<?xml version="1.0" standalone="yes"?>
<Paper uid="P97-1026">
  <Title>Sentence Planning as Description Using Tree Adjoining Grammar *</Title>
  <Section position="4" start_page="198" end_page="198" type="metho">
    <SectionTitle>
3 Linguistic Specifications
</SectionTitle>
    <Paragraph position="0"> Realizing this procedure requires a declarative specification of three kinds of information: first, what operators are available and how they may combine; second, how operators specify the content of a description; and third, how operators achieve pragmatic effects. We represent operators as elementary trees in LTAG, and use TAG operations to combine them; we give the meaning of each tree as a formula in an ontologically promiscuous representation language; and, we model the pragmatics of operators by associating with each tree a set of discourse constraints describing when that operator can and should be used.</Paragraph>
    <Paragraph position="1"> Other frameworks have the capability to make comparable specifications; for example, HPSG (Pollard and Sag, 1994) feature structures describe syntax (SUBCAT), semantics (CONTI3NT) and pragmatics (CONTEXT). We choose TAG because it enables local specification of syntactic dependencies in explicit constructions and flexibility in incorporating modifiers; further, it is a constrained grammar formalism with tractable computational properties. null</Paragraph>
    <Section position="1" start_page="198" end_page="198" type="sub_section">
      <SectionTitle>
3.1 Syntactic specification
</SectionTitle>
      <Paragraph position="0"> TAG (Joshi et al., 1975) is a grammar formalism built around two operations that combine pairs of trees, SUB-STITUTION and ADJOINING. A TAG grammar consists of a finite set of ELEMENTARY trees, which can be combined by these substitution and adjoining operations to produce derived trees recognized by the grammar. In substitution, the root of the first tree is identified with a leaf of the second tree, called the substitution site. Adjoining is a more complicated splicing operation, where the first tree replaces the subtree of the second tree rooted at a node called the adjunction site; that subtree is then substituted back into the first tree at a distinguished leaf called the FOOT node. Elementary trees without foot nodes are called INITIAL trees and can only substitute; trees with foot nodes are called AUXILIARY trees, and must adjoin.</Paragraph>
      <Paragraph position="1"> (The symbol $ marks substitution sites, and the symbol * marks the foot node.) Figure l(a) shows an initial tree representing the book. Figure l(b) shows an auxiliary tree representing the modifier syntax, which could adjoin into the tree for the book to give the syntax book.</Paragraph>
      <Paragraph position="2"> Our grammar incorporates two additional principles.</Paragraph>
      <Paragraph position="3"> First, the grammar is LEXICALIZED (Schabes, 1990): each elementary structure in the grammar contains at least one lexical item. Second, our trees include FEATURES, following (Vijay-Shanker, 1987).</Paragraph>
      <Paragraph position="4"> LTAG elementary trees abstract the combinatorial properties of words in a linguistically appealing way. All predicate-argument structures are localized within a single elementary tree, even in long-distance relationships, so elementary trees give a natural domain of locality over which to state semantic and pragmatic constraints.</Paragraph>
      <Paragraph position="5"> The LTAG formalism does not dictate particular syntactic analyses; ours follow basic GB conventions.</Paragraph>
    </Section>
    <Section position="2" start_page="198" end_page="198" type="sub_section">
      <SectionTitle>
3.2 Semantics
</SectionTitle>
      <Paragraph position="0"> We specify the semantics of trees by applying two principles to the LTAG formalism. First, we adopt an ONTO-</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="198" end_page="199" type="metho">
    <SectionTitle>
LOGICALLY PROMISCUOUS representation (Hobbs, 1985)
</SectionTitle>
    <Paragraph position="0"> that includes a wide variety of types of entities. Ontological promiscuity offers a simple syntax-semantics interface. The meaning of a tree is just the CONJUNCTION of the meanings of the elementary trees used to derive it, once appropriate parameters are recovered. Such fiat semantics is enjoying a resurgence in NLP; see (Copestake et al., 1997) for an overview and formalism. Second, we constrain these parameters syntactically, by labeling each syntactic node as supplying information about a particular entity or collection of entities, as in Jackendoff's X-bar semantics (Jackendoff, 1990). A node X:X (about x) can only substitute or adjoin into another node with the same label. These semantic parameters are instantiated using a knowledge base (cf. figure 7).</Paragraph>
    <Paragraph position="1"> For Jackendoff, noun phrases describe ordinary individuals, while PPs describe PLACES or PATHS and VPs describe ACTIONS and EVENTUALITIES (in terms of a Reichenbachian reference point). Under these assumptions, the trees of figure 1, are elaborated for semantics as in</Paragraph>
    <Paragraph position="3"> explore more complicated analyses in this general framework. For example, in (Stone and Doran, 1996), we use reference to properties, actions and belief contexts (Ballim et al., 1991) to describe semantic collocations (Pustejovsky, 1991) and idiomatic composition (Nunberg et al., 1994).</Paragraph>
    <Section position="1" start_page="199" end_page="199" type="sub_section">
      <SectionTitle>
3.3 Pragmatics
</SectionTitle>
      <Paragraph position="0"> Different constructions make different assumptions about the status of entities and propositions in the discourse, which we model by including in each tree a specification of the contextual conditions under which use of the tree is pragmatically licensed. We have selected four representative pragmatic distinctions for our implementation; however, the framework does not commit one to the use of particular theories.</Paragraph>
      <Paragraph position="1"> We use the following distinctions. First, entities differ in NEWNESS (Prince, 1981). At any point, an entity is either new or old to the HEARER and either new or old to the DISCOURSE. Second, entities differ in SALIENCE (Grosz and Sidner, 1986; Grosz et al., 1995). Salience assigns each entity a position in a partial order that indicates how accessible it is for reference in the current context. Third, entities are related by salient PARTIALLY-ORDERED SET (POSET) RELATIONS to other entities in the context (Hirschberg, 1985). These relations include part and whole, subset and superset, and membership in a common class. Finally, the discourse may distinguish some OPEN PROPOSITIONS (propositions containing free variables) as being under discussion (Prince, 1986). We assume that information of these four kinds is available in a model of the current discourse state.</Paragraph>
      <Paragraph position="2"> The applicability conditions of constructions can freely make reference to this information. In particular, NP trees include the determiner (the determiner does not have a separate tree), the head noun, and pragmatic conditions that match the determiner with the status of the entity in context, as in 3(a). Following (Gundel et al., 1993), the definite article the may be used when the entity is UNIQUELY IDENTIFIABLE in the discourse model, i.e. the hearer knows or can infer the existence of this entity and can distinguish it from any other hearer-old entity of equal or greater salience. (Note that this test only determines the status of the entity in context; we ensure separately that the sentence includes enough content to distinguish  inite articles, a, an, and 0, are used for entities that are NOT uniquely identifiable.</Paragraph>
      <Paragraph position="3"> S trees specify the main verb and the number and position of its arguments. Our S trees specify the unmarked SVO order or one of a number of fancy variants: topicalization (TOP), left-dislocation (LD), and locative inversion (INV). We follow the analysis of TOP in (Ward, 1985).</Paragraph>
      <Paragraph position="4"> For Ward, TOP is not a topic-marking construction at all.</Paragraph>
      <Paragraph position="5"> Rather, TOP is felicitous as long as (1) the fronted NP is in a salient poset relation to the previous discourse and (2) the utterance conveys a salient open proposition which is formed by replacing the tonically stressed constituent with a variable (3(c)). Likewise, we follow (Prince, 1993) and (Birner, 1992) for LD and INV respectively.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="199" end_page="203" type="metho">
    <SectionTitle>
4 SPUD
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="199" end_page="201" type="sub_section">
      <SectionTitle>
4.1 The algorithm
</SectionTitle>
      <Paragraph position="0"> Our system takes two types of goals. First, goals of the form distinguish x as cat instruct the algorithm to construct a description of entity x using the syntactic category cat. Ifx is uniquely identifiable in the discourse model, then this goal is only satisfied when the meaning planned so far distinguishes x for the hearer. Ifx is hearer new, this goal is satisfied by including any constituent of type cat. Second, goals of the form communicate p instruct the algorithm to include the proposition p. This goal is satisfied as long as the sentence IMPLIES p given shared common-sense knowledge.</Paragraph>
      <Paragraph position="1"> In each iteration, our algorithm must determine the appropriate elementary tree to incorporate into the current description. It performs this task in two steps to take advantage of the regular associations between words and trees in the lexicon. Sample lexical entries are shown in figure 4. They associate a word with the semantics of the word, special pragmatic restrictions on the use of the word, and a set of trees that describe the combinatory possibilities for realizing the word and may impose additional pragmatic restrictions. Tree types are shared between lexical items (figure 5). This allows us to specify the pragmatic constraints associated with the tree type once, regardless of which verb selects it. Moreover, we can determine which tree to use by looking at each tree ONCE per instantiation of its arguments, even when the same tree is associated with multiple lexical items.</Paragraph>
      <Paragraph position="2"> Hence, the first step is to identify applicable lexical entries by meaning: these items must truly and appropriately describe some entity; they must anchor trees that can substitute or adjoin into a node that describes the entity; and they must distinguish entities from their distractors or entail required information. Then, the second step identifies which of the associated trees are applicable, by testing their pragmatic conditions against the current representation of discourse. The algorithm identifies the combinations of words and trees that satisfy the most communicate goals and eliminate the most distractors.</Paragraph>
      <Paragraph position="3"> From these, it selects the entry with the most specific semantic and pragmatic licensing conditions. This means that the algorithm generates the most marked licensed form. In (Stone and Doran, 1996) we explore the use of additional factors, such as attentional state and lexical preferences, in this step.</Paragraph>
      <Paragraph position="4"> The new tree is then substituted or adjoined into the existing tree at the appropriate node. The entry may specify additional goals, because it describes one entity in terms of a new one. These new goals are added to the current goals, and then the algorithm repeats.</Paragraph>
      <Paragraph position="5"> Note that this algorithm performs greedy search. To avoid backtracking, we choose uninflected forms. Morphological features are set wherever possible as a result of the general unification processes in the grammar; the inflected form is determined from the lemma and its associated features in a post-processing step.</Paragraph>
      <Paragraph position="6"> The specification of this algorithm is summarized in the following pseudocode:  until goals are satisfied: determine which uninflected forms apply; determine which associated trees apply; evaluate progress towards goals; incorporate most specific, best ( form, tree ): perform adjunction or substitution; conjoin new semantics; add any additional goals;</Paragraph>
    </Section>
    <Section position="2" start_page="201" end_page="201" type="sub_section">
      <SectionTitle>
4.2 The system
</SectionTitle>
      <Paragraph position="0"> SPUD's grammar currently includes a range of syntactic constructions, including adjective and PP modification, relative clauses, idioms and various verbal alternations.</Paragraph>
      <Paragraph position="1"> Each is associated with a semantic and pragmatic specification as discussed above and illustrated in figures 4 and 5. These linguistic specifications can apply across many domains.</Paragraph>
      <Paragraph position="2"> In each domain, an extensive set of inferences, presumed known in common with the user, are required to ensure appropriate behavior. We use logic programming to capture these inferences. In our domain, the system has the role of a librarian answering patrons' queries.</Paragraph>
      <Paragraph position="3"> Our specifications define: the properties and identities of objects (e.g., attributes of books, parts of the library); the taxonomic relationships among terms (e.g., that a service desk is an area but not a room); and the typical span and course of events in the domain (e.g., rules about how to check out books). This information is complete and available for each lexical entry. Of course, SPUD also represents its private knowledge about the domain. This includes facts like the status of books in the library.</Paragraph>
    </Section>
    <Section position="3" start_page="201" end_page="203" type="sub_section">
      <SectionTitle>
4.3 An example
</SectionTitle>
      <Paragraph position="0"> Suppose our system is given the task of answering the  following question: (1) Do you have the books for Syntax 551 and  Pragmatics 590? Figure 6 shows part of the discourse model after processing the question. The two books, the set they comprise (introducing a poset relation), and the library are mentioned in (1). Hence, these entities must be both hearer-old and discourse-old. As in centering (Grosz et al., 1995), the subject is taken to be the most salient entity. Finally, the meaning of the question becomes a salient open proposition.</Paragraph>
      <Paragraph position="1"> On the basis of the knowledge in figure 7, a rhetorical planner might decide to answer by describing state have27 as an S and lose5 likewise. To construct its reference to have27, SPUD first determines which lexical and syntactic options are available. Using the lexicon and information about have27 available from figure 7(b), SPUD determines that, of lemmas that truthfully and appropriately describe have27 as an S,/have/has the most specific licensing conditions. The tree set for/have/includes unmarked, LD and TOP trees. All are licensed, because of the poset relation R between book19 and books and the salient open proposition O. We choose TOP, the tree with the most specific condition--TOP requires R and O, while LD requires only R and the unmarked form has no requirements.</Paragraph>
      <Paragraph position="2"> Thus, a topicalized /have/ tree, appropriately instantiated as shown in figure 8, is added to the description. The tree refers to three new entities, the object book19, the subject library and the reference point r of the tense.  of have27.</Paragraph>
      <Paragraph position="3"> Subgoals of distinguishing these entities are introduced. Other constructions that describe one entity in terms of another, such as complex NPs, relative clauses and semantic collocations, are also handled this way by SPUD. The algorithm now selects the goal of describing book19 as an NP. Again, the knowledge base is consulted to select NP lemmas that truly describe book19. The best is/book/. (Note that SPUD would use it if either the verb or the discourse context ruled out all distractors.) The tree set for/book/includes trees with definite and indefinite determiners; since the hearer can uniquely identify book19, the definite tree is selected and substituted as the leftmost NP, as in figure 9.</Paragraph>
      <Paragraph position="4"> The goal of distinguishing book19 is still not satisfied, however; as far as the hearer knows, both book19 and book2 could match the current description. We consult the knowledge base for further information about book19. The modifier entry for the lexical item/syntax/ can apply to the new N node; its tree adjoins there, giving the tree in figure 10. Note that because trees are lexicalized and instantiated, and must unify with the existing derivation, SPUD can enforce collocations and idiomatic  book19.</Paragraph>
      <Paragraph position="5"> composition in steps like this one.</Paragraph>
      <Paragraph position="6"> Now we describe the library. Since we ARE the library, the lexical item/we/is chosen. Finally, we describe r. Consulting the knowledge base, we determine that r is now, and that the present tense morpheme applies. For uniformity with auxiliary verbs, we represent it as a separate tree, in this case with a null head, which assigns a morphological feature to the main verb. This gives the  tree in figure 11, representing the sentence: (2) The syntax book, we have.</Paragraph>
      <Paragraph position="7"> All goals are now satisfied. Note that the semantics has been accumulated incrementally and straightforwardly in parallel with the syntax.</Paragraph>
      <Paragraph position="8"> To illustrate the role of inclusion goals, let us suppose that the system also knows that book19 is on reserve in the state have27. Given the additional input goal of communicating this fact, the algorithm would proceed as before, deriving The syntax book we have. However, the new goal would still be unsatisfied; in the next iteration, the PP on reserve would be adjoined into the tree to satisfy it: The syntax book, we have on reserve. Because TAG allows adjunction to apply at any time, flexible realization of content is facilitated without need for sophisticated back-tracking (Elhadad and Robin, 1992).</Paragraph>
      <Paragraph position="9"> The processing of this example may seem simple, but it illustrates the way in which SPUD integrates syntactic, semantic and pragmatic knowledge in realizing sentences. We tackle additional examples in (Stone and Doran, 1996).</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="203" end_page="203" type="metho">
    <SectionTitle>
5 Comparison with related work
</SectionTitle>
    <Paragraph position="0"> The strength of the present work is that it captures a number of phenomena discussed elsewhere separately, and does so within a unified framework. With its incremental choices and its emphasis on the consequences of functional choices in the grammar, our algorithm resembles the networks of systemic grammar (Mathiessen, 1983; Yang et al., 1991). However, unlike systemic networks, our system derives its functional choices dynamically using a simple declarative specification of function. Like many sentence planners, we assume that there is a flexible * association between the content input to a sentence planner and the meaning that comes out. Other researchers (Nicolov et al., 1995; Rubinoff, 1992) have assumed that this flexibility comes from a mismatch between input content and grammatical options. In our system, such differences arise from the referential requirements and inferential opportunities that are encountered.</Paragraph>
    <Paragraph position="1"> Previous authors (McDonald and Pustejovsky, 1985; Joshi, 1987) have noted that TAG has many advantages for generation as a syntactic formalism, because of its localization of argument structure. (Joshi, 1987) states that adjunction is a powerful tool for elaborating descriptions. These aspects of TAGs are crucial to SPUD, as they are to (McDonald and Pustejovsky, 1985; Joshi, 1987; Yang et al., 1991; Nicolov et al., 1995; Wahlster et al., 1991; Danlos, 1996). What sets SPUD apart is its simultaneous construction of syntax and semantics, and the tripartite, lexicalized, declarative grammatical specifications for constructions it uses. Two contrasts should be emphasized in this regard. (Shieber et al., 1990; Shieber and Schabes, 1991) construct a simultaneous derivation of syntax and semantics but they do not construct the semantics--it is an input to their system. (Prevost and Steedman, 1993; Hoffman, 1994) represent syntax, semantics and pragmatics in a lexicalized framework, but concentrate on information structure rather than the pragmatics of particular constructions.</Paragraph>
  </Section>
class="xml-element"></Paper>