<?xml version="1.0" standalone="yes"?>
<Paper uid="E87-1027">
  <Title>STRING-TREE CORRESPONDENCE GRAMMAR: A DECLARATIVE GRAMMAR FORMALISM FOR DEFINING THE CORRESPONDENCE BETWEEN STRINGS OF TERMS AND TREE STRUCTURES YUSOFF ZAHARIN Groupe d'Etudes pour la Traduction Automatique</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
STRING-TREE CORRESPONDENCE GRAMMAR: A DECLARATIVE GRAMMAR FORMALISM FOR DEFINING
THE CORRESPONDENCE BETWEEN STRINGS OF TERMS AND TREE STRUCTURES
YUSOFF ZAHARIN
</SectionTitle>
    <Paragraph position="0"> Groupe d'Etudes pour la Traduction Automatique B.P. n deg 68 Universit~ de Grenoble</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
38402 SAINT-MARTIN-D'HERES
FRANCE
ABSTRACT
</SectionTitle>
    <Paragraph position="0"> The paper introduces a grammar formalism for defining the set of sentences in a language, a set of labeled trees (not the derivation trees of the grammar) for the representation of the interpretation of the sentences, and the (possibly non-projective) correspondence between subtrees of each tree and substrings of the related sentence. The grammar formalism is motivated by the linguistic approach (adopted at GETA) where a multilevel interpretative structure is associated to a sentence. The topology of the multilevel structure is 'meaning' motivated, and hence its substructures may not correspond projectively to the substrings of the related sentence.</Paragraph>
    <Paragraph position="1"> Grammar formalisms have been developed for various purposes. Generative-Transformational Grammars, General Phrase Structure Grammars, Lexical Functional Gr-mmar, etc. were designed to be explanatory models for human language performance, while others like the Definite Clause Grammars were more geared towards direct interpretability by machines.</Paragraph>
    <Paragraph position="2"> In this naper, we introduce a declarative grammar formalism for the task of establishing the relation between on one hand a set of strings of terms and on the other a set of structural representations a structural representation being in a form amenable to processing (say for translation into another language), where all and only the relevant conten~.s or 'meaning' (in some sense adequate for the purpose) of the related string are exhibited. The grammar can also be interpreted to perform analysis (given a string of terms, to produce a structural representation capturing the 'meaning' of the string) or to perform generation (given a structural representation, to produce a string of terms whose meaning is captured by the said structural representation).</Paragraph>
    <Paragraph position="3"> It must be emphasised here that the grammar writer is at liberty (within certain constraints)to design the structural representation for a given string of terms (because its topology is independent of the derivation tree of the grammar), as well as the nature of the correspondence between the two (for example, according to certain linguistic criteria). The grammar formalism is only a tool for expressing the structural representation, the related string, and the correspondence.</Paragraph>
    <Paragraph position="4"> The formalism is motivated by the linguistic approach (adopted at GETA) where a multilevel in~rpretative structure is associated to a sentence.</Paragraph>
    <Paragraph position="5"> The multilevel structure is 'meaning' motivated, and hence its substructures may not correspond projectively to the substrings of the related sentence The characteristic of the linguistic approach is the design of the multilevel structures, while the grammar formalism is the tool (notation) for expressing these multilevel structures, their related sentences, and the nature of the correspondence between the two. In this paper, we present only the grammar formalism ; a discussion on the linguistic approach can be found in \[Vauquois 78\] and \[Zaharin 87\].</Paragraph>
    <Paragraph position="6"> For this grammar formalism, a structural representation is given in the form of a labeled tree, and the relation between a string of terms and a structural representation is defined as a mapping between elements of the set of substrings of the string and elements of the set of subtrees of the tree : such a relation is called a string-tree correspondence. An example of a string-tree correspondence is given in fig. I.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="165" type="metho">
    <SectionTitle>
TREE: NP
</SectionTitle>
    <Paragraph position="0"/>
    <Paragraph position="2"> The example is taken from \[Pullum 84\] where he called for a 'simple' grammar which can analyse/ generate the non-context free sublanguage of the African language Bambara given by :</Paragraph>
    <Paragraph position="4"> and at the same time the grammar must produce a 'linguistically motivated' structural representation for the corresponding string of words. For instance, the noun phrase &amp;quot;dog catcher hunter o dog catcher hunter&amp;quot; means &amp;quot;any dog catcher hunter&amp;quot; and so the structural representation should describe precisely that.</Paragraph>
    <Paragraph position="5">  In the string-tree correspondence in fig. I, there are three concepts involved : the TREE which is a labeled tree taking the role of the structural representation, the STRING which is a string of terms, and finally the correspondence which is a mapping (given by the arrows ~--.-.~&gt;) defined between substrings of STRING and subtrees of TREE (a more formal notation using indices would be less readable for demonstrational purposes). In the TREE, a node is given by an identifier and a label (eg. |:NP). To avoid a very messy diagram, in fig. l we have omitted the other subcorrespondence between substrings and subtrees, for example between the whole TREE and the whole STRING (trivial), between the subtree 4(5(6),7) and the two occurrences of the substring &amp;quot;dog catcher&amp;quot; (nontrivial), etc. We shall do the same in the rest of this paper. (Then again, this is the string-tree correspondence we wish to express for our examples - recall the remark earlier saying that the grammar writer is at liberty to define the nature of the string-tree correspondence he or she desires, and this is done in the rules, see later).</Paragraph>
    <Paragraph position="6"> We also note that the nodes in the TREE are simply concepts in the structural representation and thus the interpretation is independent of any grammar that defines the correspondence (in fact, we have yet to speak of a grammar) ; for instance, the TREE in fig. 1 does not necessitate the presence of a rule of the form &amp;quot;AP NT hunter ~ NP&amp;quot; to be in the grammar.</Paragraph>
    <Paragraph position="7"> A more complex string-tree correspondence is given in fig. 2 where we choose to define a structural representation of a particular form for each string in the language anbnc n. Here, the case for n=3 is given* The problem is akin to the 'respectively' problem, where for a sentence like &amp;quot;Peter, Paul and Mary gave a book, a pen and a pencil to Jane, Elisabeth and John respectively&amp;quot;, we wish to associate a structural representation giving the 'meaning' &amp;quot;Peter gave a book to Jane, Paul gave a len to Elisabeth, and Mary gave a pencil to John&amp;quot;.</Paragraph>
    <Paragraph position="9"> correspondence for a~bnc n At this point, again we repeat our earlier statement that the choice of such structural representations and the need for such string-tree correspondence are not the topics of discussion in this paper.</Paragraph>
    <Paragraph position="10"> The aim of this paper is to introduce the tool, in the form of a grammar formalism, which can define such string-tree correspondence as well as be interpretable for analysis and for generation between strings of terms and structural representations. The grammar formalism for such a purpose is called the String-Tree Correspondence Grammar (STCG). The STCG is a more formal version of the Static Grammar developed by \[Chappuy 83\] \[Vauquois &amp; Chappuy 85\]. The Static Grammar (shortly later renamed the Structural Correspondence Specification Grammar), was designed to be a declarative grammar formalism for defining linguistic structures and their correspondence with strings of utterances in natural languages. It has been extensively used for specification and documentation,as well as a (manua$ reference for writing the linguistic programs (analysers and generators) in the machine translation system ARIANE-78 \[Boitet-et-al 82\]. Relatively large scale Static Grammars have been written for French in the French national machine translation project \[Boitet 86\] translating French intoEnglis~ and for Malay in the Malaysian national project \[Tong 86\] translating English to Malay ; the two projects share a common Static Grammar for English (naturally). The STCG derives its formal properties from the Static Gra~mmar, but with more formal definitions of the properties. In the passage from the Static Grammar to the STCG, the form as well as some other characteristics have undergone certain changes, and hence the change to a more appropriate name. The STCG first appeared in \[Zaharin 86\], where the formal definitions of the grammar are given (but under the name Of the Tree Correspondence Gran~nar).</Paragraph>
    <Paragraph position="11"> A STCG contains a set of correspondence rules, each of which defines a correspondence between a structural representation (or rather a set or family of) and a string of terms (similarly a set or family of). Each rule is of the form : .Rule: R CORRESPONDENCE:</Paragraph>
    <Paragraph position="13"> The simplest form of such a rule is when al,...a n are terms and B is a tree. The rule then states that the string of terms ~l,...,ctn corresponds (&amp;quot;) to the tree B, while the entry cORRESPONDENCE gives the substring-subtree correspondence between the terms ~i, ,~_ and the subtrees BI,...,B_ of B. An</Paragraph>
    <Paragraph position="15"> defined by Sl Although in the example in fig. 3 above, the leaves of the TREE are labeled and ordered exactly as the terms in the STRING, this is not obligatory. For example, it is indeed possible to change the label of node 2 to something else, or to move the node to the right of node 4, or even to exclude the node altogether. In short, the string-tree correspondence defined by a rule need not be projective.</Paragraph>
    <Paragraph position="16"> Such elementary rules el...cz ~8 (with ul,..,u_ terms) can be generahsed to a form where each e.&amp;quot;(i-l,..,n) represents a string of terms, say A.. I Here, generalities can be captured if u i spec~'~ies the name of a rule which defines a strlng-A.~T. tree correspondence--i I (for some tree T. given in the said rule, but it is of httle slgnlflcance here), in which case the interpretation of the string-tree correspondence defined by el..e ~8 is taken to be AI..A ~8 (here AI..A means thenconca tenation of ~he s~rings Al,?. ,A-~. The substring-subtree correspondence will sti--~l be given by the entry CORRESPONDENCE. Fig. 4 illustrates this.</Paragraph>
    <Paragraph position="17"> The alternative to the above is to give each u. in terms of a tree (ie. without reference to any r~le), but then there is no guarantee that this tree will correspond to some string of terms. Even if it does, one cannot be certain that it would be the string of terms one wishes to include in the rule - after all, two entirely different strings of terms may correspond to the same tree (a paraphrase) by means of two different rules.</Paragraph>
    <Paragraph position="18"> We shall discard the alternative and adopt the first approach.The generalised rule ~l,..~n~8 (with each u. being the name of a rule) can be extended furthe~ by letting u. be a list of rule names, * 1 where this is Interpreted as a choice for the string-tree correspondence A.'-T. to be referred to, and hence the choice for th~ist~ing of terms A.</Paragraph>
    <Paragraph position="19"> represented by u.. In such a situation, it ma~Iso be possible thatZwe wish the topology oT the tree B to vary according to the choice of A., and this variation to be zn terms of the subtrees of the tree T.. For these reasons, we specify each ~. as a pairI(REFERENCE, STRUCTURE) where REFERENCEIis the said list of rule names and STRUCTURE is a tree schema containing variables, such that the structure represents the tree found on the right hand side of the &amp;quot;~&amp;quot; in each rule referred to in the list REFERENCE. This way, the tree 8 can be defined in terms of T i by means of the variables (for example those appearing simultaneously in both u.</Paragraph>
    <Paragraph position="20"> and 8). See the example later in fig. 5 for an i illustration.</Paragraph>
    <Paragraph position="21">  above, where RN2 refers to RNI and itself. Variables in the entry STRUCTURE are given in boxes, eg. \[\] , where each variable can be instantiated to a linear ordered sequence of trees. For a given element (REFERENCE, STRUCTURE), the instanciations of the variables in STRUCTURE can be obtained only by identifying (an operation intuitively similar to the standard notion of unification - again, see later in fig. 5) the STRUCTURE with the right hand side of a rule given in the entry REFERENCE.</Paragraph>
    <Paragraph position="22">  As an immediate consequence to the above, an STCG rule thus defines a correspondence between a set of strings of terms on one hand and a set of trees on the other (by means of a linear sequence of sets of trees). The rule RN! describes a correspondence between a single term and a tree containing a node NP dominating a single leaf (for example, it gives the respective structural representations for &amp;quot;dog&amp;quot;, &amp;quot;catcher&amp;quot;, etc.). The rule RN2 describes a correspondence between two or more terms and a single tree - note the recursive REFERENCE in the first element of RN2 (for example, it gives the structural representation for &amp;quot;catcher hunter&amp;quot; as well as for &amp;quot;dog catcher hunter&amp;quot;, see later in fig. 5).</Paragraph>
    <Paragraph position="23"> The entry STRUCTURE of an element may also act as a constraint by making explicit certain nodes in the STRUCTURE instead of just a node dominating a forest (we have no examples for this in this paper, but one can easily visualise the idea). This means that the entry STRUCTURE of an</Paragraph>
    <Paragraph position="25"> rule referred to by u. In its entry REFERENCE.</Paragraph>
    <Paragraph position="26"> .. I Whenever It is made use of, such a constralnt ensures that only certain subsets of T., and hence of I A., are referred to and used in the correspondence descrlbed by ~I..~ ~.</Paragraph>
    <Paragraph position="27"> n The string-tree correspondence in fig.  |is defined by rule RN3 below, which refers to rules RN! and RN2. We show how this is done in fig. 5. Note that if two variables in a single rule have the same label, then their instantiations must be identical. The concept of derivation as well as the derivation tree have been defined for the STCG \[Zaharin 86\], but it would be too long to explain them here. Instead, we shall use a diagram like the one in fig. 5, which should be quite self-explanatory. null</Paragraph>
    <Paragraph position="29"> Going back to fig. 2 where the string-tree 3 J 3 correspondence for a b c is given, each substructure below a node S in the TREE corresponds to a substring &amp;quot;abe&amp;quot;, but the terms in this substring are distributed over the whole STRING. In general, Jin a string-tree correspondence AI..A ~8 defined by a rule ~l..e--8, it is posslble that we w~sh to * n deflne a substrlng-subtree correspondence of ~jjBk, are disjoint the form A. .. where A. ''*'~Jm --\]~ --J1 substrings of the string A ...A and Sk is a sub--I --n tree of 8, and that 8 k cannot be expressed in terms of the respective structural representations (if</Paragraph>
    <Paragraph position="31"> handled by a rule of the form discussed so far because a structural representation (STRUCTURE) found on the left hand side can correspond only to a unit (connected) substring.</Paragraph>
    <Paragraph position="32"> We can overcome this problem by allowing a rule to define a subcorrespondence between a substructure in the TREE (in the RHS) and a disjoint sub-string in the STRING (in the LHS), where this subcorrespondence is described in another rule (ie.</Paragraph>
    <Paragraph position="33"> using a reference - SUBREFERENCE - for a substructure in the TREE, rather than uniquely for the elements in the LHS). One also allows elements in the LHS to be given in terms of variables which can be instantiated to substrings. Rule $2 (after fig.</Paragraph>
    <Paragraph position="34"> 5) gives an example of such a rule where X,Y,Z are variables.</Paragraph>
    <Paragraph position="35"> The rule $2 is of the following general type.</Paragraph>
    <Paragraph position="36"> (Recall that we wish to define a substrinq-subtree correspondence of the form A .... ~J~gk'm Where --Jl A. ,..,A, m-J are disjoint substrings of the string --J1 A ..A and B k is a subtree of B, and that ~k cannot be expressed in terms of the respective structural ,..,A.). In the rule representations (if any) of ~Jl ~ m ..~n~B, the elements ~ ''''~n are to be as before I ! except for those representing the substrings A. ,..,A. m-J which are to be left as unknowns, written --\]i say~jl .... ~Jm respectively. The correspondence A...A,~8 k is to be written in the entry CORRES--3 I --3 m |PONDENCE as ~...~. ~gk, and this is given a refer---r-~ ~the correspondence elsewhere in another rule. In F~-)|this SUBREFERENCE, if a rule ~'..~'~8 is a possibi O 2zNP I P</Paragraph>
    <Paragraph position="38"> tion of the rule is that the SUBREFERENCE gives a string-tree correspondence A'..A'~8'which precise- --I --p ) \[y defines the string-tree correspondence PSj s k ,where B k is identified with 8' and Aj ..A. is identified with A'..A' with the  1. --Jm --i --~  separation points being obtained from the predetermined identification between X...X. and ~'..e' mentioned above. --3 1 --3 m l p A STCG containing the rules $I and $2 defines the language anbnc n, and associates a structural representation like the one in fig.2 to every string in the language. Fig.6 illustrates how this grammar defines the string-tree correspondence in fig.2.  The informal discussion in this paper gives the motivation and some idea of the formal definition of the String-Tree Correspondence Grammar. The grammar stresses not only the fact that one can express string-tree correspondence like the ones we have discussed, but also that it can be done in a 'natural' way using the formalism - meaning the structures and correspondence are explicit in the rule, and not implicit and dependent on the combination of grammar rules applied (as in derivation trees). The inclusion of the substring-subtree correspondence is also another characteristic of the grammar formalism. One also sees that the grammar is declarative in nature, and thus it is interpretable both for analysis and for generation (for example, by ~nterpretlng the rules as tree rewriting rules with variables}.</Paragraph>
    <Paragraph position="39"> In an effort to demonstrate the principal properties of the formalism, the STCG presented in this paper is in a simple form, ie. treating trees with each node having a single label. In its general form, the STCG deals with labels having considerable internal structure (lists of features, etc.). Furthermore, one can also express constraints on the features in the nodes - on individual nodes or between different nodes.</Paragraph>
    <Paragraph position="40"> As mentioned, the concepts of direct derivation (=&gt;) and derlvatzon (-&gt;), as well as the derivation tree are also defined for the STCG. (Note that the rules with properties similar to the rule $2 entail a definition of direct derivation which is more complex than the classical definition). The set of rules in a grammar forms a formal grammar, ie. it defines a language, in fact two languages, one of strings and the other of trees.</Paragraph>
    <Paragraph position="41"> At the moment, there is no large applications of the STCG, but as the STCG derives its formal properties from the Static Grammar, it would be quite a simple process to transfer applications in the Static Grammar into STCG applications. Like the Static Grammar, the STCG is basically a formalism for specification, but given its formal nature, one also aims for direct interpretability by a machine. Though still incomplete, work has begun to build such an interpreter \[Zajac 86\].</Paragraph>
  </Section>
class="xml-element"></Paper>