<?xml version="1.0" standalone="yes"?>
<Paper uid="E91-1020">
  <Title>STRUCTURE-DRIVEN GENERATION FROM SEPARATE SEMANTIC REPRESENTATIONS</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
KEEPING SEMANTICS SEPARATE
FROM SYNTAX
</SectionTitle>
    <Paragraph position="0"> The integrated-semantics approach is often illustrated in a Prolog-like notation using DCG rules.</Paragraph>
    <Paragraph position="1"> The infix function symbol '/' is used in each category to separate tile syntactic from the semantic part. Rule (1) introduces complements in an llPSGstyle manner by &amp;quot;removing&amp;quot; tile complement from the VP's subcategorization list (cf. \[Pollard/Sag 1987\]). The relation between the semantics S and the semantics of Comp:l. is established in tile lexical entry for tile verb (2).</Paragraph>
    <Paragraph position="2">  - 113(1) vp(Subcat)/S --&gt; vp(\[CompllSubcat\])/S, Compl.</Paragraph>
    <Paragraph position="3"> (2) vp(\[np(_)/0bj, np(3rd-sing)/Subj\])/  kiss(Subj, 0bj) --&gt; \[kisses\].</Paragraph>
    <Paragraph position="4"> Recent work on semantic-head-driven generation \[Shieber et al. 1990, Calder et al. 1989, Noord 1990, Russell et al. 1990\] provides a very promising step towards efficient, goal-directed reconstruction of LF that is espescially suited for lexicon-centered grammar formalisms such as IIPSG or UCG. It was observed that top-down generation may not terminate. This is illustrated in (1). If the vp node is used for top-down expansion, there is nothing to prevent the subcategorization list from growing infinitely. If the Comp node is used, the constituent to be generated must completely be guessed due to the uninstantiated semantics. Since the grammar will contain recursive rules (e.g. for relative clauses), the guessing procedure will not terminate either. In view of this problem a bottom-up approach was suggested that is guided by semantic information in a top-down fashion.</Paragraph>
    <Paragraph position="5"> The benefits of integrated semantics are manifold. Elegant analyses of linguistic phenomena are possible that relate syntactic and semantic properties to each other (cf. the treatment of e.g. 'raising' and 'equi' constructions in \[Pollard/Sag 1987\]). LF is defined on purely linguistic grounds and as such, it is well-suited to tile contputationai linguist's work.</Paragraph>
    <Paragraph position="6"> llowever, if a generator based on an integrated semantics is to be used for conveying the results of some application system into NL, expressions of the application system's SRL have to be adalJted to LF. Given that tile grammar should not be rewritten, this amou,,ts to an additional'step of processing. This step may turn out to be costly since the SRL will typically contain application-dependent information that must be considered. Take, for instance, a transfer-based machine translation (MT) system (such as EUROTRA \[Arnold/des Tombe 1986\]).</Paragraph>
    <Paragraph position="7"> The results of the transfer (say, from German to English) are encoded in a semantic representation that is given to the system's generation component to produce the English target sentence. In a system capable of translating between a variety of languages, representations of this kind may themselves be subject to transfer and will therefore contain information relevant for translation. 2 SAn exception is tim MiMe2 system \[Noord et al. 1990\]. The price to pay for allowing transfer at the level of LF was to accept an &amp;quot;extremely poor&amp;quot; view of translation by just preserving the logical meaning emd--as far as possible--the way in which meaning is built compositionMiy (quotation from \[Noord et al. 1990\]).</Paragraph>
    <Paragraph position="8"> The effort of introducing an additional step of processing can be saved to a large extent by adopting a separate-semantics approach. The SRL of some application system may directly serve as an interface to the generator. 3 In the case at hand, two additional components must be introduced into the generation scenario: the definition of SRL and PA rules. Instead of mapping SRL onto LF, SRL is directly related to syntax by virtue of the PA rules.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
A STRUCTURE-DRIVEN GENERATOR
</SectionTitle>
    <Paragraph position="0"> The generator to be described in this section is a module of the Berlin MT system \[llauenschild/Busemann 1988\], which translates sentences taken from administrative texts in an EC corpus from German into English and vicc versa. 4 The syntax formalism Used is a constructive version of GPSG \[Gazdar et al. 1985\] as described in \[Busemann/Hauenschild 1988\]. The semantic representation language FAS (Functor-Argument Stuctures) \[Mahr/Umbach 1990\] is employed as an interface between three different processes: it is the target of GPSG-based analysis, for sentence-semantic transfer, and as the source for GPSG-based generation.</Paragraph>
    <Paragraph position="1"> FAS is defined by context-free rule schemata with complex categories consisting of a main category (e.g. 'clause' in Figure la), which is associated with a fixed list of feature specifications. 5 The categories are in canonical order with the functor preceding all of its arguments. In contrast to syntactic structures where agreement relations are established by virtue of feature propagation, FAS categories contain alnmst no redundant information. For instance, number information is only located at the 'det' category. The use of semantic relations (encoded by the 'role' feature), role configurations ('conf') and semantic features allows us to discriminate between different readings of words that result in different translational equivalents. Moreover, part of the thematic structure of the source language sentence is preserved during transfer and encoded by virtue of the feature 'them' with the numerical values indicating which portion should preferrably be presented first, second, third etc. The definitions of FAS for the German and English fragments mainly differ with regard to their terminal symbols.</Paragraph>
    <Paragraph position="2"> 3This interface does not correspond to the common separation between making decisions about what to say and how to say it (cf. \[McKeown/Swartout 1988\]). Rather the interface in question must be situated somewhere in the 'how to say it' component because it presupposes many decisions about sentence formulation (e.g. regarding pronominalization,  Sln the present versions there are up to seven features in a FAS category. For sake of simplicity many details irrelevant to the present discussion are omitted in the examples.</Paragraph>
    <Paragraph position="3">  The GPSG formalism used includes the ID/LP format, feature co-occurrence restrictions (FCRs) and universal principles of feature instantiation (FIPs). The ID rules are interpreted by the generator as providing the basic information for a local tree. The categories of each generated local tree are filrther instantiated by the FIPs and FCRz. Finally, the branches are ordered by virtue of the LP statelnen|.s. null Strategies for structure building and feature instantiation. The task of constructing an admissible GPSG syntactic structure call be divided up into the following suhta.sks that can be performed independently of each other, and each according to its own processing strategy: ,, Structure building (by virtue of PA rules, which in turn use ID rules) . Feature instantiaton and ordering of the branches (by virtue of FIPs, FCRs and LP statemerits) null The question arises which strategies are best suited to ellicient generation. For each subtask both a top-down and a bottom-up strategy have been investigated. As a result it turned out that structure building shouhl occur top-down whereas feature instantiation should be performed in a bottom-up manner.</Paragraph>
    <Paragraph position="4"> Before .justifying the result let us have a closer look at the sl.ructure-buiiding algorithm. Tile over-all syntactic structure (OSS) is successively construed in a top-down manner. At each level there is a set of nonterminal leaf nodes available serving as attachment points for further expansion steps (initially tile empty category is the only attachment point). An expansion step consists of  1. generating a local tree t by virtue of an ID rule, 2. unifying its mother node with one of the attachment points, 3. removing the attachment point from the current set, 4. defining tile daughters of t as the new current set of attachment points.</Paragraph>
    <Paragraph position="5">  Since lexicai entries terminate a branch of the OSS, the fourth of the above points is dropped during expansion of lexical categories: processing continues with the reduced set of attachment points.</Paragraph>
    <Paragraph position="6"> Feature instafftiation and the ordering of branches take place in a bottom-up manner after a local tree has no fuither attachment points associated with it (i.e. all of its daughters have been expanded). Then processing returns to tile next higher level of tile OSS examining the set of attachment points. Depending on whether or not it is empty, the next step is either feature instantiation or structure building. Given this interlinking of the two subtasks, all OSS is admitted by tile grammar if  its top-most local tree has passed feature instantiation. null The effects of feature instantiation with respect to the German example in Figure lb 6 can be better understood with the help of the S-expansion rules used; of. (3)-(5). t Rule (3) causes topicalization,  (4) introduces a perfect auxiliary, and (5) requires a transitive verb whose object is topicalized.</Paragraph>
    <Paragraph position="7"> (3). S , X\[+top\],S\[fin\] / X\[+top\] (4) s ,v,s\[psp\] (5) s / NP\[+top, ace\] , NP\[nom\], V\[trans\]  The solution will now be justified. First of all, note that the top-most part of an FAS expression is related to tile top-most part of the GPSG structure, and that the leaves of a FAS expression usually correspond to GPSG lexicon entries. As a consequence, the order the FAS expression is traversed determines the order in which the structure-building sub-task is performed. Why should then, in the case of FAS, the traversal occur top-down? The answer is motivated by the distribution of information in FAS expressions. In order to apply a certain ID rule deterministically, information from distant portions of tim FAS expression may be needed. For instance, the FAS specification (them : 1), which is part of one of the daughters of clause in Figure la, is interpreted as requiring topicalization of a syntactic constituent under the condition that a declarative sentence is being generated. This latter information is, however, only available at the \[illoC/ \[asnertion\] \] s part of the FAS expression (of. Figure la).</Paragraph>
    <Paragraph position="8"> Two possible methods for collecting this infornration present themselves. First, the pattern including (them : 1) could be required to cover as nmch of the FAS expression as would be needed to include i\].loc. In that case, all the information needed is present, and the traversal of the FAS expression could occur bottom-up as well as top-down. * Unfortunately the required size of the pattern is not always known in advance because the FAS syntax might allow an arbitrary number of recursively defined local trees to intervene.</Paragraph>
    <Paragraph position="9"> The second method--which was eventually adopted--requires the patterns to cover not more than one local FAS tree. In order to gather information that is locally missing, an auxiliary storage is needed. If, for instance, the illocution is matched, information about whether or not a declarative sentence is being generated is stored. Later on, (them : 1) is encountered. Now, the ID rule for to6These are not shown for the constituents of NPs.</Paragraph>
    <Paragraph position="10"> ZNote the different use of the symbol '/': here it denotes the category-valued feature 'slash'.</Paragraph>
    <Paragraph position="11"> e Square brackets are used here to indicate tree stnicture. picalization (3) is triggered iff 'declarative' can be retrieved from the storage.</Paragraph>
    <Paragraph position="12"> If the necessary information is not available yet, one must accept either a delay of a mapping or backtracking. With a top-down traversal of FAS expressions, however, such cases are sufficiently restricted to ensure efficiency. Note that a bottom-up traversal or a mixed strategy could be more efficient if the distribution of information in the SRL were different.</Paragraph>
    <Paragraph position="13"> The problems observed with top-down generatots using an integrated semantics cannot occur in the separate-semantics approach. Expansion of grammar rules can be controlled by the semantic representation if each rule application is explicitly triggered. Situations causing an infinite expansion due to an uninstantiated semantics (as with top-down expansion using the rule (2)) cannot arise at all since the separate semantics is fully specified. Let us now discuss why feature instantiation should be a bottom-up process. The FIPs apply to tim mother and/or a subset of daughters in a local tree. In general, tile more these categories are instantiated the less likely the l&amp;quot;lPs will have to choose between alternative instantiations, which would be a source for backtracking. A top-down strategy would meet a more completely instantiated mother, but still underspecified daughters. With a bottom-up strategy, howew:r, only tile mother would be underspecified. For instance, consider the GPSG account of parasitic gaps, which are handled by the Foot Feature Principle. The 'slash' feature may occur at more than one daughter and then require all occurrences of it to unify with the mother (el. \[Gazdar et al. 1985, p. 16211\]). While this is easy to handle for a bottom-up process, a top-down strategy would have to guess at which daughters to instantiate a slash value.</Paragraph>
    <Paragraph position="14"> Pattern-action rules. A PA rule is a production rule with a pattern for local FAS trees as its left-hand side and two sets of actions as its right-hand side. The information-gathering aclions (IGAs) maintain the auxiliary storage. The structure-building actions (SBAs) generate GPSG trees. Either one of these sets may be empty.</Paragraph>
    <Paragraph position="15"> In:order to minimize tim power of PA rules, the inventory of IGAs and SBAs is restricted. There are only lthree 1GAs for storing information into and removing from the auxiliary storage. The auxiliary storage is a two-dimensional array of a fixed size. It may contain atomic values for a set of features pre-determined by the PA rule writer as well as a single GPSG category. There are only five SBAs for different kinds of mapping, three of which are explained below; cf. \[Busemann 1990\] for a coml)rehensive discussion. Any SBA' will remove the stored category  from the storage and unify it with the :mother of the local tree it is about to generate.</Paragraph>
    <Paragraph position="16"> To illustrate this let us return to the topicalization example. The responsible PAl rules are shown in Figure 2. The pattern of the first one naatches any local FAS tree whose mbther is a term(them: 1). The 1GAs work as follows: Ifa specification (sent : (lecl) can be removed from the storage, the GPSG feature specification \[+top\] will be added to the stored category (by virtue of the IGA set_gpsg_features). The SBA set is empty. The second PA rule matches any local FAS tree whose first daughter is a dcfinite determiner with plural number followed by zcro or more daughters. Note that both patterns match the same local tree of the FAS expression in Figure la. There is only one IGA, which adds the number information to the stored GPSG category. The single SBA, call_id, states that a local GPSG tree is generated by virtue of the ID rule indicated and added to the OSS. Since the mother of the local tree (NP) now contains the specification \[+top\], it, can only unify with the 'slash' value introduced by the mother of rule (5). Fronting of the NP is achieved in accordance with the FIPs and LP statements.</Paragraph>
    <Paragraph position="17"> Three kinds of PA rules should be distinguished according to the effects of their SBAS. Figure 2 shows two of tl,em; the first one doesn't create structure at, all while the second one transduces a (FAS) local tree into a (GPSG) loi:ai tree. A third type of rules generates GPSG structure out of FAS feature specifications. Figure 1 shows its use to generate the non-local subtree including the perfect auxiliary fs I'v \[hab'l, s(psp)\]\] from the local FAS tree dominated by clauso(perf:+).</Paragraph>
    <Paragraph position="18"> Note that this PA rule must be applied before an attempt is started to attach the subtree fs/np(acc) \[np(nom), v(trans)\]\]. This latter subtree is generated by a PA rule whose pattern rnatches the same FAS tree as the previous one.</Paragraph>
    <Paragraph position="19"> We shall return to this problem in the following section.</Paragraph>
    <Paragraph position="20"> Controlling the ntapl&gt;ing procc.'dure.. First of all note that PA rules can comrnunicate with each other only indirectly, i.e. by modifying the content of the auxiliary storage or by successfully applying an SBA, thereby creating a situation in which another rule becomes applicable (or cannot be applied anymore). PA rules do not contain any control knowledge.</Paragraph>
    <Paragraph position="21"> A local FAS tree is completely verbalized iff a maximum number rt &gt; 1 of applicable PA rules are successful. A PA rule is applicable to a local FAS tree t iff its pattern unifies with t. An applicable PA rule is successful iff all elements of IGA can be executed and an SBA--if present--is successful.</Paragraph>
    <Paragraph position="22"> An SBA is successful iff a syntactic subtree can be attached to the OSS as described above.</Paragraph>
    <Paragraph position="23"> Since the set of PA rules is not commutative, the order of application is crucial in order to ensure that 72 is maximal. Due to the restricted power of the PA rules possible conflicts can be detected and resolved a priori. A conflict arises if more than one pattern matches a given FAS tree. All FAS trees matched by more than one pattern can be identified with help of the FAS grammar. The respective PA rules are members of the same conflict set. The elements of a conflict set can be partially ordered by virtue of precedence rules operating on pairs of PA rules.</Paragraph>
    <Paragraph position="24"> For instance, the conflict regarding the perfect auxiliary is resolved by making a precedence rule check the ID rules that would be invoked by the respective SBAs. If the mother of the second one can be unified with a daughter of the first one and not vice versa, then the first PA rule must be applied before the second one. Thus a PA rule with an SBA invoking ID rule (4) will apply before another one wifll an SBA invoking ID rule (5).</Paragraph>
    <Paragraph position="25"> Note that, in this example, the number of successful PA rules would not be maximal if the order of application was the other way around since the SBA invoking ID rule (4) would not succeed anymore. null The control regime described above guarantees termination, completeness and coherence in the following way: The traversal of a FAS expression terminates since there is only a finite number of local trees to be investigated, and for each of them a I17 finite number of PA rules is applicable. The aSS generated is complete because all local FAS trees are processed and for each a maximum rmmber of PA rules is successful. It is coherent because (1) no PA rule may be applied whose pattern is not matched by the FAS expression and (2) all attachment points nmst be expanded.</Paragraph>
  </Section>
class="xml-element"></Paper>