<?xml version="1.0" standalone="yes"?>
<Paper uid="J93-3002">
  <Title>A Computational Theory of Goal-Directed Style in Syntax</Title>
  <Section position="3" start_page="454" end_page="456" type="metho">
    <SectionTitle>
6 Thus the speaker or writer is accountable for his or her stylistic decisions, in the sense that that term is
</SectionTitle>
    <Paragraph position="0"> used in Ethnomethodology and Conversation Analysis. A person encountering friends on the street, for example, may choose to greet them or not, but is held accountable either way--failure to greet is a snub; it is not possible to opt out of the situation altogether (Heritage 1984). Similarly, a speaker or writer is accountable for all the stylistic nuances of his or her utterances; it is not possible to utter a sentence in such a way that only the propositional content counts and not the form in which it is expressed. 7 &amp;quot;Economy is expressed at two levels, lexical and syntactic, which may, however, be related: what is expressed lexically in one language may be realized through syntactic means in the other, and vice versa&amp;quot; (Vinay and Darbelnet 1958, p. 185, authors&amp;quot; translation).</Paragraph>
    <Paragraph position="1">  Computational Linguistics Volume 19, Number 3 be as general as possible, ideally applicable to the codification of style for the sentence and paragraph levels; appropriate for lexical, syntactic, and semantic style; and applicable to both English and French. The methodology will be used in conjunction with the stylistic vocabulary to guide the systematic collection and creation of stylistic rules. This grammar will provide a formal description of the syntactic patterns that differentiate the various stylistic goals.</Paragraph>
    <Paragraph position="2"> In addition to theoretical problems, there are implementation issues to be addressed as well. In Section 5, we will describe the development of a stylistic parser that uses the formal bodies of rules to analyze complex English text.</Paragraph>
    <Paragraph position="3"> In the last section, we will review the contributions of the paper and discuss completed and ongoing extensions to the research.</Paragraph>
    <Paragraph position="4"> 2. Background to the Study of Style</Paragraph>
    <Section position="1" start_page="455" end_page="456" type="sub_section">
      <SectionTitle>
2.1 Views of Style
</SectionTitle>
      <Paragraph position="0"> first appeared in the fifth century B.C. Scholars of the time conceived the notion that the rhetorical form of a writer's text should reflect his thoughts and intentions. This was to be achieved through the appropriate choice and organization of words and syntactic structures in order to &amp;quot;evoke the desired response&amp;quot; (Corbett 1971). Content and form were recognized as inseparable, and as exerting a reciprocal influence upon each other.</Paragraph>
      <Paragraph position="1"> Beginning with Corax of Syracuse, rhetoricians developed methods for systematic instruction in the art of writing. The classical scholars, from Isocrates and Aristotle to Cicero and Quintilian, established standards of rhetoric that influenced curricula for centuries. The contribution of Aristotle was particularly notable. He countered Plato's argument that rhetoric was mere sophistry, deceitful reasoning, by providing a system of instruction that treated the theory and practice of style as a valid discipline, designed to appeal simultaneously to reason and emotion.</Paragraph>
      <Paragraph position="2"> This classical approach to style flourished into the Middle Ages, for it formed part of a standard university education. The teaching of formal rhetoric in schools and universities continued to play a significant role throughout the Renaissance and into the 18th century. However, in the 19th century, the teaching of rhetoric gave way to the teaching of composition, which came to be associated merely with a set of basic prescriptive rules. This approach was in direct contrast to the original classical theory that stressed the education of the whole person as a preparation for achieving style in writing (Corbett 1971).</Paragraph>
      <Paragraph position="3">  to standard good taste and an avoidance of bad form. Textbooks (Fowler 1968; Strunk and White 1979; Kane 1983; Grevisse 1986) promote the idea that there is a universal and correct mode of expression. However, normative rules, while necessary in some degree so that communication is possible at all, may, if too inflexible, deaden the expression of a writer's individual voice. Universal rules of good style are best used, we suggest, simply as a guide to avoiding bad constructions and obstacles to clear writing, but not as a basis for a theory of stylistic effects in text.</Paragraph>
      <Paragraph position="4"> 2.1.3 Our View. In computational applications, where we expect to deal with large amounts of similar types of text, the analysis of group style is of more interest than the idiosyncratic style of any one writer. Group style can be subdivided into two major types, each associated with a different view of stylistics: literary style and utilitarian  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax (or functional) style. Utilitarian group styles are distinguished from the literary type by their association with a genre of text that has a particular function or purpose, such as medical textbooks or political newspaper writing. In such styles, the writer adjusts her language to what readers expect in a specific situation. The fact that functional group styles are somewhat more restrictive than literary group styles, as they are characteristic of a particular genre, suggests that the problem of codifying utilitarian style will be more tractable than the literary case. For this reason, we will focus on utilitarian texts in general, drawing most of our examples from newspaper and magazine journalism.</Paragraph>
      <Paragraph position="5"> (Our main source of examples is the English translations, published in The Guardian Weekly, 8 of articles from Le Monde.)</Paragraph>
    </Section>
    <Section position="2" start_page="456" end_page="456" type="sub_section">
      <SectionTitle>
2.2 The Current State of Computational Stylistics
</SectionTitle>
      <Paragraph position="0"> Most of the research to date in computational stylistics has been the development of so-called style-checkers. The UNix-based Writer's Workbench, Grammatik I and II, RightWriter, PC-Style, Punctuation and Style, Electric Webster, and CRITIQUE are examples (Raskin 1986). However, none of these systems is appropriate for our purposes.</Paragraph>
      <Paragraph position="1"> These programs enforce the basic virtues: be clear, be simple, be precise. They merely check for common grammatical errors such as number disagreement, pronoun case problems, unbalanced punctuation, split infinitives, excessive sentence length, and excessive sentence complexity. There is no systematic approach to constructing a vocabulary of style, no structured representation of stylistic rules. Stylistic knowledge may simply consist, as in the case of CRITIQUE, of an unstructured mass of rules (Jensen et al. 1986, p. 190). These programs have no real understanding of the significance of stylistic parameters.</Paragraph>
      <Paragraph position="2"> The other main kind of computational research in stylistics has been statistical (e.g., Cluett 1976, 1990; Milic 1982; Biber 1988, 1989). With this approach, a count is made of the number of occurrences in a sample text of a stylistically significant feature, such as an initial prepositional phrase. A statistical analysis is done to compare the count and type of stylistic features with the corresponding parameters of texts written by authors working in the same or different genres. But the interpretation of the statistical results must be done by a human; the computer system has no real understanding of what the variations in style mean.</Paragraph>
      <Paragraph position="3"> Although virtually all computational research in stylistics takes a non-AI approach to the representation and use of stylistic knowledge, the usefulness and degree of sophistication of some of these programs should not be underestimated. The York Computer Inventory of Prose Style (Cluett 1976, 1990) has been undergoing development since 1970. The principal contribution of the project is its convincing demonstration that the richness and subtlety of literary style could in fact be correlated, to a perhaps surprising degree of accuracy, with a catalog of syntactic patterns. But the program that performs the stylistic analysis has no understanding of the meaning of its statistical results; it relies on human interpretation of the statistics computed. Nevertheless, the York project's findings are still very useful, as they have catalogued the syntactic features that produce particular stylistic effects.</Paragraph>
      <Paragraph position="4"> Thus, most current approaches in computational stylistics have not included formal stylistic rules, which are clearly a desirable prerequisite. We now review the extent to which such rules have been developed in theoretical stylistics and used in computational stylistics.</Paragraph>
      <Paragraph position="5"> The title of Vinay and Darbelnet's (1958) book, Stylistique comparde du franC/ais et de</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="456" end_page="458" type="metho">
    <SectionTitle>
8 Formerly The Manchester Guardian Weekly.
</SectionTitle>
    <Paragraph position="0"> Computational Linguistics Volume 19, Number 3 l'anglais, is somewhat misleading, for their purpose was not just to deal with comparative stylistics in translation, but also to enumerate the more common rules and procedures in the overall translation process. But they did not construct a specific vocabulary for comparative stylistics and, in fact, virtually all of their 'stylistic' terms refer to basic syntactic constituents. Using their vocabulary of terms, Vinay and Darbelnet defined rules, expressed in ordinary natural language, that encoded knowledge about French-English comparative stylistics. The rules certainly fell short of the repertoire needed by a professional translator, but no one has yet improved upon their codification of comparative stylistics (Brian Fitch, personal communication). Their book is still used as a prescribed text in translation courses. Vinay and Darbelnet did not take an explicitly goal-directed approach, but their book does contain a great deal of information about the correlation of stylistic goals with particular lexical, syntactic, and semantic choices. Unfortunately, however, their rules proved to be too language-dependent for the framework that is proposed in this paper. But the success of a goal-directed codification, albeit an implicit one, supports the approach that we will take below.</Paragraph>
    <Paragraph position="1"> Crystal and Davy (1969) hypothesized the creation of a formal stylistic grammar, built upon a vocabulary of stylistic terms and intended to represent the sentence structures characteristic of a particular stylistic sublanguage. However, the grammar that they sketched contained virtually no terms other than purely syntactic ones and no correlations with stylistic goals; it was only a first step, although an ambitious one, toward a codification of the stylistic rules that define a sublanguage. What we have found especially useful to adapt from their work is the overall methodology, the approach to formalizing style, even the belief that style can be formalized.</Paragraph>
    <Paragraph position="2"> Kane's (1983) rhetoric and handbook of style represents the typical textbook that aims to teach the rules of classic good writing. Kane, too, did not take an explicitly goal-directed approach in his stylistic rules, but nevertheless gave numerous references to the stylistic goals associated with particular syntactic structures. The systematic construction of a stylistic vocabulary seems not to have been an explicit goal, but he used many terms unique to stylistics. However, his terminology was apparently augmented whenever the need arose to describe a new feature; there was no underlying structure, no formal representation of rules. The stylistic rules were expressed in ordinary English and achieved, as the author intended, good coverage of plain writing style.</Paragraph>
    <Paragraph position="3"> This body of rules was 'structured' to the extent that there were hierarchies of stylistic sentence types, but, in general, Kane's rules appear to be just an enumeration of stylistic tenets, not a deliberately constructed organization. Nevertheless, we have found it useful to draw upon these rules as justification for the syntactic coverage of our grammar.</Paragraph>
    <Paragraph position="4"> In computational stylistics, an application area of particular interest to us is machine translation (MT). In suggesting the applicability of computational stylistics to MT, Loffier-Laurian (1986) emphasized an important point: although corresponding group styles may exist across languages, the realization of the style can be different for each language. Tsutsumi (1990) presented a methodology for bridging stylistic gaps (stylistic differences) between the syntax of the source and target languages in MT. An important contribution of Tsutsumi's work is the recognition that computational stylistics is useful for the translation of pairs of languages that are not in the same language group.</Paragraph>
    <Paragraph position="5"> Hovy's (1988) PAULINE system was the first computational system that implemented goal-directed style. PAULINE was able to generate text that conformed to various stylistic and pragmatic constraints that it was given. The system was goal-directed, able to correlate such stylistic goals as formality, simplicity, and respect with the lexical  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax and syntactic characteristics of the text produced. Hovy's method was based on the definition of rhetorical goals of style, such as formality, force, and partiality, that control broader pragmatic goals. For example, a low degree of formality combined with high degrees of force and partiality gives a 'no-nonsense' effect.</Paragraph>
    <Paragraph position="6"> To achieve a particular rhetorical goal, PAULINE examined the options at various points during the production of text and applied the appropriate strategies. For example, to achieve formality, the generator would perform the following actions: Organization of topics: Make long complex sentences by subordinating them in relative clauses or by conjoining two or more sentence topics.</Paragraph>
    <Paragraph position="7"> Organization of sentences: Make sentences seem weighty by including a number of adverbial clauses, by placing these clauses toward the beginning of a sentence, by building parallel clauses, by using the passive voice, by using more 'complex' tenses such as the perfect tenses, and by avoiding ellipsis.</Paragraph>
    <Paragraph position="8"> Choice of words and phrases: Select formal words and phrases. Avoid doubtful grammar, popular idioms, slang, and contractions.</Paragraph>
    <Paragraph position="9"> Although structured in its correlation of the generator's actions with decision points, this approach to representing a goal-directed knowledge of style is essentially heuristic. Hovy's success in implementing a goal-based notion of style, even though limited in scope and informal in the mode of knowledge representation, encouraged our development of a computational approach to goal-directed style.</Paragraph>
    <Section position="1" start_page="458" end_page="458" type="sub_section">
      <SectionTitle>
2.3 Summary
</SectionTitle>
      <Paragraph position="0"> * From existing research in theoretical stylistics, we have seen that there is a basis for the codification of group-based, utilitarian, goal-directed stylistics.</Paragraph>
      <Paragraph position="1"> * The need for a vocabulary of stylistic terms is an accepted idea. However, the nature and structure of such a vocabulary for computational use, which must describe more than basic syntax and be amenable to systematic construction, have not yet been addressed. These issues will be subjects of the next section.</Paragraph>
      <Paragraph position="2"> * Similarly, while large bodies of stylistic rules already exist, they have not been organized into the formal structure necessary to a computational approach. Section 4 will develop a structured method for representing stylistic rules by constructing formal grammars.</Paragraph>
      <Paragraph position="3"> * The feasibility of a goal-directed analysis of style is supported by previous research, but an actual goal-based codification has not previously been attempted. In the grammar that we will develop in Section 4, we will incorporate such a goal-directed knowledge of stylistics.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="458" end_page="473" type="metho">
    <SectionTitle>
3. A Vocabulary of Style
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="458" end_page="459" type="sub_section">
      <SectionTitle>
3.1 Stylistic Goals
</SectionTitle>
      <Paragraph position="0"> If we hope to build AI-based systems to deal with matters of style, then we must provide a formal knowledge representation: we need a vocabulary of well-defined, expressive terms that will allow our intuitions about style to be stated precisely and  Computational Linguistics Volume 19, Number 3 understood clearly. We will construct a vocabulary from terms that are associated with utilitarian, group-based, goal-directed stylistics. We will start by considering the kinds of stylistic goals to be dealt with in the lexicon.</Paragraph>
      <Paragraph position="1"> When a writer composes text, she has certain stylistic goals in mind; such goals might include clarity, or informality, or even the clouding of an unpalatable message. To achieve her various goals, she will choose specific words (outplacement or firing, for example), syntactic constructions (heavily modified or sparse noun phrases), and semantic organization (Thirty percent of the class failed or The majority of students passed). In applications such as machine translation, we want to understand why a writer has used language in a particular way, what specific effects she intended to convey, and which linguistic choices were made to achieve these goals. There are a multitude of goals that could be considered; indeed, the space of possibilities has never been fully explored. For this study, we have chosen six goals that Vinay and Darbelnet  (1958) consider to be among the most commonly used, representing the opposite ends of three dimensions. The goals are listed as follows, with an intuitive description of the corresponding characteristics that a sentence of each type would display: * Clarity: Simplicity, harmony, and no ambiguity.</Paragraph>
      <Paragraph position="2"> Obscurity: Complexity, incongruity, and difficulty of understanding.</Paragraph>
      <Paragraph position="3"> * Concreteness: Specificity, with an emphasis on certain components.</Paragraph>
      <Paragraph position="4"> Abstraction: Generality, with no particular emphasis on any one sentence component.</Paragraph>
      <Paragraph position="5"> * Staticness: Uniformity, predictability, continuity.</Paragraph>
      <Paragraph position="6"> Dynamism: Deviation from the norm, unexpectedness, action.</Paragraph>
    </Section>
    <Section position="2" start_page="459" end_page="461" type="sub_section">
      <SectionTitle>
3.2 From Theoretical Stylistics to Computational Stylistics
</SectionTitle>
      <Paragraph position="0"> We will now develop a vocabulary of basic terms that will describe stylistically significant aspects of syntax. Our development is guided by much in classical rhetorical theory; often, our definitions will be extensions of those of the classical theory. We do this in two stages: first, we outline abstract properties of style suggested by classical rhetoric; second, we convert these properties into definitions of abstract elements that will be used in a grammar of style.</Paragraph>
      <Paragraph position="1"> Our starting point is the notion of stylistic norm, which is defined as the most commonly used structures in a given genre. This notion is intrinsic to the work of the majority of descriptive stylists such as Crystal and Davy (1969) and Cluett (1976, 1990).</Paragraph>
      <Paragraph position="2"> From this, we derive the notions of stylistic concord and discord as the fundamental principles of our formalization. Concord simply means conformity with the norm, and discord, deviation from the norm.</Paragraph>
      <Paragraph position="3"> Each genre has its own particular norm; for example, the style characteristic of scientific texts may seem disruptive if used in general newspaper writing. This view goes beyond the tenet of prescriptive rhetoric that the stylistic norm merely represents standard, 'good,' style, independent of the use of the text. Rather, we agree with Lanham (1974) that style should be viewed as the tailoring of a text to a specific audience and a specific situation; 'good' style can in fact be quite inappropriate.</Paragraph>
      <Paragraph position="4"> The concept of norm will be useful as a means of anchoring the development of our stylistic rules. It is simply the usual, the typical, the stable and concordant. Indeed, discord, deviation from the norm, will play an essential role in our formalization of style, for it is our contention that style arises from the construction of patterns of  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax concord and discord, particular combinations of order and disorder that create an overall harmonious arrangement (Crystal and Davy 1969; Cluett 1976, 1990).</Paragraph>
      <Paragraph position="5">  of style for computational use, we will look at the descriptive terms that stylists have developed over many centuries of use.</Paragraph>
      <Paragraph position="6"> One of the most important rhetorical influences on sentence style that has been identified is syntactic parallelism, or coordination: the balancing of syntactically similar forms to achieve a harmonious stylistic effect, an idea dating back to the rhetorical schemes of balance (Corbett 1971, p. 463).</Paragraph>
      <Paragraph position="7"> But if parallelism or coordination plays an important role in style, so too does subordination: variations in the hierarchical structure of a sentence that can produce correspondingly different stylistic effects. One part of a sentence can dominate another. As with parallelism, the importance to style of syntactic hierarchy is also a well-established principle in classical rhetoric, evident in the schemes of parenthesis, apposition, and climax (Corbett 1971, pp. 466, 468, 476). Thus, modulations in coordinate and subordinate structure are major factors in achieving stylistic expressiveness.</Paragraph>
      <Paragraph position="8"> A third major contributor to syntactic style is linear ordering: varying the order of components within a sentence to produce quite marked stylistic differences. The rhetorical terms of anastrophe and parenthesis (Corbett 1971, p. 466) attest to the stylistic importance of the ordering of sentence components.</Paragraph>
      <Paragraph position="9"> Thus, we can see that three factors influencing syntactic style are:  * Balance: Parallelism of structure.</Paragraph>
      <Paragraph position="10"> * Dominance: Structural hierarchy.</Paragraph>
      <Paragraph position="11"> * Position: Linear ordering of structure.</Paragraph>
      <Paragraph position="12">  We will adapt these factors from theoretical stylistics as the divisions of our vocabulary of style. Now, we will classify commonly used stylistic terms into these three groups in order to identify and abstract the general properties of style characteristic of each group.</Paragraph>
      <Paragraph position="13">  construction (Hendricks 1976), serial sentence (Kane 1983), parallel sentence (Kane 1983), balanced sentence (Kane 1983), and intersentence coordination (Crystal and Davy 1969). All of these terms suggest a stylistic imitation, the balancing of syntactically similar clauses. We also find evidence of stylistic parallelism below the clause level in the use of terms such as paired adjectives (Kane 1983), balanced phrases (Crystal and Davy 1969), and intrasentence coordination (Hendricks 1976). As well as identifying similar structures, stylists have recognized characteristic asymmetric constructions (Hendricks 1976) or syntactic counterpoint (Hendricks 1976). These types of structures are most often interrupted or convoluted sentences (Kane 1983). For inclusion in our vocabulary, we will choose the most representative terms in the balance group: intersentence coordination and intrasentence counterpoint.</Paragraph>
      <Paragraph position="14"> There are a variety of terms that can be classified into the dominance group. First, there are sentences that have one trivially dominant clause. These are the simple sentences (Kane 1983) that have only one main, or central, clause with no subordinate clauses. Next, there are many and varied types of sentences built up around a central, dominant clause. The complex sentence (Kane 1983) has a central clause and at least one dependent clause. The loose sentence (Kane 1983) has a central clause followed by  Computational Linguistics Volume 19, Number 3 a subordinate clause. The centered sentence (Kane 1983) has a central clause flanked by subordinate clauses. The cumulative sentence (Kane 1983) has a central clause accompanied by a series of appositive, modifying, or absolute constructions. Finally, there are more-complicated sentences with more than one dominant clause. The compoundcomplex sentence (Kane 1983) can be much more elaborate than a simple compound sentence, as it contains at least two independent clauses and at least one dependent clause. For inclusion in our vocabulary, we abstract from these three types of hierarchical terms to identify trivial single-clause dominance, complex single-clause dominance, and multiple-clause dominance.</Paragraph>
      <Paragraph position="15"> The position group appears to contain the largest number of terms, as any basic term can be qualified to an arbitrarily fine degree by the exact intrasentence position to which it refers. We will use only three qualifiers, associated with initial, medial, and final positions (Quirk et al. 1985). At the sentence level, we can have any word or phrase in initial, medial, or final position. Within the sentence, there are almost infinite variations of position terms. A small sample includes postpositive adjectives, and pre-verbal, postverbal, or post-clause adverbs (Crystal and Davy 1969), and complex premodification and postmodification (Crystal and Davy 1969). We will generalize from these many subtle variations and define terms that have either a concordant or discordant effect, at a particular position in a sentence, according to whether or not their usage is normal. From this grouping of terms into the three divisions of balance, dominance, and position, we have abstracted general properties of style that we want to describe in our vocabulary. The abstract elements of style that we will now define will classify the unstructured mass of stylistically significant sentence types into groups of sentences with similar stylistic properties.</Paragraph>
    </Section>
    <Section position="3" start_page="461" end_page="466" type="sub_section">
      <SectionTitle>
3.3 Abstract Elements of Style
</SectionTitle>
      <Paragraph position="0"> We will present the abstract elements in three groups, according to their properties of balance, dominance, and position, which we now define formally as follows: Balance: A balance term characterizes a stylistic effect created by the juxtaposition of similar or dissimilar sentence components.</Paragraph>
      <Paragraph position="1"> For example, parallelism in sentence structure is the juxtaposition of syntactically similar components.</Paragraph>
      <Paragraph position="2"> Dominance: A dominance term describes a stylistic effect created by the particular hierarchical structure of a sentence.</Paragraph>
      <Paragraph position="3"> In a simple sentence, for example, there is an effect of simplicity associated with the single and therefore, by default, dominant clause.</Paragraph>
      <Paragraph position="4"> Position: A position term describes a stylistic effect created by the particular placement of a syntactic component within a sentence.</Paragraph>
      <Paragraph position="5"> For example, in English, a postmodifying adjective, as in the house ablazG can be more emphatic than the more usual premodifying type, as in the blazing house.</Paragraph>
      <Paragraph position="6"> Now we shall propose a set of stylistic terms, correlated with this classification, that will make explicit those abstract stylistic properties that are now only implicit in existing terminology. These stylistic terms will be based on effects of concord and discord, which we formally define as follows: Concord: A stylistic construction that conforms to the norm for a given genre.  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax Discord: A stylistic construction that deviates from the norm.</Paragraph>
      <Paragraph position="7"> We will now use these notions to define abstract elements that are related, in turn, to effects of balance, dominance, and position.</Paragraph>
      <Paragraph position="8">  balance within a sentence. Here, we will be looking at relationships that tend to either perturb or reinforce the balance of a sentence. The first and simplest type of balance element is a homopoise (&amp;quot;same weight&amp;quot;): Homopoise: A sentence with interclausal coordination of syntactically similar components.</Paragraph>
      <Paragraph position="9"> In other words, there are one or more stylistic 'shapes' in a homopoisal sentence, each contributing the same type of effect to the concordant parallelism of the sentence. In the example below, two very simple clauses, with identically shaped parse trees, are conjoined: 13. The style was formed and the principles were acquired. 9 Sentences that are more complex can have their balance interrupted or perturbed by a heteropoisal (&amp;quot;different weight&amp;quot;) component: Heteropoise: A sentence in which one or more parenthetical components are syntactically 'detached' and dissimilar from the other components at the same level in the parse tree.</Paragraph>
      <Paragraph position="10"> It is difficult to formally characterize this class of structures and the underlying notion of a component being 'detached.' Quirk et al. (1985) also rely on this word to define what they call disjuncts but make no attempt to define it: Disjuncts ... have a superior role as compared with the sentence elements; they are syntactically more detached and in some respects 'superordinate,' in that they seem to have a scope that extends over the sentence as a whole. (Quirk et al. 1985, p. 613) ldeg Similarly, Nunberg (1990), in defining the related idea of the delimiter comma, can do no better than to characterize the class &amp;quot;in a rough way&amp;quot;: There is the class of elements delimited by commas, either at both ends (when the elements occur clause-internally) or at one end (roughly, when the elements are either clause-initial or -final) \[p. 36\] .... It is obviously not possible here to provide an analysis of the syntax and semantics of each of these constructions, much less to try to characterize them in such a way as to say what they have in common. In a rough way, however, we observe that they all involve constituency 9 Adapted from The Manchester Guardian Weekly, 14 February 1988, p. 15. 10 An example of a disjunct is the interrupting prepositional phrase in all frankness in the sentence Your son is not, in all frankness, succeeding in his present job (Quirk et al. 1985, p. 612).  Computational Linguistics Volume 19, Number 3 problems, so that they do not behave as constituents, for example, under such tests as verb-phrase ellipsis. What is more, all of them can be characterized as supplying material that is communicatively supererogatory to the bare propositional content of the lexical clause in which they appear.... This is all very rough, but it does suggest that the content of these comma-delimited elements plays a distinctive role in the representation of the \[rhetorical\] argument structure of the text. \[pp. 38-39\] Despite the difficulty of formal characterization, we agree with Quirk et al. and Nunberg that there is an intuitively clear class of constructions here, and, like these authors, hope that it can be conveyed to the reader by sufficient use of examples. We divide the class of heteropoise sentences into three types, depending upon the position of the parenthetical component: a heteropoise may be medial, initial, or final. Moreover, we will distinguish between two subtypes at each position: concordant and discordant.</Paragraph>
      <Paragraph position="11"> Medial heteropoise: A heteropoise in which the parenthetical component is in medial position.</Paragraph>
      <Paragraph position="12"> Concordant medial heteropoise: A medial heteropoise in which the parenthetical component is cohesively linked to the rest of the sentence. (This notion will be made precise when we have introduced the primitive elements in Section 3.4.) Discordant medial heteropoise: A medial heteropoise in which the parenthetical component is not cohesively linked to the rest of the sentence.</Paragraph>
      <Paragraph position="13"> In example 14 below, the relative clause which brought no protests interrupts the main clause, yet, as a postmodifier, it is still a part of the subject noun phrase and so the sentence is a concordant medial heteropoise. But now consider the medial adverbial phrase according to a company spokesman; as a sentence modifier, it is quite detached from the rest of the sentence and so the sentence is simultaneously a discordant medial heteropoise: 14. The measure, which brought no protests, was decided, according to a company spokesman, because of the dangers to which these employees are exposed in travelling to particularly exposed subtropical countries, u Initial heteropoise&amp;quot; A heteropoise, concordant or discordant, in which the parenthetical component is in initial position.</Paragraph>
      <Paragraph position="14"> Sentence 15 contains an initial parenthesis, an appositive noun phrase, that is cohesively linked to the rest of the sentence because it refers to the subject of the main clause. Thus, the sentence is a concordant initial heteropoise:  15. The heir to a fortune, her friend did not care about passing examinations. 12 11 Manchester Guardian Weekly, 21 February 1988, p. 14.</Paragraph>
      <Paragraph position="15"> 12 Quirk et al. (1985, p. 1314).</Paragraph>
      <Paragraph position="16">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax Final heteropoise: A heteropoise, concordant or discordant, in which the parenthetical component is in final position.</Paragraph>
      <Paragraph position="17"> Sentence 16 contains a final parenthesis, an 'extraneous' phrase, that is distinct from the main clause, as it is not cohesively linked to the rest of the sentence. Thus, the sentence is a discordant final heteropoise: 16. Stephen and Jennifer are not going to buy the house, according to a spokesman.</Paragraph>
      <Paragraph position="18"> We saw in 14 that a heteropoise can contain more than one parenthetical, and one of them might be concordant while another is discordant. Likewise, it can be the case that one parenthetical is, say, medial, while another is initial or final. The following example shows two successive medial parentheticals rather stridently interrupting the main text--the effect is heightened by the use of dashes--as well as a final parenthetical in the same sentence: 17. The idea of combined French and British patrols by nuclear submarines--a proposal once made by the Social Democrat leader David Owen--let alone the plan for &amp;quot;sharing the work&amp;quot; where the targets and missiles carried by these submarines are concerned was not even raised, according to a French military spokesman. 13  tic dominance, which is concerned with the hierarchical structure of a sentence. A common type of dominance element is the monoschematic, a very simple sentence: Monoschematic: A sentence with a single main clause with simple phrasal subordination and no accompanying subordinate or coordinate clauses.</Paragraph>
      <Paragraph position="19"> Here is an example of a canonical monoschematic sentence: 18. Posterity has not been kind to him. 14 The most common dominance element in the texts that we have studied is the centroschematic: Centroschematic: A sentence with a central, dominant clause with one or more of the following optional features: complex phrasal subordination, initial dependent clauses, final dependent clauses.</Paragraph>
      <Paragraph position="20"> Centroschematic sentences can be quite varied in structure, but their shared characteristic is a predominant component that serves as the organizational center for all other components. Such structures are built with subordination and coordination. In  13 Manchester Guardian Weekly, 7 February 1988, p. 13.</Paragraph>
      <Paragraph position="21"> 14 Adapted from The Manchester Guardian Weekly, 14 February 1988, p. 15.  Computational Linguistics Volume 19, Number 3 the first example below, the main clause is supported by a complex, but subordinate and coordinate, relative clause structure: 19. Neither these devices nor the cramped viewing rooms which are too narrow and whose ceilings are much too low for the big altarpieces manage to spoil the works. 15 Two subordinate clauses, the first of them adverbial and the second of them relative, accompany but do not dominate the main clause in the following sentence: 20. Not all that long ago, the famous collector Charles de Bestegui, when unable to get hold of certain paintings, was quite prepared to adorn the walls of his Venetian palace with copies, which happily rubbed shoulders with his numerous genuine canvases. 16 Finally, the complex but imitative postmodification in the next example incorporates a substantial amount of information without weakening the dominance of the main clause: 21. Silvia, a commanding woman in her 50's, a shrew falsely mellowed by religion, promptly organised prayer sessions on the lines of Tupperware meetings. 17 Given that we have a progression in complexity from monoschematic to centroschematic sentences, a natural extension is to the polyschematic: Polysehematic: A sentence with more than one central dominant clause and at least one dependent clause.</Paragraph>
      <Paragraph position="22"> Such sentences occur much less frequently than the monoschematic or centroschematic varieties, at least in the corpus used in this research. However, a sentence with obviously disparate components occasionally occurs, as in the following example. There are two dominant clauses (we could think.., and we should not forget ... ) and a dependent clause (if we consider the progress ... ): 22. If we consider the progress already achieved, the opposition that had to be overcome, for example, in order to open schools for girls, and the fact that Saudi Arabia is less than 60 years old, we could think that time will permit resolving the contradictions between the most liberal aspirations of one part of society and the ulemas' determination to keep the country as it is, and we should not forget to mention the Islamic fundamentalist movements which are threatening Saudi Arabia. TM  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax 3.3.3 Position Elements. The third group of abstract elements that we will look at are  the position elements. The most common types of position element describe concordant or discordant stylistic effects in particular positions. Thus, we have initial concord, medial concord, and final concord. Similarly, we have a range of discord elements. The definitions of these elements are closely tied to the primitive-level descriptions of our vocabulary, so we will delay their precise characterizations until Section 3.5, after the primitive-level descriptions are introduced. Here, it will be sufficient for the reader to know only that the concords describe constructions, at a particular position in the sentence, that conform to normal usage while the discords describe constructions that deviate from the norm.</Paragraph>
      <Paragraph position="23"> In addition to effects created by a single concord or discord in a particular position within the sentence, we can observe other kinds of effects pertaining to the relationships between concords and discords: Resolution: A shift in stylistic effect that occurs at the end of a sentence and is a move from a relative discord to a stylistic concord.</Paragraph>
      <Paragraph position="24"> We can see an example of a resolution in 23 below. There is an initial discord, created by the unusual placement of the adverb entirely, which is not normally found in the initial position and is not cohesively linked to the rest of the sentence (cf. Section 3.4.3 below). However, there is a final concord, as the subsequent main clause contains no such incongruities: 23. Entirely in the spirit of protective support, could we suggest you pass on an appropriate comment to the personnel concernedJ 9 And the complementary effect is dissolution: Dissolution: A shift in stylistic effect that occurs at the end of a sentence and is a move from a relative concord to a stylistic discord.</Paragraph>
      <Paragraph position="25"> We have already seen an example of a dissolution in one of the Biblical texts of Section 1.1, repeated below as 24. A strong initial concord, created by imitative clauses, is set against a final discord, produced by an uncommon syntactic inversion: 24. And the rain descended, and the floods came, and the winds blew, and beat upon that house; and it fell: and great was the fall of it. 2deg</Paragraph>
    </Section>
    <Section position="4" start_page="466" end_page="473" type="sub_section">
      <SectionTitle>
3.4 Primitive Elements of Style
3.4.1 Two Kinds of Stylistic Analysis. Now that a set of abstract stylistic elements has
</SectionTitle>
      <Paragraph position="0"> been defined, we must next consider how to use them in practical stylistic analysis.</Paragraph>
      <Paragraph position="1"> So far, these elements are too general to apply to the interpretation of an arbitrary sentence. We must relate these stylistic elements to more-basic syntactic elements, whose stylistic characteristics are more specific and concrete. In our reading of the literature  Computational Linguistics Volume 19, Number 3 of stylistics, we observed that two analyses of stylistic structure are possible: connective and hierarchic orderings: Connective ordering: The result of cohesive bonds drawing together components in a linear ordering.</Paragraph>
      <Paragraph position="2"> Hierarchic ordering: The result of bonds of subordination and superordination drawing together components in a nested ordering.</Paragraph>
      <Paragraph position="3"> These two complementary kinds of analysis are implicit in the work of most stylists and rhetoricians, such as Cluett (1976) and Bureau (1976) (Neil Randall, personal communication). null We will now use these notions of connective and hierarchic orderings to guide the definition of more-primitive stylistic elements that provide a precise syntactic basis to the vocabulary, yet also allow a mapping to the abstract elements.</Paragraph>
      <Paragraph position="4">  of Style. We believe that to establish stylistic classifications and construct a practical stylistic grammar, one must undertake corpus studies in the style of Cluett (1976) and Biber (1988, 1989) to support the exact classifications of stylistic effects. But we also agree with Crystal and Davy that: It is perhaps worth emphasising right away that the first step in any stylistic analysis must be an intuitive one. The stylistician is on precisely the same footing as anyone else here: he notices a linguistic feature which he feels to be stylistically significant. The difference between his approach and that of the untrained observer is that he will have a clearer idea of what is likely to be significant, and will know what to do with his observations once they are made. This last point is the stylistician's main competence: he is able to interrelate his observations within the framework of some theory, and thus piece together any general pattern of linguistic variation which may exist. (Crystal and Davy 1969, p. 12) Thus, although we have built our grammar of style upon accepted theory, it is currently a prototype for the formal representation of stylistic knowledge, not yet developed as a tool for practical stylistic analysis. 21 For, like Crystal and Davy, our initial objective is to provide a methodology, a vocabulary and apparatus, that will allow others to integrate our approach with analytic procedures to construct useful formalizations of style: All we can do is systematically point to certain significant facts in the language being analysed, suggest some theoretical principles which will account for the occurrence of these facts, emphasise the need for further analysis to validate or refute these preliminary soundings, and illustrate a procedure which will allow people to do this. (Crystal and Davy 1969, p. 13) 21 See Section 6 for a discussion of work in progress.</Paragraph>
      <Paragraph position="5">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax In the manner of Crystal and Davy, we agree that statistical verification must follow such research, but we believe that, initially, we can use well-motivated intuitions about stylistic frequencies to classify our stylistic elements: Such a \[statistical\] method would be prerequisite for any serious research work, but it is too detailed for our present purposes. Consequently we make use of the alternative method of expression, making our statements of frequency in more informal terms, using such quantifiers as 'rarely', 'commonly', 'often', 'very often', and so on. This range of adverbials in English cuts up a continuum of frequency very clearly, with very little overlap, and is readily intelligible. It should not be forgotten, however, that any statements of relative frequency in these terms can be referred if necessary to the precise statistical situation which underlies them. (Crystal and Davy 1969, p. 22)  stituents that have a particular stylistic effect. But what should we use as the basis for the definition of 'stylistic effect'? We have said that the two analyses of sentence structure, connective and hierarchic, can provide the appropriate guidance. We will define the stylistic effect of an individual component in terms of its contribution to each ordering: * For the connective view, a component acquires its stylistic effect from its degree of cohesiveness, its bonding with other components in the sentence. * For the hierarchic view, a component acquires its stylistic effect from its degree of subordination, its dependence on other components in the sentence.</Paragraph>
      <Paragraph position="6"> We introduce the terms conjunct, antijunct, subjunct, and superjunct (to be defined in the following sections) and we use superscripts on all the terms to indicate the degree of connectivity or subordination. In the connective view, we classify syntactic components as either conjunct 4 (excessively connective), conjunct 3 (strongly connective), conjunct 2 (moderately connective), conjunct 1 (mildly connective), and conjunct deg (neutral). Similarly, we use the terms antijunct deg through antijunct 4 to indicate increasingly disconnective effects; conjunct deg and antijunct deg are the same.</Paragraph>
      <Paragraph position="7"> There is a complementary vocabulary of primitive elements for the hierarchic view. The stylistic effects of syntactic components are correlated with the degree of subordination or superordination; the classifications are analogous to the connective: subjunct 4 through subjunct deg (decreasingly subordinate) and superjunct deg through superjunct 4 (increasingly superordinate); subjunct deg and superjunct deg are the same. Connective primitive elements. In assigning connective effects to syntactic components, we adapt Halliday and Hasan's (1976) work on cohesion. Although Halliday and Hasan focused on definitions of cohesion in whole texts, we can apply these definitions to cohesion within the sentence, for, as they point out: Since cohesive relations are not concerned with structure, they may be found just as well within a sentence as between sentences. They attract less notice within a sentence, because of the cohesive strength of grammatical structure; since the sentence hangs together already,  Computational Linguistics Volume 19, Number 3 the cohesion is not needed in order to make it hang together. But the cohesive relations are there all the same. (Halliday and Hasan 1976, p. 8) Halliday and Hasan enumerate five types of cohesion: substitution, ellipsis, reference, conjunction, and lexical cohesion. We will use all of these except lexical cohesion and will add a new factor, interpolation, that works against cohesion.</Paragraph>
      <Paragraph position="8"> Halliday and Hasan define substitution as the &amp;quot;replacement&amp;quot; (p. 88) of one item by another. We suggest that such intrasentence phenomena as apposition and extraposition can be considered as types of substitution. For example, in 25, either of the two postmodifying noun phrases could substitute for the head noun Silvia: 25. Silvia, a commanding woman in her 50s, a shrew falsely mellowed by religion, promptly organized prayer sessions on the lines of Tupperware meetings. 22 And, in 26, the clausal subject for anyone to escape could substitute for the anticipatory pronoun it: 26. It was considered impossible for anyone to escape. 23 Ellipsis is defined as the omission of an item. Many forms of ellipsis are possible within the sentence, including the omission of elements of clause structure. In the following sentence, the ellipsis is marked by .....</Paragraph>
      <Paragraph position="9"> 27. I'll gladly pay for the hotel, if you will __ for the food. 24 Items that have the cohesive property of reference are not interpreted semantically in their own right, but make reference to something else for their interpretation (Halliday and Hasan 1976, p. 31). In English, such items are personal pronouns, demonstratives, and comparatives. These 'directive' elements presuppose the existence of some other element from which information is to be retrieved; whether or not the other element is within the same sentence, the referential item will be cohesive. Conjunctive elements are defined as: elements \[that\] are not cohesive in themselves but indirectly, by virtue of their specific meanings; they are not primarily devices for reaching out into the preceding (or following) text, but they express certain meanings which presuppose the presence of other components in the discourse. (Halliday and Hasan 1976, p. 226) Halliday and Hasan cite the conjunction and, certain adverbs, and certain prepositional expressions as instances of conjunction. Many of these elements can promote cohesion within the sentence.</Paragraph>
      <Paragraph position="10"> To the principles of cohesion described above, we will add interpolation, a discon- null nective relation that works against cohesion. Interpolated elements are certain instances 22 Adapted from The Manchester Guardian Weekly, 7 February 1988. Translated from Le Monde. 23 Quirk et al. 1985, p. 1392.</Paragraph>
      <Paragraph position="11"> 24 Quirk et al. 1985, p. 907.</Paragraph>
      <Paragraph position="12">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax of parenthetical constructions, those that display none of the forms of cohesion listed earlierY In the examples below (repeated from Section 3.3.1), we see several different types of interpolation: 28.</Paragraph>
      <Paragraph position="13"> 29.</Paragraph>
      <Paragraph position="14"> 30.</Paragraph>
      <Paragraph position="15"> The measure, which brought no protests, was decided, according to a company spokesman, because of the dangers to which these employees are exposed in travelling to particularly exposed subtropical countries. 26 The heir to a fortunG her friend did not care about passing examinations. 27 Stephen and Jennifer are not going to buy the house, according to a spokesman.</Paragraph>
      <Paragraph position="16"> The cohesive relations described above can be ranked from the most to the least cohesive. Halliday and Hasan (pp. 226-227) consider substitution, including ellipsis, to be the most strictly cohesive relation, followed by reference, and then conjunction. We adopt this ranking, and so we classify intrasentential substitution and ellipsis as strongly connective (conjunct3), reference as moderately connective (conjunct2), and conjunction as mildly connective (conjunct 1). We classify interpolation as disconnective (antijunct2). 28 As an illustration of our method of classification of connective primitive elements, we will use the following sentences, in which the underlined components exhibit distinct stylistic effects. In the first sentence, the postmodifying construction is a strongly connective, conjunct 3 reduced relative clause, as it is an instance of ellipsis for which the presupposed items are who is. The postmodifying relative clause in 32 is a moderately connective, conjunct 2 referential element, as the relative pronoun 'points' to the preceding noun phrase the man. Lastly, the postposed adjectival in 33 is a disconnective, antijunct instance of interpolation, as it is detached from the rest of the sentence and has none of the cohesive properties. (The property of ellipsis is not present, as a postposed adjective, unlike a reduced relative clause, does not have the same element of presupposition of specific missing items.):  31. The man __ walking into the room is unfit for this task.</Paragraph>
      <Paragraph position="17"> 32. The man who is walking into the room is unfit for this task.</Paragraph>
      <Paragraph position="18"> 33. A man always timid is unfit for this task. 29 Hierarchic primitive elements. In assigning a hierarchic classification to a syntactic component, we adapt Halliday's (1985) work on subordination in functional grammar and Quirk et al.'s (1985) use of the term superordination.</Paragraph>
      <Paragraph position="19"> 25 Another type of structure that seems to work against cohesion is a disruption in normal linear ordering, as the syntactic inversion at the end of the now-familiar text, And the rains descended, and the floods came, and the winds blew, and beat upon that house; and it fell: and great was the fall of it. However, we have not yet incorporated this feature into our formalization of style.</Paragraph>
      <Paragraph position="20"> 26 Manchester Guardian Weekly, 21 February 1988, p. 14.</Paragraph>
      <Paragraph position="21"> 27 Quirk et al. (1985, p. 1314).</Paragraph>
      <Paragraph position="22"> 28 We do not use here the full range of connectivity that we defined above; however, some extensions of the theory by Green (1992a, 1992b) do.</Paragraph>
      <Paragraph position="23"> 29 Quirk et al. (1985, p. 1295).</Paragraph>
      <Paragraph position="24">  Computational Linguistics Volume 19, Number 3 Halliday distinguishes two types of subordination: embedding (or rank-shifting) and hypotaxis. These phenomena will be used to classify the hierarchic primitive elements in our stylistic grammar. Embedding is described as: the 'rank shift' by which a clause or phrase comes to function within the structure of a group, like who came to dinner in the man who came to dinner. (Halliday 1985, p. 219) There are various types of embeddings, including prepositional phrases, as in John might arrive as early as tomorrow, but Halliday focuses on those within the nominal group: \[The embedded prepositional phrase\] is the only instance of embedding other than in a nominal group. All other embedding in English is a form of nominalization, where a group, phrase, or clause comes to function as part of, or in place of (i.e., as the whole of), a nominal group. (Halliday 1985, p. 187) Examples of embedded subordination are finite and nonfinite clauses 3deg that function as postmodifiers, as in the house that lack built, or as the head of a nominal group, as in For Jack to build a house would be the best thing (Halliday 1985, p. 220). Halliday defines hypotaxis as: the binding of elements of unequal status. The dominant element is free, but the dependent element is not. \[p. 198\] In a hypotactic structure the elements are ordered in dependence, and this ordering is largely independent of the sequence. \[p. 199\] There are many types of hypotaxis, at the clause, phrase, and word levels. Examples of hypotactic subordination include a dependent clause following a dominant clause, preceding the dominant, enclosed in the dominant, or enclosing the dominant:  34. You never can tell till you try. 31 35. If wishes were horses, beggars would ride. 32 36. Picture, if you can, a winkle. 33 37. He might, he said, finish it himself. 34  It should be emphasized that the two types of subordination are quite different: It is important to distinguish between embedding on the one hand and ... hypotaxis on the other. Whereas hypotaxis \[is a relation\] BETWEEN clauses (or other ranking elements), embedding is not. Embedding is  30 If the clause is relative, then it must be defining (restrictive) to be an embedded clause; a non-defining relative clause is an instance of hypotaxis (to be defined below).</Paragraph>
      <Paragraph position="25"> 31 Halliday 1985, p. 200.</Paragraph>
      <Paragraph position="26"> 32 Halliday 1985, p. 200.</Paragraph>
      <Paragraph position="27"> 33 Halliday 1985, p. 200.</Paragraph>
      <Paragraph position="28"> 34 Halliday 1985, p. 200.</Paragraph>
      <Paragraph position="29">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax a mechanism whereby a clause or phrase comes to function as a constituent WITHIN the structure of a group, which itself is a constituent of a clause. (Halliday 1985, p. 219) Consequently, the stylistic effects of embedding and hypotaxis will be subtly different. Because an embedded structure is a constituent of the dominant structure, we will classify embeddings as strongly subordinate, subjunct 3. On the other hand, in hypotaxis, one component is dependent on another, but &amp;quot;in no sense is it a constituent part of it&amp;quot; (Halliday 1985, p. 219), so we will classify hypotactic structures as only mildly subordinate, subjunct 1.</Paragraph>
      <Paragraph position="30"> To the factors of embedding and hypotaxis, we add another factor, superordination, that accounts for structures that are somehow 'superior' to the main clause. 35 That such phenomena exist is supported by Quirk et al.'s (1985) definition of disjunct, which we repeat here from Section 3.3.1: Disjuncts ... have a superior role as compared with the sentence elements; they are syntactically more detached and in some respects 'superordinate,' in that they seem to have a scope that extends over the sentence as a whole. (Quirk et al. 1985, p. 613) Disjuncts can be divided into two main classes, style disjuncts and content disjuncts: Style disjuncts convey the speaker's comment on the style and form of what he is saying, defining in some way under what conditions he is speaking as the 'authority' for the utterance. Content disjuncts (also known as attitudinal disjuncts) make observations on the actual content of the utterance and its truth conditions. (Quirk et al. 1985, p. 615) Examples of these superordinate elements are underlined in the following sentences (the first three contain style disjuncts, the fourth, a content disjunct): 38. Frankly, I am tired. 36 39. If I may say so without offence, your writing is immature. 37 40. There were twelve people present, to be precise. 38 41. Even more important, she has control over the finances of the party. 39 In our grammar, we classify these types of superordinate elements as superjunct. As an illustration of our method of classification of hierarchic primitive elements, we will use the following sentences, which display distinct stylistic effects due to the different primitive classifications of the underlined components. In the first sentence below, the postmodifying restrictive relative clause is a strongly subordinate, subjunct 3 instance of embedding. The medial adverbial clause in 43 is a mildly subordinate,  35 We do not consider the case of a main clause being superordinate to a dependent clause (Quirk et al. 1985, pp. 988-990). We are concerned only with structures that are superordinate to the matrix clause. 36 Quirk et al. 1985, p. 615.</Paragraph>
      <Paragraph position="31"> 37 Quirk et al., p. 615.</Paragraph>
      <Paragraph position="32"> 38 Quirk et al., p. 616.</Paragraph>
      <Paragraph position="33"> 39 Quirk et al., p. 622.</Paragraph>
      <Paragraph position="34">  Computational Linguistics Volume 19, Number 3 subjunct 1 instance of hypotaxis, as it is a causal-conditional clause (Halliday 1985, p. 214). Lastly, the initial adjective in 44 is a superordinate, superjunct instance of a style disjunct.</Paragraph>
      <Paragraph position="35"> 42. The money which was our reward has been useful.</Paragraph>
      <Paragraph position="36"> 43. The money, as it was our reward_, should not be used for frivolous purposes.</Paragraph>
      <Paragraph position="37"> 44. True, the money has been useful.</Paragraph>
    </Section>
    <Section position="5" start_page="473" end_page="473" type="sub_section">
      <SectionTitle>
3.5 Concords and Discords
</SectionTitle>
      <Paragraph position="0"> In Section 3.3.3, we gave an informal description of the abstract elements of concord and discord. As these elements are closely tied to the primitive-level descriptions of our vocabulary, we delayed their precise characterizations until after the primitive elements of style were introduced. Now, we can define the nature of concords and discords for both the connective and hierarchic views. In both views, the concords represent constructions, at a particular position in the sentence, that conform to normal usage, while the discords describe constructions that deviate from the norm.</Paragraph>
      <Paragraph position="1"> In the connective view, the concords are associated with constructions that are connective, for we assume that cohesion is the default usage, while the discords are disconnective structures, as a lack of cohesion is less usual. Thus, a concord, which may be initial, medial, or final, is produced by a conjunct construction; a discord, which may also appear in any of the three positions, is caused by an antijunct construction.</Paragraph>
      <Paragraph position="2"> For the hierarchic view, concords are associated with structures that display a 'normal' degree of subordination or superordination, while discords are produced by excessive subordination or excessive superordination. Excessive subordination and superordination are difficult to characterize, but, as a first attempt, we propose that nesting of dependent structures to a depth greater than three will be considered discordant subordination and that 'detachment' of superordinate structures by 'intrusive' punctuation (e.g., long dashes) will be considered discordant superordination.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="473" end_page="493" type="metho">
    <SectionTitle>
4. A Stylistic Grammar
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="473" end_page="474" type="sub_section">
      <SectionTitle>
4.1 A Methodology for Developing a Stylistic Grammar
</SectionTitle>
      <Paragraph position="0"> So far, we have constructed a vocabulary of style. We can now define a method for using the vocabulary to systematically build up a syntactic stylistic grammar for both the connective and hierarchic views of style.</Paragraph>
      <Paragraph position="1"> To construct the stylistic grammar, we will use a methodology whose steps are as follows: * the classification of primitive stylistic elements; * the correlation of the stylistic effects of these elements with the abstract stylistic elements; and * the correlation of patterns of these abstract elements with specific stylistic goals.</Paragraph>
      <Paragraph position="2"> The grammar to be constructed in this manner will be stratified with internal branching, as illustrated in Figure 1. As the figure shows, at the bottom level there are three branches--lexical, syntactic, and semantic--each with its own vocabulary  A stylistic grammar.</Paragraph>
      <Paragraph position="3"> of primitive stylistic elements and rules for combining them. At the central level, we use a single vocabulary of abstract elements, stylistic terms that are maximally expressive. Rules relate these abstract elements to patterns of lexical, syntactic, and semantic primitive elements. This level is the unifying core of the grammar; as we will note later, the same abstract elements can describe both English and French style, and both syntactic and semantic aspects of style. Finally, at the top level, rules correlate individual stylistic goals with patterns of abstract elements. Together, these levels form a language-independent 4deg schema for a goal-directed grammar.</Paragraph>
    </Section>
    <Section position="2" start_page="474" end_page="475" type="sub_section">
      <SectionTitle>
4.2 Grammar of Primitive Elements
</SectionTitle>
      <Paragraph position="0"> The stylistic grammar that we have outlined in the previous section shows how lexical, syntactic, and semantic aspects of style would be integrated within one grammar. Some work has already been done on the development of the lexical and semantic branches (see Sections 6.1 and 6.2), but here we will focus on syntactic style. We will now trace the development of the syntactic component of the stylistic grammar, beginning with the most detailed and most closely syntactic level, the grammar of primitive elements.</Paragraph>
      <Paragraph position="1"> As we proposed earlier, we will consider two analyses of sentence structure, connective and hierarchic, each with its own stylistic classifications. In this base-level grammar of 40 We expect that a particular stylistic grammar will be applicable to a family of stylistically similar languages.  Computational Linguistics Volume 19, Number 3 primitive elements, the rules are based on syntactic and stylistic rules of Crystal and Davy (1969), Kane (1983), and Quirk et al. (1985).</Paragraph>
      <Paragraph position="2"> In the full grammars given by DiMarco (1990) and subsequently by Hoyt (1993), each syntactic category is given both a connective and a hierarchic primitive classification by corresponding rules. Here, we will present only selected rules, mainly those associated with the stylistic effects of various types of premodification and postmodification in the nominal group. We will begin with the elementary components and work up to full sentences. In the presentation of the grammar rules, we will first give the connective rules, then their hierarchic counterparts. Often, we will annotate rules with relevant examples. (Explanatory notes on the terminology are given in Appendix A.) Adjectivals. Adjectivals and adverbials are the simplest classes of primitive elements.</Paragraph>
      <Paragraph position="3"> To demonstrate that even the smallest syntactic components have individual stylistic effects, we give selections from the rules for the adjectivals.</Paragraph>
      <Paragraph position="4"> Connective view. Adjectivals can be classified within a range of connectivity. In the moderately connective case, there are a number of alternatives, two of which are shown below. The first alternative, the genitive form, is a referential element, as it presupposes the existence of the noun to which it refers. Whether or not this noun is in the same sentence, a referential personal pronoun is cohesive and so it is classified as a conjunct 2 element. The second alternative, a demonstrative determiner, is also classified as moderately connective because it too is a referential element.</Paragraph>
      <Paragraph position="5">  The less-conjunct case has only one alternative, an adjective. An adjective is usually used as a premodifier of a noun, so that it presupposes the presence of the noun.</Paragraph>
      <Paragraph position="6"> However, it is not a directive, referential element, but rather a conjunctive one. Thus, it is classified as conjunct 1.</Paragraph>
    </Section>
    <Section position="3" start_page="475" end_page="478" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> conjunct 1 adjectival adjective medieval iconographic subject Hierarchic view. Adjectivals are generally hypotactic rather than embedded elements. That is, they are linked to an element (the noun) of different status that could stand as an independent whole, whereas the adjectival could not. Therefore, we will classify most adjectivals as subjunct 1, mildly subordinating. However, nominal adjectives, which have been rank-shifted, are instances of embedding and so are subjunct 3, more strongly subordinating.</Paragraph>
      <Paragraph position="1">  rightly recognized the stylistic importance of premodification and postmodification of the nominal group. Our syntactic stylistic grammar is built upon the stylistic effects created by these types of modification. We first describe selected rules of premodification. Connective view. The rules for moderately connective, conjunct 2, and less connective, conjunct 1, premodification are self-explanatory.</Paragraph>
      <Paragraph position="2">  conjunct 1 premodification conjunct 1 adjectival The following rule for non-connective, conjunct deg premodification has three alternatives, an adverbial, a participle, and no premodification. Premodifying adverbials and participles do not display any of the explicit forms of cohesion but neither do they work against cohesion. They are not connective but neither are they disconnective.  Computational Linguistics Volume 19, Number 3 The following rule for disconnective, antijunct 2 premodification has only one alternative, a reduced sentence. A reduced sentence can be considered to be an instance of interpolation, so that it works against cohesion.</Paragraph>
      <Paragraph position="3">  We can now begin to introduce primitive stylistic categories that will be used to build up the definitions of the corresponding abstract elements that we saw in earlier sections. When premodification is conjunct, it is classified as centroschematic premodification. In turn, this type of premodification is concordant, for it is associated with cohesive effects, which we take to be normal usage.</Paragraph>
      <Paragraph position="4">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax Hierarchic view. In the full grammar, we define detailed rules for the various subjunct and superjunct types of premodification. Here, we will show only selected rules. The following rule for strongly subordinating, subjunct 3 premodification has four alternatives. The first is self-explanatory. The other three, reduced sentences, adverbials, and participles, are embedded, rank-shifted constructions, which are therefore classified as subjunct 3.</Paragraph>
    </Section>
    <Section position="4" start_page="478" end_page="478" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> subjunct 3 premodification --~ subjunct 3 adjectival reduced sentence adverbial participle The following rule for less subordinating, subjunct I premodification is self-explanatory.</Paragraph>
    </Section>
    <Section position="5" start_page="478" end_page="478" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> subjunct I premodification --~ subjunct I adjectival We include the rule below to account for the case of no premodification, which is neither subordinating nor superordinating.</Paragraph>
    </Section>
    <Section position="6" start_page="478" end_page="478" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> subjunct deg premodification no premodification Centroschematic premodification is produced by subjunct or 'neutral' (subjunct deg) premodifying constructions.</Paragraph>
    </Section>
    <Section position="7" start_page="478" end_page="479" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> centroschematic premodification subjunct 3 premodification subjunct 2 prernodification subjunct I prernodification subjunct deg prernodification In the hierarchic view, discord is produced by excessive subordination. We will classify this type of discordant premodification as subjunct 4 premodification. Now we can define the following rules for concordant and discordant premodification:</Paragraph>
      <Paragraph position="2"> Postmodification. Continuing to build upon the stylistic features that Crystal and Davy (1969) judged important, we define complementary connective and hierarchic rules for postmodification in the full grammar and give selected rules here.</Paragraph>
      <Paragraph position="3"> Connective view. The rule for conjunct 3 postmodification has three alternatives, a nominal group, a nonfinite clause, and a verbless clause. 42 As these constructions are instances of either substitution or ellipsis, they are strongly connective.</Paragraph>
    </Section>
    <Section position="8" start_page="479" end_page="479" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> conjunct 3 postmodification nominal group Paul Jones, the distinguished art critic, died in his sleep last night. non-finite clause You will look in vain for any concrete measures emerging from this summit. verbless clause Norman Jones, then a student, wrote several best-sellers. As a relative clause is a referential structure, it is classified as a moderately connective, conjunct 2 shape.</Paragraph>
      <Paragraph position="1">  Now we can define the various types of concordant and discordant postmodification, according to whether the postmodification is cohesive (and therefore concordant) or lacking cohesion (discordant).</Paragraph>
      <Paragraph position="2"> 43 We do not define all forms of parenthesis, but these can be easily enumerated by using, for example, the classifications given by Quirk et al. (1985).</Paragraph>
      <Paragraph position="3">  Hierarchic view. In the rule for subjunct 3 postmodification, the alternatives are a restrictive relative clause, a restrictive nonfinite clause, a nominal group, and a prepositional phrase. All of these are instances of embedding, as they function as a part of the nominal group, so they are strongly subordinating elements.</Paragraph>
    </Section>
    <Section position="9" start_page="479" end_page="483" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> subjunct 3 postmodification restrictive relative clause restrictive non-finite clause nominal group prepositional phrase In the rule for subjunct I postmodification, the alternatives are an adjectival, a nonrestrictive relative clause, and a nonrestrictive nonfinite clause. All of these are instances of hypotaxis, as none of them is a constituent of the nominal group, so they are classified as mildly subordinating elements.</Paragraph>
      <Paragraph position="1">  subjunct 3 postmodification subjunct 1 postmodification subjunct deg postmodification In the hierarchic view, discord is produced by excessive subordination. We will classify this type of discordant postmodification as subjunct 4 postmodification. Now we can define the concordant and discordant varieties of hierarchic postmodification as follows:  Computational Linguistics Volume 19, Number 3 The following stylistic variations of noun phrases are defined in the obvious way, according to the corresponding types of their premodification and postmodification: monoschematic noun phrase centroschematic noun phrase heteropoisal noun phrase concordant noun phrase discordant noun phrase A question that arises at this point is the projection problem; that is, the problem of how the stylistic effect of a sentence, or even a sentence component, is modified by the stylistic characteristics of the components nested within it. Our solution is to assume that if a component is concordant then all its components, to all levels of nesting, must be concordant, but if a component is discordant then at least one of its components, at some level of nesting, must have been discordant. Thus, if we have a noun phrase consisting of concordant premodification, a noun, and discordant postmodification, the whole noun phrase will be classified as discordant.</Paragraph>
      <Paragraph position="2"> We realize this is a very simplified solution, but we believe it will be acceptable in our work, for our sample corpus, although drawn from high-quality magazine writing, has few examples of nesting to more than one or two levels. Verb phrases. The basic verb phrase is defined as follows:</Paragraph>
    </Section>
    <Section position="10" start_page="483" end_page="484" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> verb phrase (adverbial) verb (adverbial) (complement)* The following stylistic variations of verb phrases are defined in the obvious way, according to the types of their components: monoschematic verb phrase centroschematic verb phrase heteropoisal verb phrase concordant verb phrase discordant verb phrase Sentences. We also define rules in the full grammar for complements, prepositional phrases, and dependent clauses. With these rules, we have all the components of a sentence and can define sentence structures of increasing complexity. The first of these structures is the major, a single main clause; the next is the complete, which adds dependent clauses to the main clause. In some rules, we introduce specializations according to the position of a certain type of stylistic element. An initial discordant complete, for example, is a sentence that has a discordant component in the initial position.</Paragraph>
      <Paragraph position="1">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax The basic major is defined as follows:</Paragraph>
    </Section>
    <Section position="11" start_page="484" end_page="485" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> major --~ (conjunction) (adjective)* (adverbial)* (prepositional phrase)* (nominal group)* noun phrase verb phrase We define the following stylistic specializations of major sentences, according to the types of their components:  We define stylistic specializations of complete sentences in a manner analogous to those for major sentences, so that our grammar includes the following varieties: monoschematic complete centroschematic complete heteropoisal complete initial heteropoisal complete medial heteropoisal complete final heteropoisal complete concordant complete discordant complete initial concordant complete medial concordant complete final concordant complete  Computational Linguistics Volume 19, Number 3 initial discordant complete medial discordant complete final discordant complete</Paragraph>
    </Section>
    <Section position="12" start_page="485" end_page="486" type="sub_section">
      <SectionTitle>
4.3 Grammar of Abstract Elements
</SectionTitle>
      <Paragraph position="0"> Having completed the construction of the bottom level of the stylistic grammar, the classification of primitive stylistic elements, we can now define the central level, the grammar of abstract elements. This will correlate the stylistic effects of the primitive elements with the abstract elements that were defined in Section 3.3. At this level, the connective and hierarchic views become integrated.</Paragraph>
      <Paragraph position="1"> A monoschematic sentence is a single main clause with optional, simple forms of subordination.</Paragraph>
      <Paragraph position="2">  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax The various types of concord and discord are defined according to the presence of a concordant or discordant component in initial, medial, or final position.  As well as observing localized effects of concord and discord, we can recognize more-global shifts from concord to discord, or discord to concord. That is, a specific part of the sentence seems incongruous, discordant, but this discord is interpreted within the context of the rest of the sentence, so that it may either be 'resolved' by a return to a usual form, or left unresolved. We have formalized these two situations in the abstract elements of resolution and dissolution. A resolution is a shift in stylistic effect that occurs at the end of a sentence and is a move from a relative discord to a concord. A dissolution is a shift in stylistic effect that occurs at the end of a sentence and is a move from a relative concord to a discord.</Paragraph>
    </Section>
    <Section position="13" start_page="486" end_page="487" type="sub_section">
      <SectionTitle>
Rule
</SectionTitle>
      <Paragraph position="0"> resolution initial discord and final concord</Paragraph>
    </Section>
    <Section position="14" start_page="487" end_page="487" type="sub_section">
      <SectionTitle>
Computational Linguistics Volume 19, Number 3
Rule
</SectionTitle>
      <Paragraph position="0"> dissolution initial concord and final discord</Paragraph>
    </Section>
    <Section position="15" start_page="487" end_page="488" type="sub_section">
      <SectionTitle>
4.4 Grammar of Stylistic Goals
</SectionTitle>
      <Paragraph position="0"> Having completed the central level in the stylistic grammar, the grammar of abstract elements, we can now define the rules at the top level that correlate patterns of the abstract elements with the writer's specific goals.</Paragraph>
      <Paragraph position="1"> Stylistic goals, such as clarity, are elusive qualities that up to now have been defined by stylists by means of examples and informal rules. However, with the vocabulary and formal grammar that we have defined, we now have a way of seeing and abstracting what these examples have in common. We can abstract from a plethora of low-level syntactic rules that stylists traditionally have used, and can now define formal rules for specific stylistic goals.</Paragraph>
      <Paragraph position="2"> Note that stylistic goals can be organized along orthogonal dimensions. For example, a writer might try to be clear, or obscure, or make no effort either way. Clarity and obscurity are thus opposite ends of a stylistic dimension. Likewise, the goals of concreteness and abstraction form a dimension, and so do staticness and dynamism. Below, we look at one end of each of these dimensions; for details of their duals, see DiMarco (1990) and Hoyt (1993).</Paragraph>
      <Paragraph position="3"> Clarity. Clarity is a very pervasive stylistic goal. Almost all the advice in textbooks of style and rhetoric is aimed at teaching the writer how to achieve clarity. Clarity, in other words, is thought to be the norm that a writer should strive for. To be clear is to be plain, precise, and predictable. In keeping with stylistic theory, therefore, we interpret the goal of clarity as adherence to the stylistic norm and incorporate accepted definitions of norm, that is, cohesive and concordant structures, in our grammar rule for clarity.</Paragraph>
      <Paragraph position="4"> Thus, we can use our stylistic grammar to give precise definitions to the kinds of sentences that Kane (1983), for example, associates with clarity: * Simple sentences, which consist of one independent clause: these are our monoschematic sentences.</Paragraph>
      <Paragraph position="5"> * Centered sentences, which consist of dependent constructions, followed by a main clause, followed by additional dependent clauses. We expand the notion of 'centered sentence' to include any sentence in which there is a dominant, concordant core: these are our centroschematic sentences.</Paragraph>
      <Paragraph position="6"> * Parallel sentences, which reduce ambiguity by stressing the same grammatical form: these are our homopoisal sentences.</Paragraph>
      <Paragraph position="7"> Hence, we define clarity as follows:  Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax Concreteness. Kane (1983) associates concreteness with sentences that suggest an effect of immediacy, in which the writer has arranged elements to reflect the natural order of events or ideas, so that syntax mirrors events. In the terms of our grammar, concreteness is associated with sentences that emphasize a particular component, which may be highlighted either because it is discordant or because it is parenthetical, as in a heteropoise.</Paragraph>
      <Paragraph position="8">  Staticness. We associate staticness with sentences in which there is little opportunity for stylistic variation, that is, sentences that verge on being 'fixed forms.' These are the monoschematic sentences, which are standard and simple structures, or the more elaborate but strictly balanced structures, the homopoisal sentences.</Paragraph>
      <Paragraph position="9">  5. STYLISTIQUE: A Syntactic Stylistic Parser  So far, we have developed a vocabulary and methodology for constructing a stylistic grammar and we have shown how these tools have been applied to the construction of an English syntactic stylistic grammar. In DiMarco (1990), we applied the same vocabulary and methodology to develop a French syntactic stylistic grammar. However, these grammars provide only a theoretical foundation for a computational theory of stylistics. Now we describe the implementation of the stylistic grammars in a computational system. In this section, we will describe the organization of STYLISTIQUE, a syntactic stylistic parser that interprets the stylistic grammar. We will present a short example to illustrate STYLISTIQUE's method of analysis.</Paragraph>
    </Section>
    <Section position="16" start_page="488" end_page="490" type="sub_section">
      <SectionTitle>
5.1 How STYLISTIQUE works
</SectionTitle>
      <Paragraph position="0"> The English and French syntactic stylistic grammars are implemented in separate parsers that form the STYLISTIQUE system, a definite clause grammar that consists of  syntactic parse. The primitive-element analysis assigns connective and hierarchic classifications to each sentence component as it is added to the parse tree. The abstract-element analysis builds higher-level patterns from these connective and hierarchic primitive elements. The stylistic-goal analysis correlates these patterns of abstract elements with the author's possible stylistic goals. It chooses one goal from each of the following three dimensions: Clarity/neutrality/obscurity; Concreteness/neutrality/abstraction; Staticness / neutrality/dynamism.</Paragraph>
      <Paragraph position="1"> The translation from the grammar rules that were shown in Section 4 to the implementation was straightforward. In most cases, there was an easy way of mapping between the theoretical rules and the PROLOG rules. For example, heteropoisal postmodification is described in the grammar as postmodification that is parenthetical and either conjunct or antijunct, such as a nominal group or a nonfinite clause. The grammar rules and corresponding PROLOG rule for one alternative in the definition of postmodification that is both heteropoisal and concordant is shown in Figure 3. 48 Although STYLISTIQUE can usually run unaided, it does require punctuation in the input sentences to assist in disambiguation during parsing. The underlying stylistic grammar is highly complex, and so, lacking the aid of semantics, the stylistic parser can produce a number of syntactically grammatical, but not necessarily stylistically correct, parses for a given sentence. The standard PROLOG cut might have constrained backtracking and reduced the number of correct parses, but it seemed too crude and unpredictable to be easily applied to so complex and interdependent a system of rules 47 The English parser is based on a conventional parser written by Kern Luther and Rick MacLean at the University of Toronto, with supplementary rules suggested by Crystal and Davy (1969) and Quirk et al. (1985). The French parser is also based on Luther and MacLean's code, with supplementary rules adapted from Dubois and Dubois-Charlier (1970) and Galichet (1970). 48 The prefix &amp;quot;c&amp;quot; in the PROLOG rule indicates that the structure is concordant.  Sample grammar and PROLOG rules.</Paragraph>
      <Paragraph position="2"> as STYLISTIQUE. As a simpler solution, we introduced explicit markers of punctuation that provide a form of partial disambiguation, as they do what the system should have done unaided but couldn't, because of the limitations mentioned above. STYLISTIQUE requires this prior disambiguation mostly for reasons of efficiency. For example, most parsers are unable to handle the conjunction and by any means other than trying all possible parses (Snarr 1984). The problem is analogous for or and for the comma (which can be considered a word). In our parser, the situation is complicated by the same conjunctions and the same punctuation markers playing different roles according to their level in the sentence structure. As a consequence, different conjunctions and punctuation markers are used at the sentence, clause, complement, and noun phrase level.</Paragraph>
    </Section>
    <Section position="17" start_page="490" end_page="490" type="sub_section">
      <SectionTitle>
5.2 A Sample Stylistic Parse
</SectionTitle>
      <Paragraph position="0"> A corpus of 75 sentences (52 English, 23 French) was used to test the parser. These sentences were chosen to demonstrate a substantial degree of stylistic variation. The following short example illustrates the kind of analysis that STYLISTIQUE produces for the sentence: 45. True, posterity has been kind.</Paragraph>
      <Paragraph position="1"> In subsequent sections, we will give an interpretation of the following parse tree for this sentence. 49  cordant main clause, the major, with no subordinate clauses. It begins with a style disjunct, true, which is an elliptic adjectival and therefore considered to have a connective, concordant effect, even if used in the initial parenthetical position. After the initial disjunct, the sentence continues with the bare noun posterity, which, lacking both premodification and postmodification 5deg is a minimal, and therefore concordant, noun phrase. The sentence ends with the basic verb phrase has been kind, consisting of only the copula 51 been, and the concordant, conjunct 1 adjective kind; this is an inherently concordant verb phrase.</Paragraph>
      <Paragraph position="2"> The sentence is concordant from the hierarchic view as well, for it has the form of a concordant initial heteropoisal complete sentence. This indicates that the sentence begins with a parenthetical construction, which in this case is the disjunct, true, a superordinate adjectival. The bare noun posterity, lacking both premodification and postmodification, is a monoschematic noun phrase. The verb phrase has been kind is basic and therefore monoschematic.</Paragraph>
      <Paragraph position="3"> 50 STYLISTIQUE must assign a primitive classification to the postmodification so that this information can be passed up to the higher levels of the noun phrase. As a consequence, the absence of postmodification must be marked as (trivially) centroschematic postmodification, in the connective view, and (trivially) monoschematic postmodification, in the hierarchic view.</Paragraph>
      <Paragraph position="4"> 51 The stylistic effects of auxiliary verbs are not taken into account, and so these verb forms are not recorded in the stylistic parse.</Paragraph>
      <Paragraph position="5"> Chrysanne DiMarco and Graeme Hirst Goal-Directed Style in Syntax  ements are initial and medial concords. The significant dominance elements are centroschematic and monoschematic. That is, this sentence has one dominant shape, which is, in fact, the whole sentence: it is monoschematic but also trivially centroschematic. STYLISTIQUE produces the most detailed analysis it can and does not prune extraneous information. The single important balance element is an initial heteropoise.</Paragraph>
      <Paragraph position="6"> In the hierarchic view, STYLISTIQUE finds no significant position elements, but recognizes one dominance and one balance element: the sentence is centroschematic and an initial heteropoise. It is the initial disjunct, true, that introduces a superordinate effect; this feature makes the sentence slightly too complex to be monoschematic.  clarity/obscurity dimension. The presence of the concords in the connective view, together with the connective and hierarchic centroschematic structures, gives the sentence an effect of clarity. In a less obvious manner, the presence of an initial disjunct affects stylistic goals on the other dimensions. Because a superordinate parenthetical component is present, the sentence is a heteropoise and therefore considered to be concrete. On the staticness/dynamism dimension, the sentence has no definite leaning in either direction.</Paragraph>
      <Paragraph position="7"> To summarize, this is a simple clear sentence with the slight incongruity of an initial parenthesis to relieve its blandness.</Paragraph>
    </Section>
    <Section position="18" start_page="490" end_page="493" type="sub_section">
      <SectionTitle>
5.3 Limitations of the Implementation
</SectionTitle>
      <Paragraph position="0"> STYLISTIQUE'S limitations arise from the following characteristics of the system: * The implementation of the grammar is incomplete.</Paragraph>
      <Paragraph position="1"> * The grammar is too coarse-grained.</Paragraph>
      <Paragraph position="2"> In the first case, the incomplete implementation of the grammar, STYLISTIQUE can sometimes produce analyses that are inaccurate, though not incorrect, because, due to time constraints, only about 90% of the full English and French syntactic stylistic grammars was implemented.</Paragraph>
      <Paragraph position="3"> In the second case, the coarse grain of the stylistic grammar reflects theoretical limitations. The effect of coarse-grainedness was observed in the parses of some of the sentences in the sample corpus: Some sentences were parsed as simultaneously having both an initial concord and an initial discord. For example, both the following sentences were analyzed in this inconsistent manner: 46. To tell everybody is the best thing.</Paragraph>
      <Paragraph position="4"> 47. Telling lies is wrong.</Paragraph>
      <Paragraph position="5"> Some position elements were incorrectly identified: for example, a final discord was recorded as an initial discord.</Paragraph>
      <Paragraph position="6"> Some sentences with obviously different degrees of cohesiveness received the same analysis at the abstract-element level.</Paragraph>
      <Paragraph position="7"> These anomalous results are not due to programming errors, but to the excessive ab- null Computational Linguistics Volume 19, Number 3 stractness of the abstract elements and the resulting coarse grain of the grammar. It was a deliberate choice to define elements that were quite abstract. A major limitation of previous work in stylistics was the use of terms that were so specific and so numerous that it was not possible to identify general stylistic features that were common to sentences that were stylistically similar, but not obviously syntactically similar. However, our stylistic terms have occasionally erred on the side of being too abstract. For example, an initial discord in a STYLISTIQUE analysis may correctly identify the presence of a discord in the sentence but, because the scope of the element can be so broad, STYLISTIQUE sometimes cannot distinguish between a true initial discord, which occurs at the start of a sentence, and a noninitial discord that seems to be 'initial' because it affects the whole sentence.</Paragraph>
      <Paragraph position="8"> The excessive abstractness of the stylistic elements was further demonstrated by the rules in Section 4.4, where we used the same patterns of elements to define both clarity and staticness. That there may be different types of monoschematic sentences, some merely clear, others static as well, has been overlooked in a stylistic analysis that relies on maximally expressive descriptions. Our grammar currently does not allow for such subtleties; although it fulfills our objective of a formalization of style that captures many generalities, it is not yet sufficiently expressive to distinguish all the subtleties we would wish.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>