<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1031">
  <Title>Stylistic Grammars in Language Translation</Title>
  <Section position="3" start_page="0" end_page="148" type="metho">
    <SectionTitle>
2 A Definition of Style for
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="148" type="sub_section">
      <SectionTitle>
Machine Translation
</SectionTitle>
      <Paragraph position="0"> Our approach to a definition of style is non-literary, groupbased, and, most important, goal-directed. A &amp;quot;groupbased&amp;quot; approach emphasizes the stylistic standards shared by a body of writers, rather than the characteristics of an individual author. Examples of group styles can be found in newspaper reporting or scientific writing. By &amp;quot;goal-directed', we mean that we seek to explicitly correlate specific goals of style such as clarity, informality, or abstraction with particular syntactic structures and lexical choices.</Paragraph>
      <Paragraph position="1"> Why focus on group.based style? For MT, 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. However, group style can be subdivided into two major types, literary and functional (or pragmatic). Analysing literary group style, as in a work of fiction, for example, is too ambitious a task for MT. Functional group styles, on the other hand, are correlated with particular types of situations and are a more realistic objective.</Paragraph>
      <Paragraph position="2"> To fulfill both of the complementary aims, we require a view of style that is goal-directed. A recent example with this view of stylistics is the PAULINE system described in Hovy \[1987\]. PAULINE generates text that conforms to given pragmatic and stylistic constraints; the system is goal-directed, able to correlate such stylistic goals as formallty, simplicity, and respect with the characteristics of the text produced.</Paragraph>
      <Paragraph position="3"> In French-to-English translation, as an example, we must account for the fact that French is generally more abstract, static, and precise while English tends to be concrete, dynamic, and more lenient towards vagueness (Vinay and Darbelnet \[1958\]). For the same text, therefore, the French source language may express an abstract style while the English target language must be more c~ncrete i n its lexical and syntactic choices. If we are to properly translate from French style to English stylel then we mus~ understand how syntactic structures and' lextcal choices correlate with particular stylistic goals or aspects, such as abstraction, concreteness, statieness, and' dynamism.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="148" end_page="148" type="metho">
    <SectionTitle>
3 A Computational Schema of
</SectionTitle>
    <Paragraph position="0"> Stylistics in Translation We haw; restricted our interest in style for MT purposes to concerns ~hat are pragmatic, group-based, and goaldirected. The next question iq: how is this restricted definition to be reflected in a computationM schema of stylistics in machine translation? The schema that we propose is outlined in Figure 1. The process of stylistic translation is language--independent but l~'ench-to-English translation is the particular instantiation that we have studied in detail and from which we will give examples.</Paragraph>
    <Paragraph position="1"> In the first stage of the process, French internal stylistics is considered. For each stylistic translation unit, i.e., the current section oPS text within which the style rcmains constant, in the French source text, we determine its stylistic contempt, i.e., a correlation of lexicai choices and syntac.tic structures with a particular stylistic goal and a particular styli.'*tic intensity.</Paragraph>
    <Paragraph position="2"> In the second stage, we use knowledge of French-English comparative stylistics. For each stylistic translation unit in the French source text, given its stylistic con-text, we detor,nine the appropriate English stylistic goal correspohding to the possibly different French style. Default rules of comparative stylistics are uscd but these can be broken if the situation warrants.</Paragraph>
    <Paragraph position="3"> The third and final stage requires an understanding of English internal stylistics. For each stylistic translation unit in the unedited English translation, we edit lexicai choices mid syntactic structures to achieve an appropriate style that corresponds to the French source style, but is good English style as well.</Paragraph>
    <Paragraph position="4"> The main feature of the schema is that the translation of style, according to our definition, requires three distinct bodies ot! stylistic knowledge: laYench internal stylistics, French-:E;nglish comparative stylistics, and English internat stylistics. The internal stylistics of a language tells us how to choose syntactic structures and lexical items to express a particular aspect of style, such as clarity or economy. From comparative stylistics, we gain an understanding of when the target text should express a different style from the source text and when the styles should remain the same.</Paragraph>
    <Paragraph position="5"> The need for three kinds of stylistic knowledge may not initially appear obvious. Why not just edit the output of a MT system to remove awkward syntax and achieve a more natural flow of words? The answer is that to preserve the author's stylistic intent, while meeting the stylistic demands of the target l~aguage, we must determine the author's specific stylistic goals and consider how these different aspects can be realized syntactically and lexically in the tin'get language. Thus, we must analyse the original source text to make certain we understand the source language style before trying to translate.</Paragraph>
    <Paragraph position="6"> To implement the computational schema of stylistics in translation shown in Figure 1, knowledge about stylistics must be ~,.ccumulated, created, and organized into a formal representation.</Paragraph>
    <Paragraph position="7"> Our first task was the creation of a vocabulary of English stylistics. The definition of concepts and the attempt to orgaafize them into a recognizable structure were necessary to understanding the problem of translating style.</Paragraph>
    <Paragraph position="8"> The development of the vocabulary proceeded in tandem with the collection and creation of stylistic rules.</Paragraph>
    <Paragraph position="9"> A further complication in the codification of French, English~ and comparative stylistics is the need to build differentiNets of rules for lexieal choice, syntactic structure, and semantic structure. This knowledge goes to make up stylistic grammars -- formal representations of the stylistic rules. These grammars provide a systematic description of the lexicai, syntactic, and semmltic patterns that differentiate the various stylistic goals.</Paragraph>
    <Paragraph position="10"> With these format bodies of rules to analyse the original source and uncdited target texts, the strategy to be used by a French and English stylistic parser can be planned. Ultimately, M1 components will be integrated in a post-editor that will accomplish the actual translation of style.</Paragraph>
  </Section>
  <Section position="5" start_page="148" end_page="152" type="metho">
    <SectionTitle>
4 Stylistic Grammars
</SectionTitle>
    <Paragraph position="0"> As a first step towards constructing a grammar of style, we defined the basic vocabulm-y shown in Table 1.</Paragraph>
    <Paragraph position="1"> The naive approach towards structm'il~g an extended vocabulary of stylistics would be to organize terms by syn tactic category. For example, the concepts of detached adjccti'val clause, piled-up adje.ctives, and adjectival phr~,~e are all .significant in stylistics. However, not all terms'; associated with the same syntactic category necessarily shat'e the same stylistic significance.</Paragraph>
    <Paragraph position="2"> We have therefore taken a more stylistically relevant approach by introducing stylistic elements, which provide both a re.cans of structuring the vocabulary and a link bc.tween these vocalmlary terms and stylistic goals. Furthermore, a single vocabulary of stylistic elements will be used to describe lexical, syntactic, and semantic realizations.</Paragraph>
    <Paragraph position="3"> Style, we believe, is achieved through the effects inherent in individual components (absolute shape), the effects created by the absolute position of components (whether, for example, the component is initial, medial, or final), and by relationships (relative shape) between the eomponent~ of a stylistic constituent. These relationships fall into at least two basic types: balance and harmony, including the concordant and discordant elements in a constituent, and dominance, the number of central shapes in a constituent.</Paragraph>
    <Paragraph position="4"> Each type of relationship is expressed by stylistic elements. Table 2 shows samples. All the stylistic elements that we have defined, about twenty, have been recognized in our sample English data. Given these elements, we have a means of organizing the terms in our stylistic vocabulary: each term is associated with one or more stylistic elements of the w~rious types.</Paragraph>
    <Paragraph position="5">  i.e., having a particular stylistic effect.</Paragraph>
    <Paragraph position="6"> Stylistic texture: A stylistic constituent with an overall stylistic effect, formed by particular types of relationships between its individual shapes. Stylistic equivalence class: A set of stylistic constructions having the same type of stylistic shape. Interval: The stylistic distance or difference between any two shapes in a stylistic constituent. Modulation: A shift within a stylistic constituent from one type of stylistic effect, e.g., concord, to another, e.g., discord.</Paragraph>
    <Paragraph position="7">  A stylistic shape expressing a unity of style, agreement, accord, stability, and not requiring resolution.</Paragraph>
    <Paragraph position="8"> A modulation with a narrow interva~tween two shapes, each expressing consonance__ A stylistic shape expressing a disunity of style, disagreement, contention, conflict, incongruity, and requiring resolution. A modulation with a wide interval.</Paragraph>
    <Paragraph position="9"> A stylistic texture in which the initial and terminal shapes belong to the same stylistic equivalence class.</Paragraph>
    <Paragraph position="10"> A stylistic texture in which two or more successive shapes belong to the same equivalence class.</Paragraph>
    <Paragraph position="11"> A terminal modulation which moves from a stylistic discord to a relative concord. A terminal modulation which moves from a stylistic concord to a relative discord. A stylistic constituent having a stylistic shape but syntactically incomplete. A stylistic constituent with a single dominant shape and no accompanying subordinate shapes.</Paragraph>
    <Paragraph position="12"> A stylistic constituent in which the components are organized around a single dominant shape.</Paragraph>
    <Paragraph position="13"> A stylistic constituent organized around two or more dominant shapes.  Now, for our goal-directed stylistic grammar, we propose a branching stratifieationM modeh As the foundation, we have three branches: lexieal, syntactic, and semantic stylistic grammars, each with its own vocabulary of stylistic shapes and rules relating its type of shapes.</Paragraph>
    <Paragraph position="14"> At the central level, we use a single vocabulary of constituent elements of style, as in Table 2, to build rules relating these elements to patterns of lexical, syntactic, and semantic stylistic shapes. This level is the unifying core of the overM1 stylistic grammar for we believe that people also use the same principles, the same constituent elements, to express style both in English and French.</Paragraph>
    <Paragraph position="15"> Finally, at the top level, we construct rules to correlate individual stylistic goals with patterns of stylistic elements. Together, these levels form a language-independent goal-directed stylistic grammar for language translation.</Paragraph>
    <Section position="1" start_page="150" end_page="151" type="sub_section">
      <SectionTitle>
4.1 A Syntactic Stylistic Grammar
</SectionTitle>
      <Paragraph position="0"> For each of the branches in our goM-directed stylistic grammar, we must define a vocabulary of stylistic shapes and rules for putting these shapes together, ttowcver, a &amp;quot;stylistic shape&amp;quot; wil\[ be defined differently for tile lexical, syntactic, and semantic branches.</Paragraph>
      <Paragraph position="1"> For the ,;yntactic stylistic grammar, we have built a catalogue of sentence components (syntactic stylistic shapes) organized into equivalence classes by stylistic effect, a quality inherent in a component which makes it a stylistic shape. But what, in syntactic terms, gives each component a particular stylistic effect? How should we define &amp;quot;stylistic shape&amp;quot; at the syntactic level? We have adapted Quirk and Greenbamn's \[1979\] use of syntactic integration as the basis for our definition of syntactic stylistic shape. They catalogue most adverbials, some prepositional phrases, and some clauses as either adjunct, conjunct, or disjunct deI)ending on tile degree of inherent integration. We have chosen to base the syntactic stylistic effect of all sentence components on their charac-teristic integration. We have therefore expanded the catalogue to define a syntactic stylistic vocabulary classifying components as either adjunct (strongly integrating), conjunct (less integrating, but still having a connective effect), disjunct (neutral), or &amp;quot;antijunct&amp;quot; (having a disconnective effect).</Paragraph>
      <Paragraph position="2"> We then adopted Crystal and Davy's \[1969\] &amp;quot;stylistic&amp;quot; grammar as the foundation for our syntactic stylistic grammar. Their grammar is built upon a vocabulary of stylistically significant syntactic components. It also recognizes the role of pre- and post-modification in stylistic effects, a feature we have expanded in our grammar, a po&gt; tion of which is shown in Table 3.</Paragraph>
      <Paragraph position="3"> Now, having a base syntactic stylistic grammar which builds sentences out of syntactic stylistic shapes defined by their characteristic integrating effect,, we can c.onstruct, at the next level, a grammar which relates constituent eb ements of style to patterns of syntactic stylistic shapes.</Paragraph>
      <Paragraph position="4"> Syntactic exmnples for this level are given in Table 4.</Paragraph>
      <Paragraph position="6"> where x and y are instances of Dependent clauses and are in the same stylistic equivalence class.</Paragraph>
      <Paragraph position="7">  And, finally, at tile top level, we define rules which &lt;:or relate stylistic goals with patterns of constituent eh'ments. Syntactic clarity and syntactic c(mcreteness, for example, are correlated positively and negatively with tile stylistic elements shown in Table 5.</Paragraph>
      <Paragraph position="8"> Taken together, tables 3, 4 and 5 provide a portion of our goal-directed stylistic grammar for English internal stylistics.</Paragraph>
      <Paragraph position="9"> The use of this grammar can be illustrated by the analysis of the following two samt)le sentences: they arc diaschematic and thus express clarity. Tiffs first examph, illustrates imitative dependent clauses: This painter did his beat to adjust to the tastes of the day, softening his line and painting gracefully, converting to sfumato and casting about for art-lovers even as far away as in Spanish America \[Manchester Guardian, February 14, 1988, p.15\]. And, in this example, we see imitative postnmdification: Silvia, a commanding woman in her 50's, a shrew falsely mellowed by religion, promptly organiaed prayer sessions on the lines of Tupperware meetings \[adapted from Manch-</Paragraph>
    </Section>
    <Section position="2" start_page="151" end_page="152" type="sub_section">
      <SectionTitle>
Imitative Postmodification
</SectionTitle>
      <Paragraph position="0"> x, (Conjunction), y, (Conjunction), (z) where z, y, z are instances of Postmodification and are not instances of Antijunct Postmodification and arc in the same stylistic equivalence class.</Paragraph>
      <Paragraph position="1">  ester Guardian, February 7, 1988, p.16\]. Note that our stylistic grammar also gives patterns of stylistic elements which tend to oppose a certain stylistic goal. For example, an initial discord, a disjunct dependent \[non-finite\] clause, followed by a medial discord, an antijunct postmodification, a postposed adjective, tends to suppress clarity in the following .~entence: To illustrate the benefits of assertiveness, politicians, eloquent, capture our interest.</Paragraph>
      <Paragraph position="2"> In these next two sentences, we see an effect of syntactic concreteness, an attempt to express the immediacy of the world in the syntactic structure. The first sentence contains a medial counterpoise: Now demons, whatever else they may be \[nominal relative clause\], are full of interest. The second example contains a dissolution: And the rain descended \[simple concordant major\] and the floods came, and the winds blew, and beat upon the house; and it fell: \[concordant dependent clause due to the imitation\] and great was the fall of it \[discordant major due to the inversion of the subject and predicate\] \[Matthew, 7:27\].</Paragraph>
      <Paragraph position="3"> We can now begin to see how our vocabulary of constituent elements~ the central level of our stylistic grammar, will apply to the development of a semantic stylistic grammar. In the following two sentences, for example, we see seman~,ie counterpoise, the main idea interrupted by a secondary one: The university attended by the Prime Minister, one of the finest law schools in the country, is the alma mater of many prominent politicians. Compare this with: The university attended by the Prime Minister, a set of buildings with the architectural charm of a prison, is the alma mater of many prominent politicians. In the first case, the offset supports the main idea so that there is an overall effect of stylistic concord but, in the second example, the intruding phrase is semantically and stylistically discordant, In more subtle ways, we can correlate stylistic elements with patterns of semantic structure. In the following sentences, note the difference in order of components between 11 a regardg dans le jardin par la porte ouverte and its English translation, He gazed out of the open door into the garden. In French, the result comes first, then the means, while English follows the order of the images, like a film of the action \[Vinay and Darbelnet 1958\]. We would say that that are three types of semantic stylistic shapes composing the sentence: he gazed, out of the open door, and into the garden. In the English sentence, the modulations between shapes seem small, so that there is a gradual unfolding of images, an effect of concreteness, while the French structure contains a pattern of modulations characteristic of abstraction.</Paragraph>
      <Paragraph position="4"> So far, we have put forward stylistic grammars as the formal representation of English and French internal stylistics. This approach has led us to organize a stylistic vocabulary describing sentence components in terms of stylistic shapes and then to define rules relating stylistic elements to patterns of these shapes. The next step, associating stylistic goMs with stylistic elements allows us, in effect, to correlate syntactic structures with these goals.</Paragraph>
      <Paragraph position="5"> With an explicit correlation between syntactic grammar rules, stylistic grammar rules, and stylistic goals, we can now propose a strategy :for the English and French stylistic parser that we will construct. The syntactic rules do the driving but partial stylistic decisions can affect the direction of the syntactic parse. That is, once a stylistic rule has been partially recognized, we can narrow the search space for constructs to complete the rule. As the syntactic rules analyse sequences of syntactic structures, these structures can then be recognized as sequences of stylistic elements, the components of stylistic rules. In tandem, the stylistic rules will pick up a sequence of stylistic elements mid associate it with a stylistic goal. The idea of tandem rules in language analysis is developed by Hirst \[1987\]. This approach to integrating syntax and semantics has been adapted to syntax and stylistics.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>