly dependent on the broad choice of meaning realized in the broad tone. The range of such delicate variations appears to be potentially different for the various meani,o~ (see further below). In the systems given below, note ~ high probability of choosing \[assertive\] followed by \[neutral\].</Paragraph>
      <Paragraph position="3"> giver &amp; spoken -&gt; 70% assertive / 15% deferring (1.21) / 15% withjeservation (1.22).</Paragraph>
      <Paragraph position="4"> assertive. &gt; 2% very strong (L23) / 8% strong (1.24) / 60% neutral (1.25) / 30% mild (1.26).</Paragraph>
      <Paragraph position="5"> In an intermediate level model (such as Prototype Generator 2 (PG2), which is the most advanced version of GENESYS currently implemented) we need only relatively simple rules such as the following: 1.21 : deferring: if fills 'Z' and on first p~ spoken and (simplex sit or\]inal~co~,rdinated situation) then '2-' by 'NT.</Paragraph>
      <Paragraph position="7"> if fills '~' and on_first_pa~ss spoken and (simplexsit or fmaI_co-ordinated skuation) then 'L2' by 'hiT '.</Paragraph>
      <Paragraph position="8"> And so on, for \[very strong\] (realized by '21'), \[strong\] (realized by ~/'), \[neutral\] (realized by T) and \[mild\] (realized by '1-').</Paragraph>
      <Paragraph position="9"> Here we are using a numerical notation for tones that goes back to an earlier tradition even than Hall\[day's description (1967, 1970), though it has much in common with Hallida/s. (Readers from the American traditiOn used to an iconic representation may hate some adjustments to make in interpreting the notation. But there should be act fundamental diffu:ulty; Hallida/s description has been widely used (and indeed tested) on American and Canadian English.) I give next a brief summary of the differences between the scheme for tones used here and Hall\[day's well-lmown scheme (1967, 1970).</Paragraph>
      <Paragraph position="10"> Tench's (and so my) numbers '1' and '2' correspond to HoJliday's usage, as do the use of '+' and '-'. But Halt\[day% 'Tone 3' is seen as a variant of our Tone 2; Hallida/s Tone 4 (a fall. rise) is represented by '12'; and his 'Tone 5 (a ris~fall) is shown as '21'. Tench's general descriptions of the tones in words (1987) imply  four pitch level.% and I therefore use the following labels for the model implemented: base, low, mid and high. The four levels in turn provide a framework for describing three types of pitch change. It will be helpful for what follows to set them out as three 'scales'; these descriptions of the tones are in effect source material for writing realization rules. (1 have given these scales informal semantic labels; these are not intended to correspond directly to the features in the MOOD network encountered so far, but to evoke features from various parts of the network, including the many options dependent on \[directive\]). F;uaily, let me remind yon that we are not at this point trying to account for all tones, but only for those that carry the MOOD of a matrix simplex or final clause. This dear separation of the ways in which tones are generated is a key feature of the present proposals. We shall come shortly to some of the ways of generating appropriate tones for some of the other positions in which tones occur.</Paragraph>
      <Paragraph position="11"> The 'assertive' scale (Tones 21 and 1): Tone 21: rise-fall (rise to high plus fall to base) Tone 1 +: high-fall (fall from high to base)  Tone 1: mid-fall (fall from mid to base) Tone 1-: low-fall (fall from low to base) Also (see below): Tone 21.: low rise.fall (lower version of Tone 21)* The 'deferring' scale (Tone 2): Tone 2+: high.rise (rise from base to high) Tone 2: mid-rise (rise from base to mid) Tone 2-.: low-rise (rise from base to low) The 'implication' scale (Tone 12): Tone 12: fall-rise Also (see below): Tone 12-: low fall-rise (lower version of 12)* * Tench suggests that these are variants of Tone 21 and 12 that additionally signal 'emotional  involvement'.</Paragraph>
      <Paragraph position="12"> As will be clear, Tench and I propose a modification to HaUiday% basic set of contrasts in TONE, such that Hallida/s Tone 5 is seen as an extreme form of Tone 1. This fits naturally with the semantics of these tones. In a somewhat similar way, Tench treats Hallida/s Tone 3 (a low rise) as a variant of Hall\[day's Tone 2, under the rubric of 'deference to the listener', and we adopt this too in COMMUNAL But note that, while that kind of semantic description holds good for Tone 2s (Hall\[day% Tone 3s) in the sentence-final position, I shall suggest other means of generating them in non-final positions. In the present system there are no 'double tone groups', such as Hall\[day's Tone 13 (i.e. a Tone I to realize the main MOOD meaning, followed by a Tone 3 (here 2) for 'supplementary information'.) Such final Tone 2s will be generated in a similar way to that to be illustrated in section 3.5 below for initial Tone 3s (and, as we shaU see, 12s). Finally, note that I include h~e one option that Tench includes under 'stat~ of information'. This is his 'implication', re~l;~C/d in Tone 12, i.e. a fall-rise. This is Halliday~s Tone 4, which he characterizes as (among other descriptions) 'with reservation'. This tone occurs both as a carrier of MOOD and otherwise; it is with the former tiutt we are concerned here. It seems plausible to treat it as a variant that can be chosen as an alternative to the basic falling and rising tones recognized by both Halliday and Tench, and 1 have therefore incorporated it in the overall MOOD network.</Paragraph>
    </Section>
    <Section position="3" start_page="167" end_page="167" type="sub_section">
      <SectionTitle>
3.4. Focus of Information
</SectionTitle>
      <Paragraph position="0"> 3.4.1. The Line of Approach to the Problem I shall present here a somewhat novel approach to the relationship between the two sets of phenomena described by both Halliday and Tench as TONALITY and TONiCITY, TONALITY is typically thought of as 'cutting up a string of words into intonation ,mirA' ~ench's term; Halllday's is 'tone groups'), with each intonation unit realizing one information unit. The problem, when one is approaching the question from the angle of generation, is that there is no string of words to eat up - not, that is, until the senten~ has been generated. We therefore need to look for a semanflC/ approach to the problem, My proposal is that it is helpful to start not with TONALITY but TONICITY.</Paragraph>
      <Paragraph position="1"> TONICITY is the placing of the tonic on a syllable. The item so markfd is shown to be being presented as new Information - and !h{s iS a semantic concept. ('New' information is information presented as 'not recoverable.') But a further problem arises, in that linguists reccgni*e both 'marked' and '~1~marked' tonicity. 3.4.2. Generating Marked Tonidty Marked tonidty occurs when the item cont~;-i-g the tonic syllable is presented by the speaker as 'contrastivcly new'. Unmarked toniclty occurs when there is no marked tonicity (which is by far the most mual case); we shatl return to this shortly. Marked tonidty is handled in GENESYS ha the following way. In principle, any pathway through the system network that results in the generation of a formal item will lead to a system of the following form (where is the current terminal feature): x- &gt; notcontrastively_new / contrastively_new.</Paragraph>
      <Paragraph position="2"> The realiTafion of \[contrastively..new\] is that a contrastive tonic is conflated with the element of structure that the item expotmds. TI~ fimple version implemented in PG2 is as follows:</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="167" end_page="168" type="metho">
    <SectionTitle>
'INFORMATION FOCUS'-&gt;
</SectionTitle>
    <Paragraph position="0"> 99% no element marked as contrastively.new \] 1% elementmar'fked_as..~ont'rastively..new.</Paragraph>
    <Paragraph position="1"> element marked as contrastively new-&gt; 50% ~ontrast~e newness on_.~larity (18.1) / 50% contrastiv~newnes~on process (18.2) / 0% other.</Paragraph>
    <Paragraph position="2">  Realization rule 18.1 states the complex set of conditions for conflating a eontrastive tonic ('CT') with the appropriate element; for the POLARITY system ('positive' vs. 'negative') this is typically the Operator (which may have to be supplied by a 'do-support' rule) but it may be any one of several others, depending on whether or not the clause is moodless and, if not, whether a directive, and if not, what auxiliaries are realized, etc. The rule for presenting the 'process' (realized in the Main verb) as 'contrastively_nmv' is however extremely simple: 18.2 : contrastlve newness_ on_.process : 'CT' by 'M'.</Paragraph>
    <Paragraph position="3"> In the case of our c~mple, the choke is not to present any clement as conttrasfively new.</Paragraph>
    <Paragraph position="4">  How, then, should we generate unmarked tonicity? The answer is simple: as the default i.e. when there is no contrastive tonic, in other words, I want to suggest that ~mmarked tonicity is a formal phenomenon of intonation that does not express an active choice in meaning.. The relevant facts are well known, i,e., roughly, that what we hero term a nuclear tonic ('biT') fails on the last lexical item in the information unit. The question is: 'How can we define the intonation unit, in semantic terms?' The only contender as a semantic unit, in the GENF.SYS framework, is the situation, i.e. the semantic unit typically rvalizcd in the claw. The actual decision as to which item the unmarked tonic shall be assigned to gets made relatively late in the generation process. In GENESYS we simply have a list of the few dozen items generated by the lexicogrammar that cannot carry the unmarked tonic: roughly, the 'grammatical items' of English. Essentially, then, this default rule will insert one, and only one, nuclear tonic in each sentence. This will hold even when there are two or more co-ordinated clauses in that sentence, and/or one or more  This is a concept not disting~j;shed as a separate phenomenon in Halliday's ~eatment of intonation, but which Tench does treat separately. This dear separation of two semantically distinct phenomena was a significant help in developing the generative model proposed here. However the concept of 'status of information' is quite highly generalised, in the seine that it is not manifested in just one part of the overall network (as for example MOOD is). Specifically, we fred this option at many of the points where a unit is generated that is not the final matrix clause in the sentence. Many of these (thou# by no means all) have already been implemented in GENESYS, and the following are a representative sample.</Paragraph>
    <Paragraph position="5"> 3.5.2. The Co-ordination of Situations and Things One major source of multiple intonation units is co-ordination. Thus, when GENESYS generates co-ordinated clauses (realizing co-ordinated sltuatione) such as &amp;quot;Either Ivy loves Ike, or she loves Fred, or she doesn't love anybody.', she first recognizes at an abstract level that separate information Units are being assi~ed and then inserts, depending on whether the output is to be spoken or written, either (1) commas or (2) intonation trait boundaries and an appropriate tone such as Tone 2. We shall trot re.educe here the surprisingly large system network and realization rules for this area of the grammar, which merit a paper to themselves. All that needs to be said is that to develop a model of clause co-ordination that incorporates most of the phenomena of naturally occurring texts is a major task and that it took several mouths of work to buikl our current system. In terms of the above example, it generates, if \[spoken\] has been selected: J either Ivy loves Ike/NT/2 \[ or she loves Fred/NT/2 \[ or she doesn't love anybody/NT/1 \] While the patterns of the networks and their realizations are different for the co-ordination of nominal grOUpS, they are handled in a s;milar way. The system accommodates the perhaps surprising fact that, in the case of nominal groups, there is typically one more intonation unit than there would be commas. As in the MOOD network, there is a greater number of delicate choices re~_li~ed in intonation than there is in punctuation, So the feature \[spoken\] is again used as an entry condition to the system in the 'CO ORDINATION SIT'network, toensure that the ~stera is not unread unless \[spoken\] has been chosen. Here the speaker chooses in the system of \[unmarked co ordination spoken\] vs.</Paragraph>
    <Paragraph position="6"> Ice ordination with reservationS.. The first is realised by a Irene 2 base-to-mid rme; Halliday's Tone 3), and the second by a Tone 12 (fall-rise; HaUiday's Toue 4), 3.5.3. Thematized Circumstances: Situations, Things, Qualities Another major source of additional intonation units is the thematizatlon of time and drcumstanee. These meanings are realized in Adjuncts of various types. They may occur in various places ha the clause, and here we shall consider just those that appear at the beginning of a clause. So far GENESYS includes eleven types, each of which may be realized by either a clause or a group (three different classes of groups being reco~ized: nominal, prepositional and quantityquall~ groups). Note, theah that we have now identi~d a second major source of what has been termed 'clause C/omb;ningL A similar approach is needed for 'dame final' dames, i.e. clauses that fill any of the ecleven types of Adjunct built into GENESYS so far, and that come late in the clause. (This is a different approach to clause-combining from that in Halliday 198.5 and so from that in the Nigel grammar at ISI; here such clauses are simply treated as embedded - so far with gains in generalizations rather than losses.) Let us take as an e~mple the concept of time position, which is one of five types of 'circumstance of time' reco~niTed in GENF.SYS the others being repetition, duration, periodic frequency, and usuality. While GENESYS will happily generate dames such as &amp;quot;until he leaves the company&amp;quot; to specify a time position, iu the case of our example Ivy has chosen the simpler structure of the prepositional group, i.e. &amp;quot;until next month&amp;quot;. The first system to consider is:</Paragraph>
  </Section>
  <Section position="5" start_page="168" end_page="169" type="metho">
    <SectionTitle>
'TIME POSITION TI-IEMATIZATION&amp;quot;-&gt;
</SectionTitle>
    <Paragraph position="0"> 99%-~thematize-d timeposition (20.2) / i% thematizedJim'~ Ixffifion (20.3).</Paragraph>
    <Paragraph position="1"> Because the answer modifies the presuppositions that the Personnel Officer brought to his question (i.e. that Peter Piper had a fixed address), Ivy decides to thematize the part of her reply that expresses this, i,e. her specification of the 'time position'. This is realized by placing the 'time position Adjunct' at an early place in the clause.</Paragraph>
    <Paragraph position="2"> (Note that this is not a 'movement rule'; there are no such rules in this generator, and no element is located until it can be located in its correct place.) The next two systems are:  thematized time_position-&gt; 80% tim~..pos~io a as separate information uniT/20% time_.lXm&amp;quot;/ion as_part of main information unit. -- - - spoken &amp; time_position as separate information unit-&gt; - 20% high!i~ted thcmatized time_position /  S0% ~utnil3heTaatizccLti~ ix,/ition.</Paragraph>
    <Paragraph position="3"> The first of the two systems appfies whether or not the MODE is spoken or written (the written realization being a comma). But the writing system cannot make the distinction offered in the second, so that here again the feature \[spoken\] from the erR|hal MODE system is used as an entry condition. In our example Ivy chooses to resent the specification of the time position &amp;quot;nntil last week&amp;quot;) as a separate information umt, and furthermore to klshlight it (by using a Tone 12 (a fall-rise).</Paragraph>
    <Paragraph position="4"> But you may have noticed that these features have no realization rules. How, then, do these choices get realized? The answer is that these features act as conditional features on the realization rules for the units that are generated, after re-entry to the overall network, on a subsequent pass though it. The reason for including the system at the rsnk of situation is that in this way we can capture the genexalisa6on that these options are relevant whatever the unit - a clause or some kind of group - that fills the Adjunct.</Paragraph>
    <Paragraph position="5"> In our case the sub-network that we fred ourselves in on re-entry is the network for</Paragraph>
  </Section>
  <Section position="6" start_page="169" end_page="170" type="metho">
    <SectionTitle>
'MINIMAL RELATIONSHIP PLUS THING',
</SectionTitle>
    <Paragraph position="0"> i.e. the netv~k from which prepositional groups are generated. Here we enter the following system (where the suffix 'mrpt' echoes the name of the overall system):</Paragraph>
    <Paragraph position="2"> Here \[time..mrpt\] will be In'e-selected by the choice at the higher rank. The part of its realization rule concerned with intonation may appear, once again, somewhat complex, but once again it seems to correspond to the relative (but always limited) complexity of the facts of how English works: 90.002 : time_mrpt : if (on_previous_pass time_position as separatC/ information unit and on first_l~ass'spoken )thenill current unit pgp then e &lt; &amp;quot;\['9, if on prc'~ous pass highlighted_thematized time position then '12' by 'T ~,</Paragraph>
    <Paragraph position="4"> neutral thematized time_position then '2' by 'T.</Paragraph>
    <Paragraph position="5"> As you will see, these rules insert appropriate intonation boundaries and tones. The tonic ('T') is already waiting in the starting structure of the tnepositional group, so that the rule ~;mply conflates the actual tone with it. Let us assume that Ivy, in order to highlight still further the thematization of the words &amp;quot; month', selects the highfighting rather than the neutral option. (The nominal group &amp;quot;next month&amp;quot; is generated by a further re-entry.) Finally, the system supplies the initial intonation unit boundary for any unit without one. If we assume that the rest of items generated (in components not considered in this paper) are &amp;quot;he will be living at eleven Romilly Crescent, Canton&amp;quot; the full output for our example is: \] until next month/T/12 \[ he will be living at  eleven Romilly Crescent/T/2 I Canton/NT/1 i 3.5.4. Other Sources of Intonation  Other sources of intonation occur in specialist mini-grammars such as those for dates and addresses. These can be quite complex, and may insert several tonics, each with an appropriate tone. Our worked example illustrates one such case: note the Tone 2 on &amp;quot;Crescent&amp;quot;. Yet other types will be included in the next version of GENESY$, including (1) Adjuncts (which may be filled by clauses or groups) that are placed after the nuclear tonic of a clause and which carry 'supplementary information', and (2) 'nonrestrictive relative clauses' (i.e. ones that carry, once again, 'supplementary information'),</Paragraph>
    <Section position="1" start_page="169" end_page="170" type="sub_section">
      <SectionTitle>
3.6. Summary of the lexicogrammatical
</SectionTitle>
      <Paragraph position="0"> generation of intonation We have now completed a fairly fuU specification of the major aspects of intonation included at the present stage of the development of the GENESYS model.</Paragraph>
      <Paragraph position="1"> To summarize: GENESYS offers the choice, on entering the first system that results in the generation of a scntenco, between \[written\] and \[spoken\]. The importance of this apparently trivial system is that the choice made in it determines whether or not one can go on to enter quite a number of more 'delicate' systems whose choices are realized in intonation. Its features also act as conditions on the realization of features chosen in the same network, or in one entered on a subsequent pass. The result is that the realization at the level of form will be in  terms of either intonation or punctuation. We have seen how choices in MOOD, in</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="170" end_page="170" type="metho">
    <SectionTitle>
INFORMATION FOCUS and in various types of
</SectionTitle>
    <Paragraph position="0"> 'status of informaBon' contribute together to the specification of intonation, and we have seen some of the details of how this can be implemented.</Paragraph>
    <Paragraph position="1"> The result is an integrated model that avoids the psychologically implausible approach whereby one first generates a syntax tree and a string of words at its leave% and then 'adds on' the intonation.</Paragraph>
    <Paragraph position="2"> Instead, it treats intonation as one of three modes of re8i|Tation (the other two being syntax and items),, generating the various aspects of iatonatmn at appropriate points in the generation of syntax and items.</Paragraph>
    <Paragraph position="3"> R may be helpful to conclude by specifying explicitly the final stages of this process. First the generator looks for a eontrastive tonic ('CT') with which to confiate the tone., and then, if there isn't one, it provides as a default a nuclear tonic ('NT') for the flail matrix clause, i.e.. the intonational element of structure with which the tone reollzln~ the me a,;,g of MOOD is conflated. The other intonation units specified by various types of Information status are fitted around this central framework, receiving tones appropriate to their status. Where they are clauses these tones will be eonflated with a nuclear tonic (unless, of course, there is a contrastive tonic), and where they are groups the tones will be conflated with a simple tonic A nuclear tonic is thus one that is potentially capable of r0ceiviag the type of tone that re~!i~,C/s a MOOD option. It should be made clear that, in every case of the location of a nuclear tonic or a simple tonic, the element with which it is conflated must be one that is not exponndcd by an item from the list of inherently weak items. (Any such item may of course still receive a tonic by being contrastively stressed, as in l he has/C'r/t+ eaten it I.)</Paragraph>
  </Section>
class="xml-element"></Paper>