<?xml version="1.0" standalone="yes"?>
<Paper uid="P83-1016">
  <Title>A Framework for Processing Partially Free Word Order*</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
SRI International
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> The partially free word order in German belongs to the class of phenomena in natttral language that require a close interaction between syntax and pragmatics. Several competing principles, which are based on syntactic and on discourse information, determine the \[ineac order of noun phrases. A solution to problems of this sort is a prerequisite for high-quality language generation. The linguistic framework of Generalized Phrase Structure Grammar offers tools for dealing with-word order variation. Some slight modifications to the framework allow for an analysis of the German data that incorporates just the right, degree of interaction between syntactic and pragmatic components and that can account for conflicting ordering statements. null I. Introduction The relatively free order of major phrasal constituents in German belongs to the class of natural-language phenomena that require a closer interaction of syntax and pragmatics than is usually accounted for in formal linguistic frameworks.</Paragraph>
    <Paragraph position="1"> Computational linguists who pay attention to both syntax and pragmatics will find that analyses of such phenomena can provide valuable data for the design of systems that integrate these linguist ic components.</Paragraph>
    <Paragraph position="2"> German represents a good test case because the role of pragmatics in governing word order is much greater than in English while the role syntax plays is greater than in some of the so-called free-word-order languages like Warlpiri. The German data are well attested and thoroughly discussed in the descriptive literature The fact that English and German are closely related makes it easier to assess these data and to draw parallels.</Paragraph>
    <Paragraph position="3"> The .~imple analysis presented here for dealing with free word order in German syntax is based on the linguistic framework of Generalized Phrase Structure Grammar (GPSG}, especially on its Immediate Dominance/Linear Precedence formalism {ID/LP), and complements an earlier treatment of German word order) The framework is slightly modified to accommodate the relevant class of word order regularities.</Paragraph>
    <Paragraph position="4"> The syntactic framework presented in this paper is not hound to any particular theory of discourse processing; it enables syntax to interact with whatever formal model of pragmatics one might want to implement. A brief discussion of the framework's implication~ for computational implementation centers Upon the problem of the status of metagrammatical devices.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. The Problem
</SectionTitle>
    <Paragraph position="0"> German word order is essentially fixed: however, there is some freedom in the ordering of major phrasal categories like NPs and adverbial phrases - for example, in the linear order of subject (SUB J), direct object (DOBJ), and indirect object (lOB J) with respect to one another. All six permutations of these three constituents are possible for sentences like (In). Two are given as {Ib) and (It).</Paragraph>
    <Paragraph position="1"> (la) Dann hatte der Doktor dem Mann die Pille gegeben.</Paragraph>
    <Paragraph position="2"> Then had the doctor the man the pill given (lb) Dann hatte dec Doktor die Pille dem Mann gegeben.</Paragraph>
    <Paragraph position="3"> Then had the doctor the pill the man given (It) Dann hatte die Pille der Doktor dem Mann gegeben.</Paragraph>
    <Paragraph position="4"> Then had the pill the doctor the man given All permutations have the same truth conditional meaning, which can be paraphrased in English as: Then the doctor gave the man the pill.</Paragraph>
    <Paragraph position="5"> There are several basic principles that influence the ordering of the three major NPs:  ,,r the :tutbor and should not be interpreted as representative of the views of the Nati.,nal Science Foundation or the United States government. I have benefited fr,~rn discussions with and comments from Barbara Grosz, Fernand,, Pcreira. Jane Robinson. and Stuart Shieber.</Paragraph>
    <Paragraph position="6"> tThe best overview of the current GPSG framework can be found in Gazdar and Pullum (1982). For :t description of the II)/LP format refer to Gazdar and Pullum (Ig8l} and Klein (1983), for the ID/LP treatment of German t,, tszkoreit (\]g82a. lgB2b} and Nerbonne (Ig82).</Paragraph>
    <Paragraph position="7">  The order in (la) is based on the unmarked order, (lb) would be appropriate in a discourse situation that makes the man the focus of the sentence, and (1c) is an acceptable sentence if both doctor and man are focussed upon. l use focus here in the sense of comment, the part of the sentence that contains new important information. (lc) could be uttered as an answer to someone who inquires about both the giver and recipient of the pill (for example, with the question: Who gave whom the pill?l. The most complete description of the ordering principles, especially of the conflict between the unmarked order and the topic-commeni, relation, can be found in Lenerz (1977).</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. Implications for Processing Models
</SectionTitle>
    <Paragraph position="0"> Syntactic as well as pragmatic information is needed to determine the right word order; the unmarked-order principle is obviously a syntactic statement, whereas the topic-comment order principle requires access to discourse information. deg, Sometimes different ordering principles make contradictory predictions. Example (lb) violates the unmarked-order principle; (In) is acceptable even if dem Mann \[the man\] is the focus of the sentence~ 3 The interaction of ordering variability and pragmatics can be found in many languages and not only in so-called free-word-order languages. Consider the following two English sentences: (2a) I will talk to him after lunch about the offer.</Paragraph>
    <Paragraph position="1"> (2b) I will talk to him about the offer after lunch.</Paragraph>
    <Paragraph position="2"> Most semantic frameworks would assign the same truth-conditional meaning to (2a) and (2b), but there are discourse situations in which one is more appropriate than the other. (2a) can answer a que~-tion about the topic of a planned afternoon meeting, but is much less likely to occur after an order to mention the offer as soon as possible. 4 Formal linguistic theories have traditionally assumed the existence of rather independent components for syntax, semantics, and pragmatics, s Linguistics not only could afford this idealization but has probably temporarily benefited from it.</Paragraph>
    <Paragraph position="3"> However, if the idealization is carried over to the computational implementation of a framework, it can have adverse effects on the efficiency of the resulting system.</Paragraph>
    <Paragraph position="4">  ti, .n~.d di.&lt;o-,urse situations under any circumstances. In English. marked inton arian can usually overwrite default discourse role assignments associated w~.. the order of the constituents.</Paragraph>
    <Paragraph position="5"> $Scvera\[ more recent theories can account for the interaction among some of the components. Montague Grammar (Montague. 1974) and its successors (incl. GPSG) link semantic and syntactic rules. Work on presuppositions (Karttunen and Peters. 1979), discourse representations (Kamp, If80) and Situati~,n Semantics (Barwise and Perry. 1981) narrows the gap between .,,.'mantics and pragmatics.</Paragraph>
    <Paragraph position="6"> If we as.~ume that a language generation system should be able to generate all grammatical word orders and if we further assume that, every generated order should be appropriate to the given discourse situation, then a truly nonintegrated system, i.e., a system whose semantic, syntactic, and pragmatic components apply in sequence, has to be inel~cient. The syntax will first generate all possibilities, after which the pragmatic component will have to select the appropriate variant. To do so, this component will also need access to syntactic information.</Paragraph>
    <Paragraph position="7"> In an integrated model, much unnecessary work can be saved if the syntax refrains from using rules that introduce pragmatically inappropriate orders. A truly integrated model can discard improper parses very early during parsing, thereby considerably reducing the amount of syntactic processing.</Paragraph>
    <Paragraph position="8"> The question of integrating grammatical components is a linguistic problem. Any reasonable solution for an integration of syntax and pragmatics has to depend on linguistic findings about the interaction of syntactic and pragmatic phenomena. An integrated implementation of any theory that does not account for this interaction will either augment the theory or neglect the linguistic facts.</Paragraph>
    <Paragraph position="9"> By supporting integrated implementations, the framework and analysis to be proposed below fulfill an important condition for effcient treatment of partially free word order.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="109" type="metho">
    <SectionTitle>
4. The Framework and Syntactic Analysis
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="107" type="sub_section">
      <SectionTitle>
4.1 Tile Framework of CPSG in ID/LP Format
</SectionTitle>
      <Paragraph position="0"> The theory of GPSG is based on the assumption that nat ural languages can be generated by context-free phrase structure (CF-PS) grammars. As we know, such a grammar is bound to exhibit a high degree of redundancy and, consequently, is not the right formalism for encoding many of the linguistic generalizations a framework for natural language is expected to express. However. the presumption is that it is possible to give a condensed inductive definition of the CF-PS grammar, which contains various components for encoding the linguistic regt,laritics and which can be interpreted as a metagrammar, i.e.. a grammar for generating the actual CF-PS grammar.</Paragraph>
      <Paragraph position="1"> A GPSG can be defined as a two-leveJ grammar containing a metagrammar and an object grammar. The object grammar combines {CF-PS} syntax and model-theoretic semantics. Its rules are ordered triples (n. r. t) where n is an integer (the rule number}, r is a CF-PS rule. and t is the tramlationoft.he rule, its denotation represented in some version of intensional logic. The translation t is actually an operation that maps the translation of the children nodes into the translation of t.he parent. The nonterminals of r are complex symbols, subsets of a finite set of syntactic features or - as in the latest version of the theory (Gazd:w and Pullum, 1982) - feature trees of finite size. The rules o/' the obJect grammar are interpreted as tree-admissability conditions.</Paragraph>
      <Paragraph position="2"> The metagrammar consists of four different kinds of rules that are used by three major components to generate the object  grammar in a stepwise fashion. Figure {3) illustrates the basic structure of a GPSG metagrammar.</Paragraph>
      <Paragraph position="3">  First. there is a set of banjo rules. Basic rules are immediate domi.a.ce rule (IDR) double~, ordered pairs &lt; n,i &gt;, where n is the rule number and i is an \[DR.</Paragraph>
      <Paragraph position="4"> 1DRs closely resemble CF-PS rules, but, whereas the CF-PS rule &amp;quot;1 -- 6t 6..... 6. contains information about both immediate dominance and linear precedence in the subtree to be accepted, the corresponding IDR &amp;quot;~ -- 6t, /f~. ..... /f. encodes only information about immediate dominance. The order of the right-hand-side symbols, which are separated in IDRs by commas, has no significance.</Paragraph>
      <Paragraph position="5"> Metarule Application, maps \[DR doubles to other IDR doubles. For this purpose, metaxules, which are the second kind of rules are applied to basic rules and then to the output of metarule applications to generate more IDR doubles. Metarules are relations between sets of IDRs and are written as A = B, where A and B are rule templates. The metarute can be read as: If there is an IDR double of kind A, then there is also an IDR double of kind /3. In each case the rule number is copied from A to /3. s .Several metarules can apply in the derivation of a single II)R double; however, the principle of Finite Closure, defined by Thompson (1982}, allows every metarule to apply only once in the derivational history of each IDR double. The invocation of this principle avoids the derivation of infinite rule sets, in- null the s~me integer to the whole class of rules that were derived from the ~ame basic rules. This rule number propagation is a prerequisite for the &lt;iPSG accouht of subcategori2ation.</Paragraph>
      <Paragraph position="6"> eluding those that generate non-CF, non-CS, and noarecursive</Paragraph>
      <Paragraph position="8"> Another component maps IDR doubles to IDR triples, which are ordered triples (n,i,t) of a rule number., an IDR i, and a translation t. The symbols of the resulting IDRs axe fully instantiated feature sets (or structures} and therefore identical to object grammar symbols. Thus, this component adds semantic translations and instantiates syntactic features. The mapping is controlled by a third set of rule czten6io, principles including feature co-occurrence restrictions, feature def. ult principles, and an algorithm that assigns the right kind of translation to each rule on the basis of its syntactic information.</Paragraph>
      <Paragraph position="9"> The last component of the metagrammar maps the IDR triples to the rules of the object grammar. For each IDR triple all the object grammar triples are generated whose CF-PS rules conform with the linear precedence(LP) rules, the fourth rule set of the metagrammar. LP rules are members of the LP relation, a partial ordering on V'r I.I VN. An LP rule (a,$} is usually written as a &lt; ~/and simply states that a precedes/9 whenever both a and d occur in the right-hand-side of the same CF-PS rule.</Paragraph>
      <Paragraph position="10"> It is the separation of linear precedence from immediate dominance statements in the metagrammar that is referred to .as ID/LP format. And it is precisely this aspect of the formalism that. makes the theory attractive for application to languages with a high degree of word-urder freedom. The analysis presented in the next section demonstrates the functioning of the formalism and some of its virtues.</Paragraph>
    </Section>
    <Section position="2" start_page="107" end_page="107" type="sub_section">
      <SectionTitle>
4.2 The Analysis of German Word Order
</SectionTitle>
      <Paragraph position="0"> Uszkoreit (1982a) proposes a GPSG analysis of German word order that accounts for the fixed-order phenomena, including the notoriously difqcult problem of the position of finite and nonfinite verbs. Within the scope of this paper it is impossible to repeat, the whole set of suggested rules. A tiny fragment should sumce to demonstrate the basic ideas as well as the need for modifications of the framework.</Paragraph>
      <Paragraph position="1"> Rule (41 is the basic VP ID rule that combines ditransitive verbs like forms of gebe. (give) with its two objects:</Paragraph>
      <Paragraph position="3"> Th,~ rule .~tates that a VP can expand as a dative NP (IOBJ}, an attn.-alive NP (DOBJ), and a verb. Verbs that can occur in dilrnnsitive VPs, like geben (give). are marked in the lexicon with the rule number 5. Nothing has been said about the linear order of these constituents. The following metarule supplies a &amp;quot;flat&amp;quot; sentence rule for each main verb VP rule \[+NOM 1 stands for the nominative case, which marks the subject.</Paragraph>
      <Paragraph position="4">  by a PS rule derived from ID rule (6): (7} der Doktor dem Mann die Pille gegeben the doctor the man the pill given I shall not list the rules here that combine the auxiliary halle and the temporal adverb dann with (7) to arrive at sentence (la), since these rules play no role in the ordering of the three noun phrases. What is of interest here is the mapping from ID rule (5) to t.he appropriate set of PS rules. Which LP rules are needed to allow for all and only the acceptable linearizations? The position of the verb is a relatively easy matter: if it is the finite matrix verb it precedes the noun phrases; in all other cases, it follows everything else. We have a feature MC for matrix clause as well as a feature co-occurrence restriction to ensure that +MC will always imply +FIN (finite). Two LP rules are needed for the main verb:  (Kart.tunen and Peters, 1979) 8 or a function from discourse situations to the appropriate truth-conditional meaning in the spirit of Barwise and Perry (1981). The analysis here is not concerned with choosing a formalism for an extended semantic component, but rather with demonstrating where the syntax has to provide for those elements of discourse information that influence the syntactic structure directly.</Paragraph>
      <Paragraph position="5"> Note, that the new LP rules do not resolve the problem of ordering-principle conflicts, for the violation of one LP rule is enough to rule out an ordering. On the other hand, the absence of these LP rules would incorrectly predict that all permutations are acceptable. The next section introduces a redefinition of LP rules that provides a remedy for this deficiency.</Paragraph>
    </Section>
    <Section position="3" start_page="107" end_page="109" type="sub_section">
      <SectionTitle>
4.3 The Modified Framework
</SectionTitle>
      <Paragraph position="0"> Before introducing a new definition of LP rules, let me suggest, anot.her modification that will simplify things somewhat.</Paragraph>
      <Paragraph position="1"> The I,P rules considered so far are not really LP rules in the sense in which they were defined by their originators. After all. LP rules are defined as members of a partial ordering on &amp;quot;v~,C/ U VT'. Our rules are schemata for LP rules at best, abbreviating the huge set of UP rules that are instantiations of these schemata.</Paragraph>
      <Paragraph position="2"> This definition is an unfortunate one in several respects. It not. only creates an unnecessarily large set of rules IVN contains thousands of fully instantiated complex symbols) but also suppresses some of the important generalizations about the language. Clearly, one could extract the relevant generalizations even from a fully expanded LP relation, e.g., realize that there is no LP rule whose first element has -MC and its second element NP. However, it should not be necessary to extract generalizations from the grammar; the grammar should express these generalizat.ions directly. Another disadvantage follows from the choice of a procedure for arriving at the fully expanded LP rela-Lion. Should all extensions that are compatible instantiations of (Sa), (Sb). and (9a)-(9e} be LP rules: If so. then (10) is an</Paragraph>
      <Paragraph position="4"> ,.\ feature FOCUS has been added that designates a focused consf it,eat. Despite its name FOCUS is a syntactic'fcature, justified by syntactic Pacts, such as its influence on word order. This syntactic feature needs t,o be linked with the appropriate discourse information. The place to do this is in the rule exteusioq component, where features are instantiated and semantic translations added to ID rules. It is assumed that in so doing the translation part of rules will have to be extended anyway so as to incorporate non-truth-conditional aspects of the meaning.</Paragraph>
      <Paragraph position="5"> For example, the full translation could be an ordered pair of truth-conditional and non-truth-conditional content, extending Karttunen and Peters's treatment of conventional implicature Yet nothing can be a matrix verb and definite simultaneously, and NPs cannot be finite. (101 is a vacuous rule. Whether il is a LP rule at all will depend on the way the nonterminal vocabulary of the object grammar is defined. If it only includes the nonterminals that actually occur in rules then (10) is not as LP rule. \[n this case we would need a component of the metagrammar, the feature instantiation principles, to determine 8T,~ be more precise. Karttunen and Peters actuaJly make their translati,,ns ordered triples of truth-conditiona.l content, impllcatures, and an inhcrhance expression that plays a role in h~.ndling the projection problem for presuppositions.</Paragraph>
      <Paragraph position="6">  another compouent of the metagrammar, the LP component. 9 LP will be redefined as a partial order on 2 p, where F is the set of syntactic features I0 The second and more important change can best be described by viewing the LP component as a function from a pair of symbols (which can be characterized as feature sets) to truth values, telling us for every pair of symbols whether the first can precede the second in a linearized ru!e. Given the LP relation  {(al,~/t),(a~,B~.) ..... (a~,~)} and a pair of complex symbols (3',6), the function can be expressed as in (11).</Paragraph>
      <Paragraph position="7"> (11} cl A c,~ A ... A c,~ where c~ ---- ~(~; _C 6 A #; C: 3')  for 1 &lt; i &lt; n ~,Ve call the conjunct clauses LP conditions; the whole conjunction is a complex LP condition. The complex LP condition allows &amp;quot;T to precede /~ on the right-hand side of a CF-PS rule if every LP condition is true. An LP condition ct derived from the LP rule (a~,//i) is true if it is not the case that 3 has the features ;/~ and 6 has the features aC/. Thus the LP rule NP &lt; VP stanch for the following member of the LP relation {{+N,-V, +2B~R}, l-N, +V, +2BAR}).</Paragraph>
      <Paragraph position="8"> The LP condition following from this rule prevents a superset of {-N, +V, +2BAR} from preceding a superset of l-N, +V, +2BAR}, i.e., a VP from preceding an NP.</Paragraph>
      <Paragraph position="9"> But notice that there is nothing to prevent us from writing a fictitious LP rule such as (12} +PRONOUN &lt; -ACCUSATIVE German has verbs like Ichrcn that take two accusative noun phr~.ses as complements. If {12) were an LP rule then the resulting LP condition defined as in ( l 1 ) would rule out any occurrence of two prouominalized sister NPs because either order would be rejected.l 1 It. is an empirical question if one might ever find it useful to write LP rules as in (12}, i.e., rules a &lt; ~/, where a U 3 could be a ~ubset of a complex symbol. Let me introduce a minor redefinition of the interpretation of LP, which will take care of cases such as (12) and at the same prepare the way for a more substantial modification of LP rules. LP shall again be interpreted as a function from pairs of feature sets (associated with complex symbols} to truth values. Given the LP relation {(a1,,'Jl),(oo..;\]'.,} ..... (a.,~q~) and a pair of complex symbols 0The widety uscd notation for nomnstantiated LP rules and the feature instantiati,,n principles could be regarded an meta, met.Lgrammatical devices that inductively define a part of&amp;quot;the metagrammar.</Paragraph>
      <Paragraph position="10"> 10Remember that, in an .~-synta.x. syntactic categories abbreviate feature sets NP ~ {+N, -V, +2BAR}. The definition can emily be extended to work on feature trees instead of feature sets.</Paragraph>
      <Paragraph position="11"> 1 lln principle, there is nothing in the original ID/LP definition either that would prevent the grammar writer from abbreviating a set of LP rules by  (121. It is not quite clear, however, which set of LP rules is abbreviated by (r&amp;quot;).</Paragraph>
      <Paragraph position="12"> (3',/~), the function can be expressed as in (13).</Paragraph>
      <Paragraph position="13"> (13) ct A c2, A ... A cn where</Paragraph>
      <Paragraph position="15"> for l &lt; i &lt; n That means 3' can precede 6 if all LP conditions are true. For instance, the LP condition of LP rule (12) will yield false only if &amp;quot;t is +ACCUSATIVE and # is +PRONOUN, and either 3, is -PRONOUN or 6 is -ACCUSATIVE (or both).</Paragraph>
      <Paragraph position="16"> - Now let. us assume that, in addition to the kind of simple LP rules just introduced, we can also have complex LP rules consisting of several simple LP rules and notated in curled brackets a.s in (14}:  The LP condition associated with such a complex LP rule shall be the disjunction of the LP conditions assigned to its members. LP rules can be generally defined as sets of ordered pairs of feature sets {(at,Bt),(a~,~) ..... (am,~/m)}, which are either notated with curled brackets as in (10), or, in the case of singletons, as LP rules of the familiar kind. A complex LP rule {{at, dl), (no_, ,%) ..... {am, B,n)} is interpreted as a LP condition of the following form {(o 1 C 6 A~t C -~)V(a~ C 6 At/= C_ -,)v . vt~.,C6A~,,C_~))--((a, C_3,A3, c_ ~}v(a.. c_ &amp;quot;l A ,'t= C 6)V ... V(am C 3, A dm ~ 6)}. Any of the atomic LP rules within the complex LP rule can be violated as long as the violations are sanctioned by at least one of the atomic LP rules. Notice that with respect to this definition, &amp;quot;regular&amp;quot; LP rules, i.e., sing{elons, can be regarded as a speciaJ case of complex I,P rules.</Paragraph>
      <Paragraph position="17"> \[ want C/o suggest that the LP rules in {Sa}, (8h), and (l-I} arc a subset of the LP rules of German. This analysis makes a number of empirical predictions. For example, it predicts that  In (17) the sub-LP-rules +DAT &lt; +ACC and -FOCUS &lt; +FOCUS are violated. No other sub-LP-rule legitimizes these violations and therefore the sentence is bad.</Paragraph>
      <Paragraph position="18"> This agrees with the findings of Lenerz (1977), who tested a large number of sample sentences in order to determine the interaction of the unmarked syntactic order and the ordering preferences introduced by discourse roles. There are too many possible feature iustantiatious and permutations of the three noun phrases to permit making grammaticality predictions here for a larger sample of ordering variants. So far 1 have not discovered any empirical deficiencies in the proposed analysis.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="109" end_page="109" type="metho">
    <SectionTitle>
5. Implications for Implementations
</SectionTitle>
    <Paragraph position="0"> The theory of GPSG, a,s described by its creators and as outlined in this paper, cannot be used directly for implementation. The number of rules generated by the metagrammar is just too large. The Hewlett-Packard system (Gawron etal., 1982} as well as Henry Thompson's program, which are both based on a pre-ID/LP version of GPSG, use metarules as metagrammatical devices, but with feature iustantiation built into the processor.</Paragraph>
    <Paragraph position="1"> Agreement checks, however, which correspond to the work of the metagrammatical feature instantiation principles, are done at parse time. As Berwick and Weinberg (1982\] have pointed out, the cont ext-freeness of a grammar might not accomplish much when the number of rules explodes. The more components of the metagrammar that can be built into the processor (or used by it as additional rule sets at parse time), the smaller the resulting grammar will be. The task is to search for parsing algorithms that. incorporate the work of the metagrammar into context-free phrase structure parsing without completely losing the parsing time advantages of the latter. Most PSG parsers do feature handling at parse time. Recently, Shieber (forthcoming) has extended the Earley algorithm (Earley 1970) to incorporate the linearization process without a concomitant loss in parsing c~ciency. The redefinition of the LP component proposed in this paper can be intrusted easily and efficiently into Shieber's extension.</Paragraph>
    <Paragraph position="2"> If the parser uses the disjunctive LP rules to accept all ordering variants that are well-formed with respect to a discourse, there still remains the question of how the generator chooses among the disjuncts in the LP rule. It would be very surprising if the different orderings that can be obtained by choosing one LP rule disjua:t over another did in fact occur with equal frequency. Although there are no clear results that might provide an answer to this question, there are indications that certain disjuntas &amp;quot;win out&amp;quot; more often than others. However, this choice is purely stylistic. A system that is supposed to produce high-quality output might contain a stylistic selection mechanism that avoids repe, hions or choose~ among variants according to the tyt:e of text or dialogue.</Paragraph>
  </Section>
class="xml-element"></Paper>