<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1410">
  <Title>Reinterpretation of an existing*NLG system in a Generic Generation Architecture</Title>
  <Section position="5" start_page="69" end_page="69" type="metho">
    <SectionTitle>
Abstract Semantic Abstract semantic representa-
</SectionTitle>
    <Paragraph position="0"> tions are the first level at which semantic predicates are associated with arguments. At this level, semantic predicates and roles are those used in the API to query the knowledge base and arguments are knowledge base entities.</Paragraph>
    <Paragraph position="1"> Semantic (Concrete) semantic representations provide a complete notation for &amp;quot;logical forms&amp;quot; where there is no longer any reference to ,the knowledge base. The representations are based on systems such as SPL (Kasper, 1989) and DRT (Kamp and Reyle, 1993).</Paragraph>
    <Paragraph position="2"> 4More details can be found in (Cahill et al., 1999) and at the RAGS project web site: http ://www. itri . brighton, ac. uk/rags.</Paragraph>
    <Paragraph position="3"> and satellite or changing the rhetorical relation to one within a permitted set.</Paragraph>
  </Section>
  <Section position="6" start_page="69" end_page="69" type="metho">
    <SectionTitle>
Abstract Document Document structure defines
</SectionTitle>
    <Paragraph position="0"> the linear ordering of the constituents of the Rhetorical Representation with a POSITION feature, as well as two other features, TEXT-LEVEL, which takes values such as paragraph or sentence; and LAYOUT, which takes values such as wrapped-text and vertical list. It takes the form of a tree, usually, but not necessarily, isomorphic to the Rhetorical Representation and linked to it, but with these three features at the nodes instead of rhetorical relations. null</Paragraph>
  </Section>
  <Section position="7" start_page="69" end_page="69" type="metho">
    <SectionTitle>
Abstract Syntactic Abstract Syntactic Represen-
</SectionTitle>
    <Paragraph position="0"> tations capture high-level aspects of syntactic structure in terms of notions such as lexical head, specifiers, modifiers and complements. This level of representation is compatible with approaches such as LFG f-structure, HPSG and Meteer's Text Structure.</Paragraph>
  </Section>
  <Section position="8" start_page="69" end_page="70" type="metho">
    <SectionTitle>
3 Partial and Mixed Representations
</SectionTitle>
    <Paragraph position="0"> For all of the RAGS levels partial representations are possible. Without this, it is not possible for a module to pass any result to another until that result is completely determined, and this would impose an unwanted bias towards simple pipeline architectures into the model. There are many cases in NLG where a representation is built collaboratively by several modules. For instance, many systems have a referring expression generation module whose task is to complete a semantic representation which lacks those structures which will be realised as NPs. Such a functionality cannot be described unless partially complete semantic representations can be communicated.</Paragraph>
    <Paragraph position="1"> In addition, mixed representations are possible, where (possibly partial) representations at several levels are combined with explicit links between the elements. Many NLG modules have to be sensi- null tive to a number of levels at once (consider, for .......... instance, -aggregatiomxeferring,expmssion.,generation and lexicalisation, all of which need to take into account rhetorical, semantic and syntactic constraints). The input to most reusable realisation systems is also best viewed as a mixture of semantic and abstract syntactic information.</Paragraph>
    <Paragraph position="2"> The extra flexibility of having partial and mixed representations turned out to be vital in the reconstruction of the CGS system. (Mellish et al., 2000).</Paragraph>
  </Section>
  <Section position="9" start_page="70" end_page="71" type="metho">
    <SectionTitle>
4 The CGS system
</SectionTitle>
    <Paragraph position="0"> The Caption Generation System (CGS) generates explanatory captions of graphical presentations (2-D charts and graphs). Its architecture is a pipeline with several modules, shown in the left hand part of Figure 1. An example of a diagram and its accompanying text are given in Figure 2. The propositions are numbered for ease of reference throughout the paper.</Paragraph>
    <Paragraph position="1"> The input to CGS is a picture representation (graphical elements and its mapping from the data set) generated by SAGE plus its complexity metric.</Paragraph>
    <Paragraph position="2"> The text planning module (Moore and Paris (1993)) plans an explanation in terms of high level discourse goals. The output of the planner is a partially ordered plan with speech-acts as leaves.</Paragraph>
    <Paragraph position="3"> The ordering module receives as input the discourse plan with links specifying the ordering relations between sub-trees and specifies an order for them based on heuristics such as that the description should be done from left to right in the visual space. The aggregation module &amp;quot;only conjoins pairs of contiguous propositions about the same grapheme type 5 in the same space&amp;quot; (Mittai et al., 1999) and inserts cue phrases compatible with the propositions e o ( .=., &amp;quot;whereas&amp;quot; for contrastive ones). The internal order of the sentence constituents is determined by the centering module using an extension of the centering theory of Grosz and colleagues (Grosz et al., 1995).</Paragraph>
    <Paragraph position="4"> The referring expression module uses Date and Reiter's (Dale and Reiter, 1995) algorithm to construct the set of attributes that can uniquely identify a referent. There are'two, situations where the text planning module helps specifically in the generation of referring expressions: (1) when the complexity for expressing a graphic demands an example and 5&amp;quot;Graphemes are the basic building blocks for constructing pictures. Marks, text, lines and bars are some of the different grapheme classes available in SAGE.&amp;quot; (IVlittal et al., 1999).</Paragraph>
    <Paragraph position="6"> labels for the RAGS representations refer to the following: I = conceptual; II = semantic; III = rhetorical; IV = document; V = syntactic.</Paragraph>
    <Paragraph position="7"> it signals this both to SAGE (for highlighting the corresponding grapheme) and to the rest of the text generation modules; and (2) when in a specific situation the referring algorithm would need several interactions for detecting that an entity is unique in * a certain visual space and.the planning could detect it in the construction of the description of this space.</Paragraph>
    <Paragraph position="8"> When this occurs, the text planner &amp;quot;circumvents the problem for the:.referring ,expression :module at the planning stage itself, processing the speech-acts appropriately to avoid this situation completely&amp;quot;.</Paragraph>
    <Paragraph position="9"> After lexicalisation, which adds lexeme and major category information, the resulting functional descriptions are passed to the FUF/SURGE realiser that generates texts like the caption of Figure 2.</Paragraph>
    <Paragraph position="11"> charts, the y-axis indicates the houses. (7) In the first chart, the left edge of the bar shows the house's selling price whereas (8) the right edge shows the asking price. (3) The horizontal position of the mark shows the agency estimate. (4) The color shows the neighbourhood and (5) shape shows the listing agency. (6) Size shows the number of rooms. (9) The second chart shows the number of days on the market.</Paragraph>
  </Section>
  <Section position="10" start_page="71" end_page="74" type="metho">
    <SectionTitle>
5 Reinterpretation of CGS in RAGS
</SectionTitle>
    <Paragraph position="0"> Our reinterpretation of the CGS system defines the interfaces between the modules of CGS in terms of the RAGS data structures discussed above. In this section we discuss the input and output interfaces for each CGS module in turn as well as any problems we encountered in mapping the structures into RAGS structures. Figure 1 shows the incremental build-up of the RAGS data levels across the pipeline. Here we have collapsed the Abstract Rhetorical and Rhetorical and the Abstract Semantic and Semantic. It is-interesting to note that the build up of levels of representation does not tend to correspond exactly with module boundaries.</Paragraph>
    <Paragraph position="1"> One of the major issues we faced in' our reinterpretation was where to produce representations (or partial representations) whose emergence was not defined clearly in the descriptions of CGS. For instance, many decisions about document structure are made only implicitly by the system. In most cases we have opted to produce all types of representations at the earliest point where they can conceivably have any content. This means, for instance, that our reimplementation assumes an (unimplemented) text planner which produces an Abstract Rhetorical Representation with Abstract Semantic leaves and an Abstract Document Representation.</Paragraph>
    <Paragraph position="2"> Text Planner The input to the Longbow text planner discussed in section 4 above is a representation of a picture in SAGE format (which has been annotated to indicate the types of complexity of each grapheme) together with a goal, which can typically be interpreted as &amp;quot;describe&amp;quot;. It outputs an essentially fiat sequence of plan operators, each of which corresponds in the output* text .to .a.speech act. In our reinterpretation, we have assumed that this fiat structure needs to be translated into an Abstract Rhetorical Representation with (at least) minimal structure. Such a structure is implicit in the plan steps, and our interpretation of the rhetorical structure for the example text corresponds closely to that of the post-processing trace produced by CGS.</Paragraph>
    <Paragraph position="3">  . . .Z., However, we are still not entirely sure exactly CGS creates this structure, so posed it at the very beginning, onto the text planner.</Paragraph>
    <Paragraph position="4"> Already at this stage it is necessary about where we have imoutput of the to make use of mixed RAGS representations. As well as this Abstract Rhetorical Representation, the text planner has to produce an Abstract Document Representation, linked to the Abstract Rhetorical Representation. This is already partially ordered - although the exact value of POSITION features cannot be specified at this stage, the document tree is constructed so that propositions are already grouped together. In addition, we make explicit certain default information that the CGS leaves implicit at this stage, namely, that the LAYOUT feature is always wrapped text and that the TEXT-LEVEL feature of the top node is always paragraph.</Paragraph>
    <Paragraph position="5"> Ordering The ordering module takes the Abstract Document Representation and the Abstract Rhetorical Representation as input and outputs an Abstract Document Representation with the POSITION feature's value filled,for all :the nodes, .That is, it fixes. * the linear order of the final output of the speech acts. In our example, the ordering is changed so that steps 7 and 8 are promoted to appear before 3, 4, 5 and 6. The resulting structure is shown in figure 36 .</Paragraph>
    <Paragraph position="6"> 6In this and the.following diagrams, objects are represented by circles with (labelled) arrows indicating the relations be--Aggregation Although aggregation might seem like a self-contained process within NLG, in practice it can make changes at a number of levels of representation and indeed it may be the last operation that has an effect on several levels. The aggregation module in our reinterpretation thus has the final responsibility to convert an Abstract Rhetorical Representation with Abstract Semantic Representation leaves into a Rhetorical Representation with Semantic Representation leaves. The new Rhetorical Representation may be different from before as a result of speech acts being aggregated but whether different or not, it can now be considered final as it will no longer be changed by the system. The resulting Semantic Representations are no longer Abstract because further structure may have been determined for arguments to predicates. On the other hand, referring expressions have not yet been generated and so the (Concrete) Semantic Representations cannot be complete. The reconstruc,.tion createspartia.i Semantic Representations with &amp;quot;holes&amp;quot; where the referring expressions (Semantic Representations) will be inserted. These &amp;quot;holes&amp;quot; are linked back to the knowledge base entities tfiat they correspond to.</Paragraph>
    <Paragraph position="7"> Because Aggregation affects text levels, it also affects the Abstract Document Representation, which has its TEXT-LEVEL feature's values all filled at this tween them. Dashed arrows indicate links between different levels of representation.</Paragraph>
    <Paragraph position="8">  point. It may also need to change the structure of the Abstract Document Representation, for instance, adding in a node for a sentence above two, now aggregated, clause nodes.</Paragraph>
    <Paragraph position="9"> Centering Because Centering comes before Referring Expression generation and Realisation, all it can do is establish constraints that must be heeded by the later modules. At one stage, it seemed as if this required communicating a kind of information that was not covered by the RAGS datatypes. However, the fact that an NP corresponds (or not) to a center of some kind can be regarded as a kind of abstract syntactic information. The reconstruction therefore has the centering module building a partial (unconnected) Abstract Syntactic representation for each Semantic Representation that will be realised as an NP, inserting a feature that specifies whether it constitutes a forward- or backward-facing center, approximately following Grosz et al (Grosz et al., 1995). This information is used to determine whether active or passive voice will be used. An example of such a partial Abstract Syntactic Representation is given in Figure 4.</Paragraph>
    <Paragraph position="10"> Referring Expression In our reconstruction of the CGS system, we have deviated from reproducing the exact functionality for the referring expression module and part of the lexical choice module. In the CGS system, the referring expression module computes association lists which can be used by the lexical choice module to construct referring expressions suitable for realisation. In our reconstruction, however, the referring expression module directly computes the Semantic Representations of referring expressions.</Paragraph>
    <Paragraph position="11"> We believe that this is a good example of a case where developing a system with the RAGS data structures in mind simplifies the task. There are undoubtedly many different ways in which the same results could be achieved, and there are many (linguistic, engineering etc.) reasons for choosing one rather than another. Our particular choice is driven by the desire for conceptual simplicity, rather than any strictly linguistic or computational motivations. We considered for each module which RAGS level(s) it contributed to and then implemented it to manipulate that (or those) level(s). In this case, that meant a much more conceptually simple module which just adds information to the Semantic Representations. null Lexical Choice In CGS, this module performs a range of tasks, including what we might call the later.stages of_referring expression generation and lexical choice, before converting the plan leaves into FDs (Functional Descriptions), which serve as the input to the FUF/SURGE module. In the reconstruction, on the other hand, referring expressions have already been computed and the Rhetorical Representation, with its now complete Semantic Representations, needs to be &amp;quot;lexicalised&amp;quot; and  tion in our terms involves adding the lexeme and major category information to the Abstract Syntactic Representations for the semantic predicates in each Semantic Representation. The FUF/SURGE input format was regarded as a combination of Semantic and Abstract Syntactic information, and this can easily be produced from the RAGS representations. The combined Semantic and Abstract Syntactic Representations for the plan step &amp;quot;These two charts present information about house sales from data set ts-1740&amp;quot; is shown in Figure 5. The boxes indicate suppressed subgraphs of the lexemes corresponding to the word in the boxes and triangles indicate suppressed subgraphs of the two adjuncts.</Paragraph>
  </Section>
class="xml-element"></Paper>