<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2019">
  <Title>Generating French with a Reversible Unification Grammar</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
O. Intr~cluction
</SectionTitle>
    <Paragraph position="0"> In this paper, we describe the linguistic solutions to some of the problems encountered in writing a reversible French grammar. This grammar is primarily intended to be one of the components of a machine translation system built using ELU, 1 an enhanced PATR-II style unification grammar linguistic environment based on the LID system described in Johnson and Rosner (1989), but it is also part of our more general experimentation with fully reversible grammars.</Paragraph>
    <Paragraph position="1"> The requirement that it be reversible imposes a stringent criterion of linguistic adequacy on a grammar, siuce it is not allowed to overgenerate while it must at the same time provide a large coverage for analysis (Dymetman and IsabeUe (1988)). Formally, grammars that are fully reversible must be completely declarative, since no reierence can be made in the grammar rules to the process (analyzer or synthesizer) which will use them. The unification formalism makes itt possible to write such grammar statements, because due to the associativity and commutativity of the unitication operation, the result of unifying feature structures is independent of the order in which they are unitied (Appelt (1989)).</Paragraph>
    <Paragraph position="2"> Writing reversible grammars, however, presents problems which do not arise in the traditional grammars used for either analysis or generation. In addition, the progress accomplished recently in building generators for unification grammars has already revealed some of the problems posed by unification-based reversible grammars. 2 As shown by Russell et al. (1990), even though the grammar rules do not refer to the generation process, the generation algorithm imposes particular constraints on the grammar formalism. 3 This paper concentrates particularly on the problems encountered in the generation of French, specifically in the analysis to be given to clitics.</Paragraph>
    <Paragraph position="3">  one pre~;nted in Saint-Dizier (1989), since his grammar is neither reversible nor purely declarative, as the rules are annotated with ' generation points'.</Paragraph>
    <Paragraph position="4"> We first briefly describe the aspects of the generation algorithm and of the grammar formalism which are relevant to the particular problems under discussion, then present the facts of French syntax which pose those problems and the solutions we have adopted.</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1. The Generator
</SectionTitle>
    <Paragraph position="0"> The generation algorithm of ELU is based on the algorithm described in Shieber et al. (1989) and was developed at ISSCO by J. Carroll. 4 Generation is head-driven: each role has a &amp;quot;semantic head&amp;quot; (see Shieber (1988)), which is specified by the ~ammar writer, and the head daughter of a rule is generated before its siblings. The depth-first algorithm defines a downward path through the semantic heads of rules.</Paragraph>
    <Paragraph position="1"> This algorithm does not require that the grammar be semantically monotonic. Non-monoto\]tficity is obtained by having the generator distinguish two types of rules in the grammar, &amp;quot;chaining&amp;quot; arKl &amp;quot;nonchaining&amp;quot; rules, and by introducing the notion of &amp;quot;pivot&amp;quot;. Following from this distinction, it employs both bottom-up and top-down processing.</Paragraph>
    <Paragraph position="2"> The partition of the set of grammar rules into chaining and non-chaining rules is pre-compiled from the specification of what counts as the &amp;quot;semantics&amp;quot; of a feature structure. In a chaining rule, the mother and the head daughter have identical semantics; chaining rules are used bottom-up from the &amp;quot;pivot&amp;quot;, which is defined as the lowest point in the path through the head daughters of chaining rules at which the semantics of the feature structure remains unchanged. In a non-chaining rule, the mother and the head daughter have different semantics; non-chaining rules are used top-down from the pivot.</Paragraph>
    <Paragraph position="3"> The efficiency of the ELU generator depends in a large part on the restrictors defined by the grammar writer. Computing the pivot, i.e. creating a teachability table for chaining rules, and bottom-up processing 4 Cf. Russell et al. (1990) for a description of the differences between the two algorithms.</Paragraph>
    <Paragraph position="4"> 106 1 are both controlled by pre-compiled &amp;quot;linking&amp;quot; inforo mation, which is encoded as sets of restrictor values.</Paragraph>
    <Paragraph position="5"> A restrictor is a specification of a value wlfich can be computed from a feature structure (syntactic category, for example, is often defined as a resUictor). Before attempting unification between two feature structures, the values tbr the restrictors are checked in both of them; if these values are not compatible, unitication would be bound to fail mid is not tried. As linking information is only relevant for chaining rules, it is only used bottom-up during processing, ~md since by definition, chaining rules have the same semantics for their heads, linking information must be syntactic.</Paragraph>
    <Paragraph position="6"> Restrictors are also used heavily m the selection of lexical items, so the attributes chosen as restrictors have to be good discriminauts between i~ature structures. 5 &amp;quot;Ihe generation algorithm by itself guarantees neither the completeness nor the coherence of tile resulting feature structure. The responsibility tbr preventdeg ing the generation of structures which unify with the input, bu~ me incomplete (i.e. ensuring completeness) rests with the grmnmar writer: any structure which needs to be generated in its entirety should not be represented as an uncotx~trained ti~ature structure, but must be specified as another data type, i,e. a list, a tree, or a user-defined type expression. Tt~e graulmar writer and the generator share tile responsibility for preventing additions to the input structure (i.e.</Paragraph>
    <Paragraph position="7"> preserving coherence): the gramm~ writer must again select the appropriate data types, and the generator &amp;quot;tTcezes&amp;quot; uninstantiated variables that occur in the input.</Paragraph>
    <Paragraph position="8"> The choice of appropriate data types as well as of good restrictors is therclore crucial to ensure flint the grammar is not only efficient but usable in generation.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. The G, rammar Formalism.
</SectionTitle>
    <Paragraph position="0"> The syntactic ieplesentations built by file parser are trees where each node is a directed acyclic graph consisting of atuibute-value pairs (i.e. a feature struc~ ture which allows reentrancy). 'File semantic represemations used ~s input by the generator are feature structures derived from the syntactic trees.</Paragraph>
    <Paragraph position="1"> The gr~unmar rules consist of context-free phrase structure rules annotated with corrstraint equatiotrs expressing relations between the categories mentioned ill the rule. The ELU tormalism provides a generalization of the template facility of PATR-II, the &amp;quot;relational abstractions&amp;quot;, which are statements abstracting over sets of constraint equations. These statements 5 Restrictors are also used to restrict the search space in parsing (see Shieber (1985)). &amp;quot;fbe use of linking information in generation was first proposed by van Noord (1988).</Paragraph>
    <Paragraph position="2"> may receive multiple and mcursive definitions. To give multiple definitions to a relational abstraction permits collapsing what i~l an unextended PATR-Iike formalism would be several distinct rules, and is a powerful way to capture linguistic generalizations.</Paragraph>
    <Paragraph position="3"> Multiple definitions, however, give rise to a high degree of non-determinism during processing. Therefbre, while the parser expands multiple definitions whenever they are encountered, the generator uses a lazier approach and only expands them when they are needed. Nevertheless, tiffs strategy is not sufficient, and the problem posed by the non-determinism of relational abstractions is the most complex and severe of the grammar/generator interactions described in Russell et al. (1990), because of its adverse effects on the restliclion of top-down generation.</Paragraph>
    <Paragraph position="4">  3. French Critics Any French gr~wnmar must account for the position aid ordering of preverbal clitics. While full complement aud modifier ptwases occur to tile right of the mah~ verb of a clause, up to three elitics may occur in front of a verb, as in (1).</Paragraph>
    <Paragraph position="5"> (1) 11 m'y en a fait p,'u-t.</Paragraph>
    <Paragraph position="6">  he me there of it informed He infot~ned me of it there.</Paragraph>
    <Paragraph position="7"> Moreover, tile clitics must appear in a fixed order, which, as shown ill (2), is independent of the semantics of the critics.</Paragraph>
    <Paragraph position="8"> (2) a. Ils vous l'y ont dolm~e.</Paragraph>
    <Paragraph position="9"> they to you it there gave They gave it to yott there.</Paragraph>
    <Paragraph position="10"> b. *Ils leur l'y ont dotm6e.</Paragraph>
    <Paragraph position="11"> they to them it there gave 77wy gm, e it to them there.</Paragraph>
    <Paragraph position="12"> This f~ed order can be represented by the traditional table given in (3). 6  (3) Ordering of French ditics me le lui y en  te la leur se les nous vous in most accounts of the distribution shown in (3), tile problem is simplified, because only subcategolized complements are de'tit with. A French preverbal elitic, however, is not necessarily a subcategorized complement of the verb; adverbials and parts of complement phrases can also cliticize, and the grammatical category of some clitics is that of adverbs or 6 In (3), se stands for any of the so-called 'R-clitics', i.e. the reflexive and reciprocal pronouns, as well as the inherent reflexive ~md the middle marker, as explained in more detail below.</Paragraph>
    <Paragraph position="13"> 2 107 quantifiers.</Paragraph>
    <Paragraph position="14"> The contrast between (4.a) and (4.b) shows that a clitic can be either a full complement, or part of a complement. In (4.a), en is the full prepositional object of the verb parler, while in (4.b), en represents the partitive prepositional phrase which is the complement of the object of vouloir.</Paragraph>
    <Paragraph position="15"> (4) a. I1 en parlait souvent.</Paragraph>
    <Paragraph position="16"> he often talked about it \[cf. I1 parlait souvent de ce Uvre\] \[he often talked about that book\] b. J' en veux deux.</Paragraph>
    <Paragraph position="17"> I want two of them \[cf. Je veux deux de ces pommes\] \[I want two of these apples\] The contrast between (5.a) and (5.b) shows that a elitic can either be subcategorized or not. In (5.a), y is the subcategorized complement of the verb aller, while in (5.b), y is a locative adverb, which is not sub-categorized by the verb dormir.</Paragraph>
    <Paragraph position="18"> (5) a. I1 y aUait souvent.</Paragraph>
    <Paragraph position="19"> he often went there \[cf. I1 allait souvent dans cette ville\] \[he often went to that city\] b. I1 y dormait souvent.</Paragraph>
    <Paragraph position="20"> he often slept there.</Paragraph>
    <Paragraph position="21"> \[cf. I1 dormait souvent dans cet hdtel\] \[he often slept in that hotel\] Besides the personal pronouns and the adverbs given in Table (3), there are other lexical items which are not usually considered in the treatment of French clitics, but whose behavior is closely related. 7 The negative elements pas, plus, jamais, rien, and the quantifiers tant, autant, plus, moins and tout 8 also cliticize and may appear preverbally.</Paragraph>
    <Paragraph position="22"> While ,all the clifics of Table (3) must appear in front of the traditional AUX constituent (i.e. before any of the verbal elements of the VP), the examples in  (6) slhow that the elements of this second set appear inside AUX, more precisely after the first tensebearing verbal form.</Paragraph>
    <Paragraph position="23"> (6) a. I1 n'en avaitjamais 4t6 persuad6,  he had never been sure of it b. I1 n' en avait jamais rien cru.</Paragraph>
    <Paragraph position="24"> he had never believed any of it c. Je n'y en ai jamais autant vu.</Paragraph>
    <Paragraph position="25"> I had never seen so many of them there 7 They are, however, the subject of work in theoretical linguistics, see e.g. Perlmutter (197l), Emonds (1975), Kayne (1975), and more recently Pollock (1989). Interestingly, though it was developed in a different framework and for different reasons, our treatment of those elements is compatible with the latter's analysis (cf. also fn.9).</Paragraph>
    <Paragraph position="26"> 8 The quantifier tout has actually several forms, inflected for gender and number: tout/tousltoute/toutes. There are thus at least two slots for clitics inside a French VP, and neither of these slots correlates with argumenthood. The quantifiers rien and autant which appear inside AUX in (6.b) and (6.c) are (paJtts) of the argument of the verbs croire and voir, and so is the quantifier en, which is in front of the AUX. On the other hand, the adverb y is not an argument in (6.c), nor is jamais in (6.a-c).</Paragraph>
    <Paragraph position="27"> Therefore, the lexical entry of every clitic element must specify not only that it is a clitic but whether it appears in front of or inside the AUX constituent.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4. Generation
</SectionTitle>
    <Paragraph position="0"> Theoretically, the fundamental problem posed by clitics stems from their dual nature, syntactic' and morphological, and partly consists in deciding whether to treat them by syntactic or by morphological processes. 9 Descriptively, there are three issues to be addressed: argument-binding, linear ordering relatior~s, and categorial status of the clitics. All three give rise to problems in generation due to nondeterminism, for which the solution is to ensure that the lexical verb is instantiated as soon as possible.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1. Subcategorizatlon
</SectionTitle>
      <Paragraph position="0"> The unification formalism makes it very natural to encode syntactic information in the lexicon and with a lexicalist approach, our treatment of arguments is straightforward: we make the standard use of a sub-categorization list to encode the complements a verb requires. Since any complement phrase may be realized as a clitic, this fact is not mentioned in the sub-categorization list. to  9 E.g., restrictions on coordination show that clitics are not independent syntactic constituents.</Paragraph>
      <Paragraph position="1"> (i) * I1 me et te connait.</Paragraph>
      <Paragraph position="2"> he knows me and you  Cf. the various analyses presented in Borer (1986). More recently (Rizzi and Roberts (1989), Kayne (1990)), the question has been reformulated in terms of the type of mechanism (adjunction or substitution) involved in cliticization and of whether clitics are phrasal heads or not. With the lexicalist approach adopted in our grammars both types of processes can be referred to in the lexicon, but it is of course still desirable that the two be clearly separated.</Paragraph>
      <Paragraph position="3"> l0 This analysis contrast with that of Baschung et al. (1987), or B6s et al. (1989), which treats separately complements appearing to the left and complements appearing to the right of the verb. Their reason for doing so is that they take the variants shown in (i) and (ii) to indicate a relatively free order of complements (sulx-ategorized or not) in French. (i) il a donn6 (hier) un livre ~t Marie (hier).</Paragraph>
      <Paragraph position="4"> yesterday he gave a book to Marie (ii) il a donn6 (hier) ?l Marie un livre (bier).</Paragraph>
      <Paragraph position="5"> yesterday he gave a book to Marie While the ordering of full complements inside the VP poses some problems for generation, it is a separate question from that of cliticization, and the two should receive principled solutions of their own.</Paragraph>
      <Paragraph position="6"> 108 3 During analysis, art element found in the VP is checked against Subcat, the subcategorization list of the predicate. If it does not unify with any element of Subcat, ~t is treated as a VP modifier and added to Mods, the list of modifiers. From the point of view of generation, clitics realize elements from either the Subcat list or from the Mods list.</Paragraph>
      <Paragraph position="7"> For instance, we partly follow the lexicalist analysis of Grimshaw (1982) for the R-clitics represented by se. That is, we consider that the R-clitic is not an argument of &amp;quot;iuheienfly reflexive&amp;quot; (7.a,b) and &amp;quot;middle&amp;quot; verbs (7.c), but a morphosyntaclic marker, t i  (7) a. i~1 s'est 6vanoui.</Paragraph>
      <Paragraph position="8"> he fainted b. I1 se le demandait.</Paragraph>
      <Paragraph position="9"> he was wondering about it c. II s'est cassd.</Paragraph>
      <Paragraph position="10"> it broke But in reciprocal and true reflexive constructions, such as (8.a,b), we treat the R-clitic as a pronoun which is an argument of the verb. 12 (8) a. Ils se sont regardds.</Paragraph>
      <Paragraph position="11"> they looked at each other~themselves b. Ils se les sont donnds.</Paragraph>
      <Paragraph position="12"> they gave them to each other~to themselves  Thereibre, because the verbs in the examples of (7) are marked in the lexicon as being iitherently reflexive, an R-clitic is generated from Subcat without being bound to the list of semantic arguments. In (8) on the other hand, the verbs are respectively transitive and ditransitive: in their case, a semantic argument is both bound to ,an element of Subcat and re,alized as a reflexive pronoun because of its own semantic features. In (9.a-c) se is, as in (7), the inherent reflexive marker mid is generated from Subcat. In  (9.a) en is the partitive phrase of a subcategorized argument; y in (9.b) is a subcategorized locative argument from Subcat and in (9.c), it is a VP adverb from Mods.</Paragraph>
      <Paragraph position="13"> (9) a. I1 s'en est cass6 deux.</Paragraph>
      <Paragraph position="14"> two of them broke b. Ils s'y trouvaient.</Paragraph>
      <Paragraph position="15"> they were there c. lls s'y vendaient.</Paragraph>
      <Paragraph position="16">  they were soM there As described in Russell et at. (1990), problems arise iin generation because of non-determinism and becau~e of the unavailability of some syntactic information to the generator. The subcategorization list II In (7.a), there is no non-reflexive verb e~'anouir, and in (7.b), die reflexive verb has a different semantics than the non-reflexive verb from which it is lexically derived. 12 In this respect, our analysis also differs from that presented in Wehrli (1986).</Paragraph>
      <Paragraph position="17"> mechanism typical of unification grammars is a source of both these kinds of problems. Subcategorization lists are relational abstractions with multiple definitions; therefore, they introduce non-detenninism in tile expansion of the rules in which they are invoked. Moreover, they exemplify the type of syno tactic information typically found ill lexical entries; tiffs infommtion is not available to the generator until the lexical head has been instantiated, but if it was avMlable at a higher point in the path through the rules it would help constrain the top-down search.</Paragraph>
      <Paragraph position="18"> In particular, here, separating the elements found inside the VP into arguments and modifiers can only be (lone alter the lexical head has been instantiated mid its subcategorization list is available. As shown by the two meanings of the verb trouver given in the lexical entries (10.a,c) and exemplified in (10.b,d), the semantics of the verb (its argument list) may change according to its subcategorization list.</Paragraph>
      <Paragraph position="20"> he finds it there \[cf. II le trouve dans les Alpes.\] \[he finds it in the Alps\]</Paragraph>
      <Paragraph position="22"> d. I1 s'y trouve.</Paragraph>
      <Paragraph position="23"> it is located there \[cf. I1 se trouve dans les Alpes.\] \[it is located in the Alps\] In (10.d) the clitic y is ,an argument (i.e. it is bound to one of the variables in the arguments list), while in (10.b) it is not (i.e. it is added to the modifiers list). Even though the two possiblities ,are mutually exclusive, if the subcat list is not available at the VP level, the search must proceed top-down and tim VP is expanded top-down and non-deterministically. Recall that when the semantics for the head daughter of a rule does not change, the rule is a chaining rule which is used bottom-up, but if the semantics of the head changes, then the rule is a non-chaining rule, which is used top-down and defines a pivot.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2. Linear ordering
</SectionTitle>
      <Paragraph position="0"> As was shown by the examples of (2), the linear ordering among preverbal clitics is independent of their semantics; it is also independent of the syntactic features of their dominating clause, i.e. negation, inversion, etc, A perspicuous way to express citric ordering is to have one relational abstraction with separate definitions stating the different precedence constraints holding between two preverbal clitics.</Paragraph>
      <Paragraph position="1"> The simplified definitions for Precede(C1,C2) given in (11) would account for most of the distribution</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.3. Categorial status
</SectionTitle>
      <Paragraph position="0"> A characteristic property of clitics is that they do not have a maximal projection and remain X deg constituents, with their own syntactic category feature coming from the lexicon. To express the fact that a dative pronoun or the clitics y and en actually stand for a PP can be done by building a PP in the lexicon, e.g. with a relational abstraction such as Make-PP(CI,PP).</Paragraph>
      <Paragraph position="1">  The relational abstractions Precede and Make-PP constitute an elegant collapsing of syntactic and lexical rules which is useful in analysis: the grammar rules which rewrite VPs containing clitics need not specify all the various possibilities. However, as with the relmional abstractions encoding subcategorization facts, its multiple definitions render Precede nondeterministic. The non-determinism of Make-PP, which is due to the fact that some clitic forms are ambiguous, is no less severe. During generation, the evaluation of the equations is delayed until the semantics for the head has been instantiated, and if the lexical head is not instantiated early enough, rules which involve these relational abstractions are tried repeat13 There are other constraints not accounted for by (11), e.g. the one requiring that an ambiguous acc/dat form cannot be interpreted as an accusative in front of a dative: (i) ,k Elie nous lui prdsentera.</Paragraph>
      <Paragraph position="2"> she us to him will introduce Similar constraints exist among the clitic elements appearing in post-verbal position.</Paragraph>
      <Paragraph position="3"> edly even if they cannot apply.</Paragraph>
      <Paragraph position="4"> In conclusion, for the purpose of generation, we need an analysis where the semantic head of the VP is not necessarily the lexical main verb, but is the element which will be sure to be instantiated as early as possible. In an analysis reminiscent of current work in the Government-Binding framework, 14 where a clause is IP (Inflectional Phrase), the maximal projection of INFL, we take as the semantic head for our rules the element which bears tense. This element, I, may be either the main verb or an auxiliary which takes the main verb as complement.</Paragraph>
      <Paragraph position="5"> With this analysis of the structure of VP, the semantics of the head daughter I remains the same along the path through the semantic heads so that the pivot of the structure, i.e. the point at which bottom-up generation can start from, is at the end of path. At that point, either I is the main verb (V-raising has applied) and it can be instantiated immediately, or I is an auxiliary (V-raising hasn't applied) and the main verb is its sister, which can be reached through other chain rules.</Paragraph>
      <Paragraph position="6"> We can deal with clitics in two ways:  Besides being descriptively more adequate, since the ordering constraints hold between the clitics themselves, not between a clitic and a verbal constituent, the second approach is to be preferred because a list ensures completeness of the resulting feature structure. Moreover, the whole list of clitics can he built without instantiating the lexical verbal head. With the two clitic positions and taking I as the head, the syntactic structure for a VP is as in (15).</Paragraph>
      <Paragraph position="7">  Clitic elements are marked as to whether they must appear to the left or to the fight of I. If V-raising hasn't applied, as in the examples of (6), the two critic lists will be in front of the main verb, on either side of I. If V-raising has applied, the two clitic fists will still be on either side of I, and of the main verb, as in (16). (16) a. I1 ne l'en persuaderajamais.</Paragraph>
      <Paragraph position="8"> he will never convince her of it b. I1 n' en croit jamais rien.</Paragraph>
      <Paragraph position="9"> he never believes any of it c. Je n ' y en voit jamais autant.</Paragraph>
      <Paragraph position="10"> I never see so many of them there</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. Conch~slon
</SectionTitle>
    <Paragraph position="0"> We have shown with die example of French clitics how some problems inherent in the writing of reversible grammars arise, ~md what aspects of the formalism are responsible for them. The solutions we propose are motivated by internal considerations ,and provides a coherent syntactic account of the phenomena under consideration, i.e. clitic placement and so-called &amp;quot;adverb climbing&amp;quot; (although space prevents us from showing tile details here, riley also deal adequately with vefibal negation). These solutions make full use of the properties and adwmtages of die lexicalist approach to gr,'unmars while circumventing (some of) the dange~ it presents.</Paragraph>
    <Paragraph position="1"> ** I ,'un grateful to Susan Warwick and Graham Russell for the time they have spent helping me understand EHJ and its generator. Neither of them, of course, is responsible for any mistake in this paper.</Paragraph>
  </Section>
class="xml-element"></Paper>