<?xml version="1.0" standalone="yes"?>
<Paper uid="P90-1026">
  <Title>Asymmetry in Parsing and Generating with Unification Grammars: Case Studies From ELU</Title>
  <Section position="4" start_page="0" end_page="206" type="metho">
    <SectionTitle>
2. Generating with Unification Gram-
</SectionTitle>
    <Paragraph position="0"> mars The goal of employing a single, minimally augmented, grammar for both parsing and generation has become more accessible with the introduction of declaratve grammar formalisms (cf. Kay, 1985). In the context of machine translation, for which the ELU system has been developed, the use of the same grammar for both tasks is highly desirable; indeed much of the work on bidirectional grammars has been carried out in centres working on MT (cf. Busemann, 1987; van Nonrd, to appear;, Dymetmann &amp; Isabelle, 1988; and Wedekind, 1988).</Paragraph>
    <Paragraph position="1"> Regardless of the application, however, the ability to generate with a grammar is extremely useful as a method of checking its adequacy.</Paragraph>
    <Paragraph position="2"> Despite the objective of reversibility, all of the systems mentioned here impose generation-specific restrictions on their grammars, either by limiting the form of possible rules or by augmenting them with annotations. DymeUnann &amp; Isabelle (1988) require the grammar writer to specify for each role the order in which daughters should be generated; however, an order that might be correct when generating from one structure can lead to non-terminating search with another. Busemann (1987) and Saim-Dizier (1989) describe methods of generation which rely on the parsing of a control structure using a specialized grammar to build the syntax of a sentence; it is questionable to what extent the latter two systems can be considered to operate with bidirectional grammars.</Paragraph>
    <Paragraph position="3"> Constraints imposed by Wedekind (1988) and van Noord (to appear) exclude certain linguistic analyses from generation. In order tO overcome the high degree of non-determinism inherent in the top-down approach, Wedekind stipulates that a daughter of a rule must be 'connected' (i.e. that its semantics must be instantiated) before it can be generated from. Less restrictively, van Noord stipulates similar constraints on rules, i.e. that if the semantics of the mother node is known, then the semantics of the head daughter is instantiated, and additionally that if the syntax of the semantic head is known, then the semantics of each daughter is known. These restrictions limit the class of possible analyses, excluding accounts appropriate to LFG (Kaplan and Bresnan, 1982), HPSG (Pollard &amp; Sag, 1987) and UCG (7_eevat et al., 1987).</Paragraph>
    <Paragraph position="4"> The disparate state of progress in parsing and generation raises important issues concerning the adequacy of grammatical descriptions and the computational tools that interpret them. A situation exists in which a grammar may be 'correct' for analysis, but 'incorrect' for generation.</Paragraph>
    <Paragraph position="5"> Significantly, this may be the case even when the restrictions and annotations mentioned above are taken into account. Grammatical analyses developed in a purely parsing environment cannot  always be transferred slraightforwardly into a format suitable for generation. Two types of conclusion may be drawn from this: failures may be ascribed to inadequacies of current generator technology, or the grammatical analyses in question may be re-evaluated. Practical remedies will involve two related strands of research; improving methods of generation so as to IDinimiTe restrictions on the form of grammars that can be generated fzom, and identifying problematic properties of grammars. It is the second of these which the present paper chiefly addresses, though we also remark, in the next section, on some enhancements to the Shieber et al. (1989) algorithm that have been incorporated in the ELU generator.</Paragraph>
  </Section>
  <Section position="5" start_page="206" end_page="206" type="metho">
    <SectionTitle>
3. The Generator in ELU
</SectionTitle>
    <Paragraph position="0"> In this section we describe the generation algorithm in ELU, and discuss in what respects it differs from that described by Shieber et al. (1989). 3 Two notions central to this method of generation are that of the 'pivot', and that of partitioning the grammar intO 'chaining' and 'noD-chaining ' rules. Loosely, the 'pivot' of a structure to be generated from is the lowest node in a path down semantic heads of rules at which the semantics of the current generation root structure remain~ unchanged. A ch~inlng lille is one in which the semantics of the object associated with the right-hand side category that has been declared as the head unifies with that of the left-hand side category. Other rules are non-chaining roles. Rules that apply between the root and the pivot are, by definition, chaining rules; further, any rule which can be attached below the pivot is, by definition, a non-chaining rule. Rules are partitioned into these two groups drain 8 grammar compilaton. null Once the chaining rules have been identifed, the grammar compiler computes the possible sequences of such rules alon 8 a path through their mothers and semantic heads. The result is a 'teachability table', each of whose elements is a pair of restrictor value sets4 representing classes of FSs which can occur at the top and bouom of such a path; in each case, the 'bottom' restrictor set characterizes a pivot. A restrictor set is also computed for each lexical stem, in order to retrieve words efficiently during generation null The generation algorithm uses the distinction between chaining and non-chaining rules as well as</Paragraph>
  </Section>
  <Section position="6" start_page="206" end_page="206" type="metho">
    <SectionTitle>
3 Our discussion will therefore assume familiari-
</SectionTitle>
    <Paragraph position="0"> ty with this paper.</Paragraph>
    <Paragraph position="1"> 4 Restrictors are attributes selected by the writer of a grammar as being maximally distinctive; when two FSs are to be unified, their respective restrictor values axe first checked for compatibility, so as to eliminate the cost of an attempted nnificaton which is bound to fail. See Shieber (1985).</Paragraph>
    <Paragraph position="2"> that between head and non-head dauglzers, the reachability table for chaining rules, the semantic portion of the FS to be generated fi~m 5, and the restrictors for lexicon stems. The algorithm is: 1. Take all grammar rules declared as 'initial' (or all rules in the grammar if no such declaration has been made); for each of these rules whose mother unifies with the input FS, apply the role top-down, building FSs for each of the daughters, and, starting with the head daughter, execute step 2 for each one. If generation firom the daughters is successful, compute all possible word-forms (as constrained by the locally available syntactic information) for each lexical stem  generated.</Paragraph>
    <Paragraph position="3"> 2. Create a pivot COnSisting of just the semantic portion of the current FS. Non-determiniejcally perform steps 2a and 2b: a. Fmd a lexical stem which unifies with the pivot, making sure Coy checking with the reachability table) that the FS resulting from the unification can be linked through semantic heads of just chaining rules up to the current FS.</Paragraph>
    <Paragraph position="4"> b. Fmd a non-chaining rule which can have the pivot as mother, similarly making sure that the FS resulting from the unification of the pivot and the mother can be linked up to the current FS. Recursively (through 2) generate the rule's daughters, starting with the head daughter.</Paragraph>
    <Paragraph position="5"> 3. Link the pivot up to the current FS through  semantic beads of just chaining rules (at each stage, before adding a new rule in the chain, checking with the teachability table that further linking will be possible) and then recursively (through 2) generate the non-bead Os, ghters of these rules.</Paragraph>
    <Paragraph position="6"> In this algorithm non-cbaining roles are used topdown, while chaining rules are used bottom-up. Linking information is used both to check the applicability of a lexical stem or a non-chainlng role when generating top-down from a pivot, and also to control search when generating bottom-up, by ensuring that the left-hand side of any role considered still lies on a possible path through chaining rules to the current FS.</Paragraph>
    <Paragraph position="7"> One innovation of the ELU generator is that the notion 'semantic bead' is interpreted rather differently; whereas the earlier work simply defines the semantic bead of a rule as the daughter whose semantics unifies with that of the left-hand side, and thus leaves the notion undefined for non-chalnlng rules, that described here permits the grammar writer to identify one daughter in each rule as the</Paragraph>
  </Section>
  <Section position="7" start_page="206" end_page="206" type="metho">
    <SectionTitle>
5 The relevant paths being determined by the
</SectionTitle>
    <Paragraph position="0"> user's declaration semantic head. A role in which a O~ghter sluues the semantics of the mother can thus be made into a chaining rule or a non-chaining rule, according to whether that daughter is identified as the semantic head, and a rule that would otherwise have multiple semantic heads can be assigned just one. 6 A rule in which there is no such daughter will remain a non-chaining rule, but may nevertheless be annotated with a similar specification. The rationale is twofold: the ability to coerce what would otherwise be a chaining rule to a non-chaining rule grallts the grammar writer more control over generation, and the ability to specify one daughter as semantic~dly more siPSnlf~mnt than the others may be exploited in order to direct the attention of the generator towards !hat daughter.</Paragraph>
    <Paragraph position="1"> A second difference is the order of events in bottom-up generation. Instead of generating firom the non-head daughters of each chaining rule as it is attached, the pivot is firm linked to the root, so that, if backtracking is forced, effort will not have been spent on processing StrU~h-e that must be discarded. null Finally, on each occasion that top-down generation is initiated, an auempt is made to add a lexical item below the current root, rather than extending the path by application of non-chainlng rules until no such rule is applicable. Here, the motivation is that lexical information may be made available as soon as possible without forcing the grammar writer to adopt analyses that will produce bottom-up generation. This is important because global syntactic properties of a sentence are ofteu determined by lexical information.</Paragraph>
  </Section>
  <Section position="8" start_page="206" end_page="209" type="metho">
    <SectionTitle>
4. Grammars for Generation
4.1. Introduction
</SectionTitle>
    <Paragraph position="0"> In this section we examine more closely interactions between generator and grammar. These fall under two headings: (i) the presence of now deterwini.~m in the grammar, and (ii) the role of lexicalism.</Paragraph>
    <Paragraph position="1"> One aspect of non-detetmini.qm in generation, that of the ordering of role application, is partially overcome in FLU by the user specification of the bead daughter. Non-determinism with respect to the order of solving constraint equations is less well understood. The use of restrictors helps to reduce the number of feature structures to be considered.</Paragraph>
    <Paragraph position="2"> 6 Thus circumventing a problem noted by Shieber et al. (1989, f~4) in connection with such rules. Van Noord (p.c.) stipulates that any daughter which has the same semantics as the mother, but is not the semantic bead, may not branch: this constraint is clearly too strong, precluding, among other things, linguistically motivated accounts of coordination. null  However, in FLU, the use of relational abstractions as a generalization of temj~late facilities increases the problem considerably/Relational abstractions permit the grammar writer to augment the phrase structure rules with statements which may receive multiple definitions in terms of constraint equations; the 'Linear Precedence' definition in (2) below is an example. This facility is a standard ELU device for collapsing what would in an unextended PATR-like formalLqrC/ he several distinct rules, thereby capturing linguistic generalizations that would otherwise go unexpressed.</Paragraph>
    <Paragraph position="3"> It is particularly impoRant to control non-determinism in generation, since, at least when processing is initiated, there is relatively little information available to direct the search. Expanding multiple definitions as they are encountered would give rise to an n~cceptable number of alternatives, many of which might he identical, and often the information from the abstraction is not required until all but one of the alternatives have been excluded by other factors. This is not always the case, however, and when exceptions occur their effect may be drastic. We now describe one such exception to demonstrate how an elegant analysis for parsing is unsuitable for generation.</Paragraph>
    <Paragraph position="4"> 4.2. A grammar for French clitics A common technique in modem lexically-oriented grammars, and one which reflects and extends the traditional notion of 'valency', is to encode informarion about the various phrases with which a verb combines in items on a subcategorization list. The grammar then enforces a match between a member of the list and a phrase which is to combine with some projection of the verb and removes the item from the list. When a sentence is complete, i.e. the verb has 'found' all necessary phrases, a grammar may require that the list he empty, or perhaps that any remaining item is in some way specified as optional. See e.g. Shieber (1986) and Pollard and Sag (1987) for applications of this method.</Paragraph>
    <Paragraph position="5"> A complete grammar of French must account for the position and ordering of clitic pronouns.</Paragraph>
    <Paragraph position="6"> These precede the verb, while other complement phrases follow. Moreover, they appear in a fixed order, as shown in (1): (1) me le lui y en te la leur se les nous vons Up to three clitics may occur, but for the sake of this discussion, we consider only the simpler case 7 Cf. Johnson &amp; Rosuer (1989) for a fuller description of relational abstractions.</Paragraph>
    <Paragraph position="7"> of two critics as complement phrases to the verb. s There are of course many ways of accounting for their distribution; 9 the subcategorization list device seems a natural solution, since any complement phrase may be realized as a critic. The grammar rule in (2) introduces up to two clitics before the verb, their relative order determined by a relational abstraction which is defined by a number of clauses, each clause licensing one of the possible  clitic sequences.</Paragraph>
    <Paragraph position="8"> (2) vplus -&gt; CI1 C12 I-IV</Paragraph>
    <Paragraph position="10"> Some remarks on notation will be helpful: calls to relational abstractions are indicated by the exclamation mark, feature-value disjunction is indicated by the slash, and an equation of the form 'X = Y--Z', where X and Y are lists, nnifies X non-detenninistically with the result of extracting one instance of Z from Y.</Paragraph>
    <Paragraph position="11"> The effect of this rule, then, is to associate a pair of clitics with a verb, checkln~ that they are correctly ordered, and unifying the subcategorization list of the left-hand side category with a copy of that of the head verb from which objects unifying with each of the clitlcs have been removed.</Paragraph>
    <Paragraph position="12"> The problem emerges when information assumed to he held in the subcategorizafion list of 'vplus' is required in order to control further generation. For example, if 'vplus' appears as sister to another complement phrase, and the same procedure of unifying the latter with an item on the list takes place, then because the generator has suspended expansion of non-determini.~tic abstraclions, the subcategorization list itself will he uninstantiated, and therefore no information regarding the semantics of the complement phrase will he available to restrict top-down generation.</Paragraph>
    <Paragraph position="13"> s This is something of an oversimplification, as not only complement phrases, but also adverbials and parts of complement phrases are realized as clitics. See Grimshaw (1982) for a partial LFG account of these phenomena. We also ignore the issue of negation, which considerably complicates the clitic-aux-verb structure.</Paragraph>
    <Paragraph position="14"> 9 The categorial treatment proposed in Baschung et al. (1987) not only makes use of order of arguments, but also codes each clitic for all possible combinations.</Paragraph>
    <Paragraph position="15">  Modifications to the syntactic constituency assumed bere do not affect the principle; as long as the instanfiation of so central an element of the grammar as the subcategorization list is delayed, the problem will remain. An alternative type of analysis would remove the non-determinism from the grammar by factoring it out into a larger nomber of rules. This solution is not without its own disadvantages; the number of distinct rules needed by a full treatment of French critics, integrated with the placement of the various negative panicles and auxiliaries, should not be underestimated. We postpone further discussion of nondetermini.~m and delay until the conclusion and turn now to the problem of empty semantic heads, an important problem for bead-driven generation algorithms. 1o</Paragraph>
    <Section position="1" start_page="208" end_page="209" type="sub_section">
      <SectionTitle>
4.3. Empty Semantic Heads
</SectionTitle>
      <Paragraph position="0"> In German and Dutch, there are two positions in a sentence where tensed verbs may appear: in second position of a main clause, and in final position of a subordinate clause. Once again, a multitude of analyses are possible within ELU grammars. One approach is to control the distribution of verbs with grammar rules specific to clause-type; this solution gives rise to what might be felt to be an unacceptable degree of duplication in the grammar. A more elegant approach, successful for parsing, exploits the possibility of assoc/ating a word or phrase appearing in one position within a sentence with a 'gap' elsewbere.</Paragraph>
      <Paragraph position="1"> The latter analysis will be recognized as a variant of a standard Govermnent-Binding treatment, in which a tensed verb in a main clause is 'raised' from an 'underlying' sentence-final position to a 'surface' second position (see e.g. Haider (1985), Platzack (1985) for discussion of this class of analyses). The dependency may be implemented by the use of a feature, say 'v2', whose value in a verb-second construction is a feature structure representing the verb to be raised, and in other constmctions an atomic constant such as 'none', which serves to block the dependency. At the extraction site, any value of 'v2' other than 'none' may be cashed out as an empty production. Information regarding the various syntactic properties of the raised verb is passed in the normal fashion between the verb's true position and the extraction site, wbere it is able to exert the same constraints upon complement phrases that a lexically-realiTed verb would.</Paragraph>
      <Paragraph position="2"> The simplified rule set given in (3) will serve as a basis for discussion. Recall that the generator operates by partitioning the rules of the grammar 1o This problem is alluded to in Shieber et al.</Paragraph>
      <Paragraph position="3"> (1989, fn.4) and is discussed in a draft of an expanded version of the paper.</Paragraph>
      <Paragraph position="4">  into classes to be applied top-down (non-ch~inlng rules - here 'S-gap' and 'V2') and bottom-up (chaining rules - here 'TOP', 'S' and 'V').</Paragraph>
      <Paragraph position="5"> Bottom-up generation is only practical if the input structure to that phase of generation contains sufficient information, e.g. the verb with its sub-categorization list.</Paragraph>
      <Paragraph position="6">  The verb-raising analysis sketched here has the unfortunate property of supplying the generator with a semantic bead (the verb gap) about which nothing is known. At the stage when top-down processing has identified the verb gap as the starting point fog boUom-up generation, the input featm'e structure is underspecified. In particular, the subeategorization list of the missing verb is -ninstalltiated, and in the grammar in question, it is the length of this list which controls invocation of the recumive role 'S'. No bindings can be found, and the generator suspends evaluation of that equation in the hope, in-founded on this occasion, that information not yet present will later allow its solution. The result is that'S' is repeatedly added above 'S-gap', in a non-termlnating attempt to ensure completeness of the search.</Paragraph>
      <Paragraph position="7"> Van Noord (1989) describes two solutions to this problem, both of which are additions to the original program, and whose only motivation (so far) is to overcome this specific problem. The first, somewhat ad-hoc, solution allows the verb to have as one of its morphological realizations the empty string. Since word forms are generated at the end of processing by a morphological front-end, the generator can posit the same word in both positions (for the purpose of relrieving its subcategorizafion behaviour f~om the lexicon, for example). The morphological component then generates one empty string and one full word according to the position of the verb (i.e. in a main or subordinate clause). &amp;quot;\['nis mechani.~n is not available in ELU. The second solution adds an additional 'connect' clause in the Prolog program, specific to gaps, in order to assure that the gap is first instanfiated before further processing; this solution raises the issue of I~ming programs to treat specific problems as they are encountenxL There are other constructions which raise the same kind of problem; the fronting of apparently non-constitnent verbal sequences in German (Nerboone, 1986) introduces more complex dependencies, while in English the phenomena of Gapping and Verb-Phrase Ellipsis both manifest themselves syntactically in the absence from a sentence of a verb and possibly other material. Here, the difficulty is, if anything, greater, as the dependencies in question are anaphoric in nature, rather than syntactic.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>