<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1419">
  <Title>TEXTUAL ECONOMY THROUGH CLOSE COUPLING OF SYNTAX AND SEMANTICS</Title>
  <Section position="4" start_page="179" end_page="180" type="metho">
    <SectionTitle>
2 SPUD*
</SectionTitle>
    <Paragraph position="0"> An NLG system must satisfy at least three constraints in mapping the content planned for a sentence onto the string of words that realize it \[4, 13, 20\]. Any fact to be communicated must be fit into an abstract grammatical structure, including lexical items. Any reference to a domain entity must be elaborated into a description that distinguishes the entity from its distractors--the salient alternatives to it in context. Finally, a surface form must be found for this conceptual material.</Paragraph>
    <Paragraph position="1"> In one architecture for NLG Systems that is becoming something of a standard \[22\], these tasks are performed in Separate stages. For example, to refer to a uniquely identifiable entity x from the common ground, first a set of concepts is identified that together single out x from its distractors in context. Only later is the syntactic structure that realizes those concepts derived.</Paragraph>
    <Paragraph position="2"> SPUD \[26, 27\] integrates these processes in generating a description--producing both syntax and semantics simultaneously, in stages, as illustrated in (4).</Paragraph>
    <Paragraph position="4"> Lexicalized Tree-Adjoining Grammar (LTAG) \[23\]. A tree may contain multiple lexical items (cf. (4)b).</Paragraph>
    <Paragraph position="5"> Each such tree is paired with logical formulae that, by referring to a rich discourse model, characterize the semantic and pragmatic contribution that it makes to the sentence. We give a detailed example of SPUD's processing in Section 3 and describe in Section 4 the reasoning methods we use to derive computational representations like the set of distractors shown in (4). For now, a general understanding of SPUD suffices-this is provided by the summary in Figure 2.</Paragraph>
    <Paragraph position="6"> The procedure in Figure 2 is sufficiently general so that SPUD* can use similar steps to construct both definite and indefinite referring forms. The main difference lies how alternatives are evaluated. When an indefinite referring form is used to refer tO a brand-new generalized individual \[19\] (an object, for example,  * Start with a tree with one node (e.g., s, uP) and one or more referential or informational goals.</Paragraph>
    <Paragraph position="7"> * While the current tree is incomplete, or its references are ambiguous to the hearer, or its meaning does not fully convey the informational goals (provided progress is being made): - consider the trees that extend the current one by the addition (using LTAG operations) of a true and appropriate lexicalized descriptor; - rank the results based on local factors (e.g., completeness of meaning, distractors for reference, unfilled substitution sites, specificity of licensing conditions); - make the highest ranking new tree the current tree.</Paragraph>
    <Paragraph position="8">  or an action in an instruction), the object is marked as new and does not have to be distinguished from others because the hearer creates afresh &amp;quot;file card&amp;quot; for it. However, because the domain typically provides features needed in an appropriate description for the object, SPUD continues its incremental addition of content to convey them, When an indefinite form is used to refer to an old object that cannot be distinguished from other elements of a uniquely identifiable set (typically an inferrable entity \[ 19\]), a process like that illustrated in (4) must build a description that identifies this set, based on the known common properties of its elements. Several advantages of using LTAG in such an integrated system are described in \[27\] (See also previous work on using TAG in NLG such as \[11\] and \[29\]). These advantages include: * Syntactic constraints can be handled early and naturally. In the problem illustrated in (4), SPUD directly encodes the syntactic requirement that a description should have a head noun--missing from the concept-level account--using the NP substitution site.</Paragraph>
    <Paragraph position="9"> * The order of adding content is flexible. Because an LTAG derivation allows modifiers to adjoin at any step (unlike a top-down CFG derivation), there is no tension between providing what the syntax requires and going beyond what the syntax requires.</Paragraph>
    <Paragraph position="10"> * Grammatical knowledge is Stated once only. All operations in constructing a sentence are guided by LTAG's lexicalized grammar; by contrast, with separate processing, the lexicon is split into an inventory of concepts (used for organizing content orconstructing descriptions) and a further inventory of concepts in correspondence With some syntax (for surface realization).</Paragraph>
    <Paragraph position="11"> This paper delineates a less obvious, but equally significant advantage that follows from the ability to consider multiple goals in generating descriptions, using a representation and a reasoning process in which syntax and semantics are more closely linked: * It naturally supports textual economy.</Paragraph>
  </Section>
  <Section position="5" start_page="180" end_page="182" type="metho">
    <SectionTitle>
3 Achieving Textual Economy
</SectionTitle>
    <Paragraph position="0"> To see how SPUD supports textual economy, Consider first how SPUD might derive the instruction in Example (1). For simplicity, this explanation assumes SPUD makes a nondeterministic choice from among available lexical entries; this suffices to illustrate how SPUD can realize the textual economy of this example.</Paragraph>
    <Paragraph position="1"> A priori, SPUD has a general goal of describing a new action that the hearer is to perform, by making sure the hearer can identify the key features that allow its performance. For (1), then, SPUD is given two  features of the action to be described: it involves motion of an intended object by the agent, and its result is achieved when the object reaches a place decisively away from its starting point.</Paragraph>
    <Paragraph position="2"> The first time through the loop of Figure 2, SPUD must expand an s node. One of the applicable moves is to substitute a lexical entry for the verb remove. Of the dements in the verb's LTAG tree family, the one that fits the instructional context is the imperative tree of (5).</Paragraph>
    <Paragraph position="3">  away(RESULT, end(TIME), REMOVED, SOURCE) The tree given in (5a) specifies that remove syntactically satisfies a requirement tO include an s, requires a further NP to be included (describing what is removed), and allows the possibility of an explicit vv modifier that describes what the latter has been removed from. 2 The semantics in (5b) consists of a set of features, formulated in an ontologically promiscuous semantics, as advocated in \[9\]. It follows \[14\] in viewing events as consisting of a preparatory phase, a transition, and a result state (what is called a nucleus in \[14\]). The semantics in (5b) describes all parts of a remove event: In the preparatory phase, the object (REMOVED) is in/on SOURCE. It undergoes motion caused by the agent (REMOVER), and ends up away from SOURCE in the result state.</Paragraph>
    <Paragraph position="4"> Semantic features are used by SPUD in one of two ways. Some make a semantic contribution that specifies new information---these add to what new information the speaker can convey with the structure. Others simply impose a semantic requirement that a fact must be part of the conversational record--these figure in ruling out distractors.</Paragraph>
    <Paragraph position="5"> For this instruction, SPUD treats the CAUSED-MOTION and AWAY semantic features as semantic contributions.&amp;quot; It therefore determines that the use of this item communicates the needed features of the action. At the same time, it treats the IN feature---because it refers to the shared initial state in which the instruction will be executed--and the NUCLEUS feature---because it simply refers to our general ontology--as semantic requirements. SPUD therefore determines that the only (REMOVED,SOURCE) pairs that the hearer might think the instruction could refer to are pairs where REMOVED starts out in/on SOURCE as the action begins. Thus, SPUD derives a triple effect from use of the word remove--increasing syntactic satisfaction, making semantic contributions and satisfying semantic requirements--all of which contribute to SPUD's task of completing an S syntactic constituent that conveys needed content and refers successfully. Such multiple effects make it natural for SPUD to achieve textual economy. Positive effects on any of the above dimensions can suffice to merit inclusion of an item in a given sentence. However, the effects of inclusion may go beyond this: even if an item is chosen for its semantic contribution, its semantic requirements can 2Other possibilities are that SOURCE is not mentioned explicitlyl but is rather inferred from (1) the previous discourse or, as we will discuss later, (2) either the predicated relationships within the clause or its informational relationship to another clause.  still be exploited in establishing whether the current lexico-syntactic description is sufficient to identify an entity, and its syntactic contributions can still be exploited to add further content.</Paragraph>
    <Paragraph position="6"> Since the current tree is incomplete and referentially ambiguous, sPUD repeats the loop of Figure 2, considering trees that extend it. One option is to adjoin at the w the entry corresponding to from the hat. In this compound entry, from matches the verb and the matches the context; hat carries semantics, requiring that SOURCE be a hat. After adjunction, the requirements reflect both remove and hat; reference, sPUD computes, has been narrowed to the hats that have something in/on them (the rabbit, the flower).</Paragraph>
    <Paragraph position="7"> Another option is to substitute the entry for the rabbit at the object NP; this imposes the requirement that REMOVED be a rabbit. Suppose sPUD discards this option in this iteration, making the other (perhaPS . less referentially ambiguous) choice. At the next iteration, the rabbit still remains an option. Now combining with remove and hat, it derives a sentence that SpUD recognizes to be complete and referentially unambiguous, and to satisfy the informational goals.* Now we consider the derivation of (2), which shows how an informational relation between clauses can support textual economy in the clauses that serve as its &amp;quot;arguments&amp;quot;. SPUD starts with the goal of describing the holding action in the main clause, and (if possible) also describing the filling action and indicating the purpose relation (i.e., enablement) between them. For the homing action, sPUD's goals include making sure that the sentence communicates where the cup will be held and how it will be held (i.e., UPWARD). SPUD first selects an appropriate lexico-syntactic tree for imperative hold; sPUD can choose to adjoin in the purpose clause next, and then to substitute in an appropriate lexico-syntactic tree forfill. After this substitution, the semantic contributions of the sentence describe an action of holding an object which generates an action of filling that object. As shown *in \[7\], these are the premises of an inference that the object is held upright during the filling. When SPUD queries its goals at this stage, it thus finds that it has in fact conveyed how the cup is to be held. SPUD has no reason to describe the orientation of the cup with additional content.</Paragraph>
    <Paragraph position="8"> Additional examples of using SPUD to generate instructions can be found in \[3, 25\].</Paragraph>
  </Section>
  <Section position="6" start_page="182" end_page="184" type="metho">
    <SectionTitle>
4 Assessing interpretation in SPUD
</SectionTitle>
    <Paragraph position="0"> This section describes in a bit more detail how SPUD computes the effects of incorporating a particular lexical item into the sentence being constructed. For a more extensive discussion, see \[25\].</Paragraph>
    <Paragraph position="1"> * sPUD's computations depend on its representation of overall contextual background, including the status of propositions and entities in the discourse. For the purpose of generating instructions to a single hearer, we assume that any proposition falls either within the private knowledge of the speaker or within the common ground that speaker and hearer share. We implement this distinction by specifying facts in a modal logic with an explicit representation of knowledge: \[siP means that the speaker knows p; \[C/\]p means that p is part of the common ground. Each entity, e, comes with a context set D(e) including it and its distractors.</Paragraph>
    <Paragraph position="2"> Linguistically, when we have a E D(b) but not b E D(a), then a is more salient than b.</Paragraph>
    <Paragraph position="3"> This conversational background serves as a resource for constructing and evaluating a three-part staterecord for an incomplete sentence, consisting of: * An instantiated tree describing the syntactic structure of the sentence under construction. Its nodes are labeled by a sequence of variables v indicating the patterns of coreference in the tree; but the tree also records that the speaker intends v to refer to a particular sequence of generalized individuals e.</Paragraph>
    <Paragraph position="4"> * The semantic requirements of the tree, represented by a formula R(v). This formula must match facts in the common ground; in our modal specification, such a match corresponds to a proof whose conclusion instantiates \[C\]R(v). In particular, the speaker ensures that such a proof is available when v is instantiated to the entities e that the speaker means to refer to. This determines what alternative  referents that the hearer may still consider: { a E D(e) \[ \[c\]R(a) }. The semantic requirements of the tree result from conjoining the requirements Ri(vi) of the *individual lexical items from which the state is derived.</Paragraph>
    <Paragraph position="5"> The semantic contributions of the tree, represented by a formula N(v); again, this ~s the conjunction of the contributions Ni(vi) of the individual items. These contributions are added to the common ground, allowing both speaker and hearer to draw shared conclusions from them. This has inspired the following test for whether a goal to communicate G has been indirectly achieved. Consider the content of the discourse as represented by \[C\], augmented by what this sentence will contribute (assuming we identify entities as needed for reference): N(e). Then if G follows, the speaker has conveyed what is needed.</Paragraph>
    <Paragraph position="6"> When SPUD considers extending a state by a lexical item, it must be able to update each of these records quickly. The heart of sPUD's approach is logic programming \[25\], which links complexity of computation and complexity of the domain in a predictable way. For example, informational goals are assessed by the query \[c\](N(e) D G)I This leaves room for inference when necessary, without swamping sPUD; in practice, G is often a primitive feature of the domain and the query reduces to a simple matching operation. Another source of tractability comes from combining logic programming with special-purpose reasoning.</Paragraph>
    <Paragraph position="7"> For example, in computing reference, { ai E D(ei) \[ \[c\]Ri(ai) } is found using logic programming but the overall set of alternatives is maintained using arc-consistency constraint-satisfaction, as in \[6, 8\]. SPUD must also *settle which semantic features are taken to constitute the semantic requirements of the lexical item and which are taken to constituteits semantic contributions. 3 When SPUD partitions the semantic features of the lexical item, as many features as possible are cast as requirements~that is, the item links as strongly with the context as possible. In some cases, the syntactic environment may further constrain this assignment. For example, we constrain items included in a definite NP to be semantic requirements, while the main verb in an indicative sentence is usually taken to make a semantic contribution. (Exceptions to such a policy are justified in \[28\].) 3These can vary with context: consider a variant on Figure 1, where the hearer is asked &amp;quot;What just happened?&amp;quot;. One possible response-- &amp;quot;I have removed the rabbit from the hat&amp;quot; m refers successfully, despite the many rabbits and hats, because there is still only one rabbit in this scene that could have been removed from a hat. Here, where the scene is taken as shared, what is taken as a semantic requirement of remove---thatthe rabbit ends up away from the hat--is used to identify a unique rabbit. This contrasts with the previous &amp;quot;rabbit&amp;quot; example where, taking the scene in Figure 1 as shared, the command &amp;quot;Remove the rabbit from the hat&amp;quot; takes as its semantic requirement that the rabbit be in the hat and uses it for unique identification. Note that if the above scene is not taken as shared, both are then taken as semantic contributions, and &amp;quot;I have removed a rabbit from a hat&amp;quot; becomes an acceptable answer.</Paragraph>
  </Section>
  <Section position="7" start_page="184" end_page="184" type="metho">
    <SectionTitle>
5 Other Methods that Contribute to Eflicient Descriptions
</SectionTitle>
    <Paragraph position="0"> This section contrasts spoI&gt;--and its close coupling of syntax and semantics--with prior work on generating more concise descriptions by considering the effects of broader goals, 4 starting with Appelt \[1\]. Appelt's planning formalism includes plan-critics that can detect and collapse redundancies in sentence plans.</Paragraph>
    <Paragraph position="1"> However, his framework treats subproblems in generation as independent by default; and writing tractable and general critics is hampered by the absence of abstractions like those used in SPUD to simultaneously model the syntax and the interpretation of a whole sentence.</Paragraph>
    <Paragraph position="2"> \[6, 10, 12\], in contrast, use specialized mechanisms to capture particular descriptive efficiencies. By using syntax to work on inferential and referential problems simultaneously, SPUD captures such efficiencies in a uniform procedure. For example, in \[12\], McDonald considers descriptions of events in domains which impose strong constraints on what information about events is semantically relevant. He shows that such material should and can be omitted, if it is both syntactically optional and inferentially derivable:</Paragraph>
  </Section>
  <Section position="8" start_page="184" end_page="185" type="metho">
    <SectionTitle>
FAIRCHILD Corporation (Chantilly VA) Donald E Miller was named senior vice president and
</SectionTitle>
    <Paragraph position="0"> general counsel, succeeding Dominic A Petito, who resigned in November, at this aerospace business. Mr. Miller, 43 years old, was previouslY principal attorney for Temkin &amp; Miller Ltd., Providence RI.</Paragraph>
    <Paragraph position="1"> Here, McDonald points out that one does not need to explicitly mention the position that Petito resigned from in specifying the resignation sub-event, since it must be the same as the one that Miller has been appointed to. This can be seen as a special case of pragmatic overloading.</Paragraph>
    <Paragraph position="2"> Meanwhile, Dale and Haddock \[6\] consider generating interacting references, building on Haddock's work on reference resolution \[8\]. Their example NP, the rabbit in the hat, refers successfully in a context with many rabbits and many hats, so long as only one of the rabbits, 1&amp;quot;5 say, is actually in one of the hats, h3 say. Like (1), the efficiency of this description comes from the uniqueness of this rabbit-hat pair. However, Dale and Haddock construct NP semantics in isolation and adopt a fixed, depth-first strategy for adding content. Horacek \[10\], challenges this strategy with examples that show the need for modification at multiple points in an NP. For example, (6) refers with respect to the scene in Figure 3.</Paragraph>
    <Paragraph position="3">  (6) the table with the apple and with the banana.</Paragraph>
    <Paragraph position="4"> (6) identifies a unique table by exploiting its association with two objects it supports: the apple and the  banana that are on it. (Note the other tables, apples and bananas in the figure--and even tables with apples and tables with bananas.) Reference to one of these--the apple, say--is incorporated into the description 4Other ways of making descriptions more concise, such as through the use of anaphoric and deictic pronouns (or even pointing, in multi-modal contexts), are parasitic On the heater's focus of attention, which can (in large part) be defined independently of goal-directed features of text.</Paragraph>
    <Paragraph position="5">  first; then that (subordinate) entity is identified by further describing the table (higher up).5 By considering sentences rather than isolated noun phrases, SPUD extends such descriptive capacities even further.</Paragraph>
  </Section>
class="xml-element"></Paper>