<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0316">
  <Title>Building Another Bridge over the Generation Gap</Title>
  <Section position="3" start_page="139" end_page="141" type="metho">
    <SectionTitle>
SHOOTING .4
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="139" end_page="139" type="sub_section">
      <SectionTitle>
2.3 Organization of Lexical Resources
</SectionTitle>
      <Paragraph position="0"> as a Multilayered Model As stated above, we organize situation-specific lexicat resources that are provided by the LF information into four layers. The layers are called Intersituational Layer, Intrasituational Layer, Constituent Layer, and Lexeme Layer, respectively. The first three layers are motivated by the semantic constituency of expressions as provided by LFS: 1. two predicate structures (as given by LF sequences) are dealt with at the intersituational layer; 2. predicate structures (as given, e.g., by the Operx function) -- at the intrasituational layer; and 3. predicate constituent structures (as given, e.g., by the St function) -- at the constituent layer. At the lexeme layer, the LFS are unified with concrete lexemes, s LFs correspond to knowledge at varying levels of abstraction (see the discussion in \[Wanner and Bateman, 1990\]); therefore, we organize them coherently at each of the first three layers in network form in terms of the functional, semantic, and syntactic features that they represent (cf. also \[Wanner, 1992\]). For example, the Operl function is decomposed as {actor.realization, situation-oriented, actor-salience, ...}. Within this organization, the most general structures provide the representation of lexical semantics and the most deli4Evidently, semantically related situations have intersecting lexicalization clusters. In order to achieve an efficient organization of the overall lexical resources, we use inheritance techniques, cf. \[Wanner, 1992\].</Paragraph>
      <Paragraph position="1"> SObviously, our layers are in analogy to the &amp;quot;ranks&amp;quot; in systemic functional grammars, cf. \[Halliday, 1985\].</Paragraph>
      <Paragraph position="2"> cate ones -- lexicalization.</Paragraph>
      <Paragraph position="3"> A PGS is then built up b\] the lexical choice process by a successive traversal of the layers starting from the intersituational layer. How this is done is described in detail in \[Wanner, 1992\]. Here we focus on the organization of the layers and on the interface between the output of the text planning module and lexical resources.</Paragraph>
      <Paragraph position="4"> In what follows, each layer is described briefly (starting with the lexeme layer). Note that this description is necessarily extremely oversimplified; it largely abstracts from the actual representation in a systemic functional framework.</Paragraph>
      <Paragraph position="5"> Lexeme Layer. The lexeme layer contains LF - LF value pairs. Consider selected entries for Operl and $1 in the lexicalization cluster of SHOOTING: L.o~: + J Constituent Layer. The constituent layer contains feature specifications for the realization of predicate constituents and their attributes, s For example, to name the ACTOR in SHOOTING, his/her proper name; a situation non-specific realization, e.g., his/her function in relation to other entities in the knowledge base (e.g., grandfather); the situation-specific and contextually neutral lexeme marksman; or a negatively loaded situation-related expression such as murderer, killer, etc. can be chosen. Consider a sample fragment for the realization of the ACTOR in SHOOTING in network form: actor- - situation-spectr~- 1 ne 1o&amp;quot; g.o ! mg -Io * . .o * The labels in Times Roman are network features in the corresponding fragment; those in italics are pointers to &amp;quot;external features&amp;quot; that provide a more detailed specification of the feature they are associated with.</Paragraph>
      <Paragraph position="6"> &amp;quot;External&amp;quot; means 'at other layers of the same cluster' or 'in clusters of other situations'. Clusters of other situations are identified by labels in small capitals (preceded by an uparrow). The labels in bold denote roles which are introduced into the PGS when  the corresponding feature is chosen during the network traversal.</Paragraph>
      <Paragraph position="7"> The fragment shows that there are pointers to features in. other lexicalization clusters and to features at lower layers of the same lexicalization cluster (of. the pointer to S1 at the lexeme layer).</Paragraph>
      <Paragraph position="8"> Intrasituational Layer. The intrasituational layer contains feature specifications for the realization of predicates and their attributes. For example, to express ACTOR's performance of SHOOTING, e.g., V0 ( \[tO\] shoot), Operl (\[to\] take a shot, \[to\] fire a shot), etc. are available; but not, e.g., Oper2 (*\[to\] get a shot as in * The burglar got a shot from grandpa). In our functional interpretation, the difference between \[to\] shoot and the other options is that it is 'actionoriented', while the others are 'situation-oriented'. If \[to\] get a shot were grammatical, it would emphasize 'actee-salience' (as \[to\] be shot does), while the other options mentioned emphasize 'actor-salience'. In network form, this looks as follows (see \[Wanner and Bateman, 1990\] for a detailed discussion): axis: the 'causation' and the 'initial' phases precede the 'continuing' and the 'final' phases; the 'continuing' phase precedes the 'final' phase, etc. Further, the phases 'initial' and 'final' are related to SHOOTING itself. null As global, for example, the different PREPARATION actions (PREPARATION / (verbalized, e.g., by \[to\] charge the gun) and PREPARATION// (verbalized, e.g., by \[to\] cock the trigger) precede SHOOTING; ACTOR's and ACTEE'S actions that follow SHOOTING as well as the 'resulting state' of ACTEE succeed it. In the following fragment, only the 'global' organization is partially illustrated: [  - gL tazce, aag...</Paragraph>
      <Paragraph position="9"> - gl. succ.eeainl</Paragraph>
      <Paragraph position="11"> Note the difference between PGS (i.e. grammatical) roles and situation roles: the feature set { . .. actor's performance, situation-oriented, actor-salience} triggers the realization of the label of the situation (So) in the grammar as ACTEE; and the situational role ACTEE -- aS GOAL.</Paragraph>
    </Section>
    <Section position="2" start_page="139" end_page="139" type="sub_section">
      <SectionTitle>
3 Using Lexical Resources for
Discourse Organization
</SectionTitle>
      <Paragraph position="0"> To tailor RST relations in the text plan to lexical resources, the lexical choice process must know (i) how, e.g., RST relations are specified in more detail by lexical resources and (it) what the possible linguistic realizations of these finer specifications are. Since LFs have well-defined semantic, lexical and syntactic realizations, (it) is provided by our representation. In this section, we address (i); and more precisely: * how discourse structure relations are defined in lexical resources, Intersituatlonal Layer. The intersituational layer contains feature specifications for the realization of relations between predicates. For example, of the CAU-SATION relation between ACTOR's performance of the situation SHOOTING (verbalized, e.g., by \[to\] shoot or by \[to\] take a shot) and the state of ACTEE after the situation (\[to\] be dead); of the CONDITION relation between ACTOR's preparing the INSTRUMENT for functioning and ACTOR's performing of the situation (as, e.g., instantiated by \[to\] charge and \[to \] shoot; etc. In the following fragment, we present the specification of the TEMPORAL SEQUENCE relation taking SHOOTING as reference time. We represent the 'internal' temporal and the 'global' temporal organization of SHOOTING.</Paragraph>
      <Paragraph position="1"> As 'internal', we present its phases at the temporal * how these relations are linked up with the relations specified in the text plan.</Paragraph>
    </Section>
    <Section position="3" start_page="139" end_page="141" type="sub_section">
      <SectionTitle>
3.1 Towards Lexical Discourse Struc-
ture Relations
</SectionTitle>
      <Paragraph position="0"> For the definition of discourse structure relations in lexical resources (henceforth lexical discourse structure relations), we use LF sequences. As stated above, LF sequences (and LFs in general) are organized in terms of their semantic, functional, and syntactic features. null The expression of the same content via different syntactic patterns is an important task in sentence planning (eft, e.g., \[Meteer, 1992, Vander Linden et  al., 1992, Scott and de Souza, 1991\]. 7 But it is the functional content we associate with each LF sequence, which links t_he lexical resources up to global discourse structure relations.</Paragraph>
      <Paragraph position="1"> Table 1 shows the functional content we associate with some LF sequences, s For a detailed discussion of how the functional content of LF sequences is used to define and to structure lexical discourse relations following Halliday's organization of intrasentential logico-semantic relations (cf. \[Halliday, 1985\]), see \[Wanner, 1994\].</Paragraph>
      <Paragraph position="2">  . Caus2Operl A (Magn o) SynV0 (as in She commands respect -- I (really) think highly of her); 2. V0 ^ because of o SUBJECT (as in I respect her because off her work); 3. CaussOperx (as in Her work makes me respect  her).</Paragraph>
      <Paragraph position="3"> The functional content of these realizations is more detailed than that of VOLITIONAL CAUSE. For example: 1. communicates an active role of the CAUSER; 2. -- the active role of the ACTOR, and 3. -- the active role of the SUBJECT. Therefore, they not only realize VOLITIONAL CAUSE, but also specify it in more detail. To summarize, the search for a lexical discourse structure relation is done in accordance with the functional content, the communicative intention of the speaker, and the contents of the arguments of the RST relation considered. If the RST relation connects unrelated case frames 1deg (as, e.g., EVIDENCE in In winter, the days are short. It is getting light late and early dark) these case frames are realized independently without being connected by a lexical discourse structure relation. If the case frames are related, the following three variations are possible: (i) An RST relation coincides, in general, with the functional content of a lexical discourse structure relation; as, e.g., VOLITIONAL CAUSE in the following rudimentary text plan for the 'RESPECT' examples given above does:</Paragraph>
    </Section>
    <Section position="4" start_page="141" end_page="141" type="sub_section">
      <SectionTitle>
3.2 Establishing the Initial Link
</SectionTitle>
      <Paragraph position="0"> The first and the most important task in tailoring the text plan to linguistic resources is to find lexical discourse structure relations that correspond to RST relations specified in the text plan.</Paragraph>
      <Paragraph position="1"> To illustrate this task, let us consider the RST relation VOLITIONAL CAUSE as it holds between the situation's causation and the situation itself.</Paragraph>
      <Paragraph position="2"> The corresponding function in our model is causal enhancement. Apart from the LF sequence Caus A V0 given in the table above (in the cluster of RESPECT, instantiated, e.g., by An old tradition requires children 4o show respect for their parents), at least the following three also function as causal enhancement: 9 rClosely related to this topic is the task of 'compacting' the information to be corv_mtmJcated' (sometimes called 'aggregatlon' \[Hovy, 1993\]).</Paragraph>
      <Paragraph position="3"> STl~s is not to say that these functions are the only ones that are possible.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="141" end_page="142" type="metho">
    <SectionTitle>
9 The Syn function provides a synonym expression; the Caus
</SectionTitle>
    <Paragraph position="0"> indices '2' and '3' stand for 'causation by ACTEE' and 'causation by the SUBJECT', respectively.</Paragraph>
    <Paragraph position="1">  If so, the subclassification of the lexical relation determines its final realization.</Paragraph>
    <Paragraph position="2"> (it) An RST relation instantiation subsumes several distinct classes of lexical discourse structure relations; as, e.g., the instantiation of the RST relation CON-TRAST in the text plan below (this plan is also highly simplified):  This text plan may be realized either as contrastive claxification: Gaul is entirely occupied by the Romans; well, not entirely.., one small village still holds out.; or as contrastive enhancement: Gaul is almost entirely occupied by the Romans; but one small village still holds out, In this case, a sufficiently general lexical discourse structure relation which subsumes both contrastive clarification and contrastive enhancement is chosen.</Paragraph>
    <Paragraph position="3"> (iii) An RUT relation is not captured by our taxonomy (as, e.g., CONCESSION). Then, the corresponding case frames are treated as unrelated (see above).</Paragraph>
  </Section>
class="xml-element"></Paper>