<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1405">
  <Title>Controlled Realization of Complex Objects by Reversing the Output of a Parser</Title>
  <Section position="3" start_page="39" end_page="43" type="intro">
    <SectionTitle>
2. Architecture
</SectionTitle>
    <Paragraph position="0"> The overall design of the generation architecture used here is as described in Meteer 1992, following a set of principles laid out in McDonald, Pustejovsky, and Meteer 1988. It is a message-driven system that begins with Structures taken directly from the underlying system for which it is speaking, realizing them monotonically via a succession of progressively more linguistic represen-tational .levels through the use of an *extensive set of abstract linguistic resources ultimately grounded in a TAG grammar of English.</Paragraph>
    <Paragraph position="1"> &amp;quot; The source structures are represented in a Kl-one derived system called Krisp (McDonald 1994s), and the parser that produces the corpus-based lattice of realization types is Sparser (McDonald 1992, 1994a)mtwo complete, mature systems. We will introduce only as much information about them as necessary to support the rest of the discussion. Much of this paper will be devoted to an extension to Krisp, a &amp;quot;saturation lattice', that is the basis of this technique of micro-planning by reversing the parser's output.</Paragraph>
    <Paragraph position="2"> At present the new generator, to be christened ,Magellan&amp;quot;, implements only the micro-planning and later stages of generation. There is no speaker in a situation with genuine motivations, without which a authors might shed some light on the problem. If pressed, I would make a macro-planner from reverse engineered schemas with a randQm element.</Paragraph>
    <Paragraph position="3"> 3 I will use the term 'parse' and 'parser' as a convenient * short-hand for designating the full natural language understanding system that is actually being used. This does not Stretch the usual senses of the term too much since Sparser does do its semantic interpretation at literally the same time as its parsing into syntactic phrases. The key difference from the usual parser is that the end result is a set of objects in a domain model rather than just a parse tree.</Paragraph>
    <Paragraph position="5"> generation system (or certainly a fully-articulated theory of generation) is incom-plete. In its stead, as a way to exercise the micro-planner, is a windup toy--a largely graphical interface that permits the experimenter to stipulate the inpu t and decision criteria that a macro-planner would have produced and see what text results.</Paragraph>
    <Paragraph position="6"> The generation proces s starts with the Krisp units representing a full relation. We select by hand the fragment of it to be expressed and some simple information-structure parameters. Then the micro-planning mechanism described here is deployed to populate a Text Structure representation, which has been excerpted directly from Meteer's Spokesman system (1992). Spokesman's mechanisms then read out the Text Structure to create the TAG derivation tree that is the input to Mumble (Meteer et al. !987), which in turn produces a TAG-based surface structure and from that the eventual text. ~</Paragraph>
    <Section position="1" start_page="39" end_page="41" type="sub_section">
      <SectionTitle>
2.1 Categories in the domain model
</SectionTitle>
      <Paragraph position="0"> The micro-planner's task is to realize a single, highly-structured, compositional *relation as a Text Structure. To illustrate theresources it uses to do this, consider these two (made up) sentences:.</Paragraph>
      <Paragraph position="1">  (2) &amp;quot;GTE owns BBN.&amp;quot; (3) &amp;quot;BBN is a Subsidiary of GTE.&amp;quot;  These express the same information. They should be represented by the same object in the domain model. Which of the two alternative realizations of this object a speaker will choose is a question of which of the two companies they decide to make the theme.</Paragraph>
      <Paragraph position="2"> The expression below defines the type, or 'semantic category', that these texts instantiate. This is a Lisp expression that is evaluated at the time the domain model is loaded. It makes reference to several things that will already have been defined, notably thecategory's parent super-category in the taxonomic lattice, named 'owner-owns-owned', 4 and two tree families in the grammar. Once the expression has been executed, theresult is (1) a category object representing the type in the domain model; (2) a setof phrase structure rules in the semantic grammar used bythe parser (based on the 'realization' field); and  (3) a &amp;quot;saturation lattice' (based on the 'binds' field) which is described below. (The realization field is only sketched here since it's values would make little sense without the background that will be supplied 4 There is an unfortunate tendency for names like these to dominate how a person thinks aboutconcepts. Simply  because names are necessarily (if they are to be useful) comprised of regular, suggestive natural language words, they can too often cloud the mind to the possibility that there are many different ways to realize the same conceptual content (see, e.g., Elhadad et al. 1996) This is something always to be guarded against. The choice of names for all the objects in this domain model and grammar is arbitrary and strictly for the convenience of the human designer. Because they are implemented in terms of objects and pointers, the names are not even used at runtime, and serve only to provide a way to uniquely designate the objects in the written expressions that are needed to initially define them when a direct manipulation interface is not being used  This category corresponds roughly to a KLOne concept with two slots named 'parent' and 'subsidiary', whose values are restricted to objects of type (category) company. The &amp;quot;specializes' field indicates how this category is knit into the taxonomic lattice and the inheritance path of its slots, which are termed 'variables' in Krisp.</Paragraph>
    </Section>
    <Section position="2" start_page="41" end_page="43" type="sub_section">
      <SectionTitle>
2.2 Saturation lattices
</SectionTitle>
      <Paragraph position="0"> The reference model for Krisp is the Lambda Calculus, where there is a well articulated notion of expressions that have only bound a few of their variables to specific values and left the others open. Such 'partially saturated relations' have a first class representation in Krisp. This representation is supported by a lattice 'below' each of the categories of the normal taxonomic (is-a) lattice. This lattice defines types for all of the category's sets of possible partial instantiations and provides a representational anchor for the realization annotations that the parser lays down and the micro-planner uses.</Paragraph>
      <Paragraph position="1"> As shown below, a Saturation latticeconsists of a linked set of nodes that represent all the possible combinations of bound and open variables of the category, including a pseudo-variable &amp;quot;self that allows us to include the category itself. (This variable is usually realized as the verb of a clause or the head noun of a np.) Notice the use of single-letter abbreviations for the variables when they appear i n multi-variable nodes.</Paragraph>
      <Paragraph position="2"> self ('s') p reat (tip) ' s+ p+b s aidiaby(' ')b  In the present example, the lattice is relatively simple with just three levels, s At the top we have the information states where one of the three variables is bound and the other two open. Next these nodes ('lattice points') converge to form a level where each possible combination of two bound and &amp;quot;one open variable is represented. These then join to form the lattice point that represents the state where the relation is fully saturated, i.e. all of its variables are bound. This bottom node in the lattice is annotated by the various contexts in this category has appeared as a contiguous phrase: as a whole sentence, as a subordinated clause, as a reduced phrase in a conjunct, etc. The abstract resources for these contexts correspond to attachment points in Mumble and usually adjunctions in a TAG.</Paragraph>
      <Paragraph position="3"> The saturation lattice is used in the parsing direction to provide a set of indexes that anchor and organize partial results so that phrases with the same denotation are directed, through the paths of the lattice, to the same model-level object. 6 In the generation direction it is used to inform the micro-planner of the realization potential of each of the partial relation types. The basis of this information is a set of annotations on the lattice points that record what realizations the parser has seen for that combination of bound and open variables and in what context s they have occurred.</Paragraph>
      <Paragraph position="4"> These annotations are recorded or elaborated every time the parser reads a text that has instances (partial or full) of that category. For example, if we imagine that the parser has seen just examples 2 and 3, then, roughly speaking, it will have recorded that the combination of self and subsidiary ('s+b') can be realized as a VP (&amp;quot;owns BBN&amp;quot;) and that s+p can be realized as a possessive NP (&amp;quot;subsidiary of GTE&amp;quot;), but it will have no reason to believe that there is a direct (self-contained) realization of p+b since.it has never seen them together as the only content elements in one phrase. 7 Should it later read a text that includes the phrasing &amp;quot;'...BBN, a subsidiary of GT ...&amp;quot; (or for that matter &amp;quot;'...lsoQuest, a subsidiary of SRA...&amp;quot;), it will extend the annotation on the s+b lattice point to include that relative clause pattern.  If a category defines N variables then its saturation lattice has N+I factorial nodes over N+I&amp;quot; levels. For the initial financial example, which is the realization of a 14-tuple, this means its lattice could in principle contain several billion nodes distributed across 15 levels. ! t obviously does not, and the reason is simply that the lattice is 0nly instantiated as the parser finds particular combinations of variables. Because the compartmentalization of the elements of the 14 tuple is high and their actual patterns of combination relatively few, the lattice has not quite a hundred nodes as this is written.</Paragraph>
      <Paragraph position="5"> This is the way that Krisp implements the 'uniqueness principle' articulated by Maida and Shapiro 1982 whereby every individual has a single representation in thedomain model regardless of how often or in what context it Occurs.</Paragraph>
      <Paragraph position="6"> Given our knowledge of English grammar, we can imagine the gapping construction where this would occur: &amp;quot;GTE owns BBN and IBM Lotus&amp;quot;, but i t has not occurred in this corpus, therefore it is not included in the realization patterns recorded in the saturation lattice.</Paragraph>
    </Section>
    <Section position="3" start_page="43" end_page="43" type="sub_section">
      <SectionTitle>
2.3 Strands of annotations
</SectionTitle>
      <Paragraph position="0"> The annotations on the lattice points are not independent. They are linked together in strands running down through the saturation lattice that reflect the parser's derivation tree as it accdmulated successively larger portions of the text to the head-line of its maximal phrases, binding one variable after another in a particular order and thereby following a particular path down through the lattice. Each derivation tree that has been seen for a given combination of variable bindings is a separate strand. It is the micro-planner's job to choose one 0f these strands based on the properties of the individual nodes it is Comprised of (one for each binding). Having selected a strand, it then Creates (or extends) the Text Structure by taking the concrete relation that it has been given by the macro planner and using the strand as a recipe for introducing the objects in the relation into the Text Structure. They are added one by one as the micro-planner reads out the strand from top to bottom, at each step adding the object that is bound to the variable that was added at that lattice point.</Paragraph>
      <Paragraph position="1"> This use of Strands lets us capture some delicate co-occurrence Constrains for free because :the realizations of the terms in a relation (and their constituent terms) are not independent but must follow the pattern defined by the selected strand. In the ern domain consider the common alternation in the placement of the &amp;quot;fractional time period' term (&amp;quot;quarter&amp;quot;, &amp;quot;'nine months&amp;quot;, etc.) with respect to the 'financial item' term (&amp;quot;earnings&amp;quot;, &amp;quot;'turnover&amp;quot;, etc.) in a phrase that anchors the reporting period to a particular point in the calendar. We typically see phrasings like #4 or #5 but never the combination in #6.</Paragraph>
      <Paragraph position="2">  (4) &amp;quot;'..,quarterly earnings for the period ending March 31...&amp;quot; &amp;quot; (5) ...earnings for the quarter ending March 31...&amp;quot; (6) * &amp;quot;...quarterly earnings for the quarter ending March 31...&amp;quot;  The question is how is #6 to be avoided. The source for the anchor adjunct includes the fact that the 'period' is one fiscal quarter; what is the constraint mechanism that suppresses the expression of the actual time period when it has been stipulated, thematically, tO appear with the head noun? The answer is simply that the pattern of realiz~itions in #6 has never been seen by the parser and consequently there is 90 strand for it in the lattice. The parser has done all the work and the microplanne r reaps the benefitS without the need for any sort of active constraint propagation mechanism. It just reads out the template that the parser has provided.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>