File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/88/p88-1034_metho.xml
Size: 26,906 bytes
Last Modified: 2025-10-06 14:12:12
<?xml version="1.0" standalone="yes"?> <Paper uid="P88-1034"> <Title>COMBINATORY CATEGORIAL GRAMMARS: GENERATIVE POWER AND RELATIONSHIP TO LINEAR CONTEXT-FREE REWRITING SYSTEMS&quot;</Title> <Section position="3" start_page="0" end_page="278" type="metho"> <SectionTitle> 1 Introduction </SectionTitle> <Paragraph position="0"> There have been a number of results concerning the relationship between the weak generative capacity (family of string languages) associated with different grammar formalisms; for example, the thecxem of Oaifman, et al. \[3\] that Classical Categorial Grammars are weakly equivalent to Context-Free Grammars (CFG's). Mote recently it has been found that there is a class of languages slightly larger than the class of Context-Free languages that is generated by several different formalisms. In pardodar, Tree Adjoining Grammars (TAG's) and Head Grammars (HG's) have been shown to be weakly equivalent \[15\], and these formalism are also equivalent to a reslriction of Indexed Grammars considered by Gazdar \[6\] called Linear Indexed Grammars (LIG's) \[13\].</Paragraph> <Paragraph position="1"> In this paper, we examine Combinatory Categorial Grammars (CCG's), an extension of Classical Categorial Grammars developed by Steedman and his collaborators \[1,12,9,10,11\]. The main result in this paper is *This work was partially mpported by NSF gnmts MCS-82-19116-CER. MCS-82-07294, DCR-84-10413, ARO grant DAA29-84-9-0027. and DARPA gnmt N0014-85-K0018. We are very grateful to Mark Steedmm, \]C Vijay-Shanker and Remo Pare~:hi for helpful disctmiem. that CCG's are weakly equivalent to TAG's, HG's, and LIG's. We prove this by showing in Section 3 that Combinatory Categorlal Languages (CCL's) are included in Linear Indexed Languages (LIL's), and that Tree Adjoining Languages (TAL's) are included in CCL's.</Paragraph> <Paragraph position="2"> After considering their weak generative capacity, we investigate the relationship between the struclzwal descriptions produced by CCG's and those of other grammar formalisms. In \[14\] a number of grammar formalisms were compared and it was suggested that an important aspect of their descriptive capacity was reflected by the derivation structures that they produced. Several formalisms that had previously been descn2~d as mildly context-sensitive were found to share a number of properties. In particular, the derivations of a grammar could be represenled with trees that always formed the tree set of a context-free grammar. Formalisms that share these properties were called Linear Context-Free Rewriting Systems ('LCFRS's) \[14\].</Paragraph> <Paragraph position="3"> On the basis of their weak generative capacity, it appears that CCG's should be classified as mildly contextsensitive. In Section 4 we consider whether CCG's should be included in the class of LCFRS's. The derivation tree sets traditionally associated with CCG's have Context-free path sets, and are similar to those of LIG's, and therefore differ from those of LCFRS's. This does not, however, nile out the possibility that there may be alternative ways of representing the derivation of CCG's that will allow for their classification as LCP'RS's.</Paragraph> <Paragraph position="4"> Extensions to CCG's have been considered that enable them to compare two unbounded sU'uctures (for example, in \[12\]). It has been argued that this may be needed in the analysis of certain coordination phenomena in Dutch.</Paragraph> <Paragraph position="5"> In Section 5 we discuss how these additional features increase the power of the formalism. In so doing, we also give an example demonstrating that the Parenthesis-free Categorial Grammar formalism \[5,4\] is moze powerful that CCG's as defined here. Extensions to TAG's (Multicomponent TAG) have been considered for similar reasons. However, in this paper, we will not investigate the relationship between the extension of CCG's and Multicomponent TAG.</Paragraph> </Section> <Section position="4" start_page="278" end_page="279" type="metho"> <SectionTitle> 2 Description of Formalisms </SectionTitle> <Paragraph position="0"> In this section we describe Combinatory Categorial Grammars, Tree Adjoining Grammars, and Linear Indexed Grammars.</Paragraph> <Section position="1" start_page="278" end_page="278" type="sub_section"> <SectionTitle> 2.1 Combinatory Categoriai Grammars </SectionTitle> <Paragraph position="0"> Combinatory Categorial Grammar (CCG), as defined here, is the most recent version of a system that has evolved in a number of papers \[1,12,9,10,11\].</Paragraph> <Paragraph position="1"> A CCG, G, is denoted by (VT, VN, S, f, R) where VT is a finite set of terminals (lexical items), VN is a finite set of nonterminals (atomic categories), S is a distinguished member of VN, f is a function that maps elements of VT U {e} to finite subsets of C(VN), the set of categories*, where V N g C(VN) and if CI, C 2 e C(VN) then (el/c2) E C(VN) and (c1\c2) E C(VN).</Paragraph> <Paragraph position="2"> R is a finite set of combinatory rules, described below. We now give the combinatory rules, where z, y, z are variables over categories, and each Ii denotes either \ or /.</Paragraph> <Paragraph position="3"> 1. forward application: 2. backward application: u (z\u) -. z 3. generaliT~d forward composition for some n _> 1: (... I.z.) -.</Paragraph> <Paragraph position="4"> 4. generalized backward composition for some n E 1:</Paragraph> <Paragraph position="6"> z Note that f can assign categoric8 to the empty suing, ~, though, to our knowledge, this feature has not been employed in the linguistic applications C/~ C'CG.</Paragraph> <Paragraph position="7"> Restrictions can be associated with the use of the combinatory rule in R. These restrictions take the form of conswaints on the instantiations of variables in the rules.</Paragraph> <Paragraph position="8"> These can be constrained in two ways.</Paragraph> <Paragraph position="9"> 1. The initial nonterminal of the category to which z is instantiated can be restricted.</Paragraph> <Paragraph position="10"> 2. The entire category to which y is instantiated can be resuicted.</Paragraph> <Paragraph position="11"> Derivations in a CCG involve the use of the combinatory rules in R. Let the derives relation be defined as follows.</Paragraph> <Paragraph position="12"> ~c~ F ~clc2~ if R contains a combinawry rule that has czc2 --* c as an instance, and a and ~ are (possibly empty) strings of categories. The string languages, L(G), generated by a CCG, G', is defined as follows.</Paragraph> <Paragraph position="13"> {al...</Paragraph> <Paragraph position="14"> c, ~ f(aO, a, ~ VT U {~}, 1 _< i _< .} Although there is no type-raising rule, its effect can be achieved to a limited extent since f can assign type-raised categories to lexical items, which is the scheme employed in Steedman's recent work.</Paragraph> </Section> <Section position="2" start_page="278" end_page="279" type="sub_section"> <SectionTitle> 2.2 Linear Indexed Grammars </SectionTitle> <Paragraph position="0"> Linear Indexed Grammars (LIG's) were introduced by Gazdar \[6\], and are a restriction of Indexed Grammars introduced by Aho \[2\]. LIG's can be seen as an extension of CFG's in which each nonterrninal is associated with a stack.</Paragraph> <Paragraph position="1"> An LIG, G, is denoted by G = ( Vjv , VT , Vs , S, P) where VN iS a finite set of nontenninals, VT is a finite set of terminals, Vs is a finite set of stack symbols, where At .... A. E VN, l E Vs, and a E VT O {~}. The notation for stacks uses \[. */\] to denote an arbi-Wary stack whose top symbol is I. This system is called L/near Indexed Grammars because it can be viewed as a restriction of Indexed Grammars in which only one of the non-terminals on the right-hand-side of a production can inherit the stack from the left-hand-side.</Paragraph> <Paragraph position="2"> The derives relation is defined as follows.</Paragraph> <Paragraph position="3"> ~A\[Z,, ... ht\]~ ~ ~A,\[\] ... A,\[Z,,... t~\].., a,\[\]~ if A\[.. l\] -. ~,\[\]...A,\[..\]...A,\[\] ~ P otA\[lm.., ll\]~ o =~ aAl\[\]... Ai\[lm... ill\]... An\[\]/~ if A\[..\] --. A,\[\]...A,\[-. Z\]...A,,\[\] ~ P : c,a\[\]a ~ ,ma if A\[\]--.a~P The language, L(G), generated by G is</Paragraph> </Section> <Section position="3" start_page="279" end_page="279" type="sub_section"> <SectionTitle> 2.3 Tree Adjoining Grammars </SectionTitle> <Paragraph position="0"> A TAG \[8,7\] is denoted G = (VN, VT, S, I, A) where VN is a finite set of nontennlnals, VT is a finite set of terminals, S is a distinguished nonterminal, I is a finite set of initial trees and A is a finite set of auxiliary trees.</Paragraph> <Paragraph position="1"> Initial trees are rooted in S with w E V~ on their frontier. Each internal node is labeled by a member of VN.</Paragraph> <Paragraph position="2"> Auxiliary trees have tOlAW2 E V'~VNV~ oll their frontier. The node on the frontier labeled A is called the foot node, and the root is also labeled A. Each internal node is labeled by a member of VN.</Paragraph> <Paragraph position="3"> Trees are composed by tree adjunction. When a tree 7' is adjoined at a node ~/in a tree .y the tree that results, 7,', is obtained by excising the subtree under t/from and inserting 7' in its place. The excised subtree is then substituted for the foot node of 3 / . This operation is illustrated in the following figure.</Paragraph> <Paragraph position="4"> ~': $ r'.&quot; x Y&quot;: s Each node in an auxiliary tree labeled by a nonterminal is associated with adjoining constraints. These constraints specify a set of auxiliary trees that can be adjoined at that node, and may specify that the node has obligatory adjunction (OA). When no tree can be adjoined at a node that node has a null adjoining (NA) constraint.</Paragraph> <Paragraph position="5"> The siring language L(G) generated by a TAG, G, is the set of all strings lYing on the frontier of some tree that can be derived from an initial trees with a finite number of adjunctions, where that tree has no OA constraints.</Paragraph> </Section> </Section> <Section position="5" start_page="279" end_page="282" type="metho"> <SectionTitle> 3 Weak Generative Capacity </SectionTitle> <Paragraph position="0"> In this section we show that CCO's are weakly equivalent to TAG's, HG's, and LIO's. We do this by showing the Inclusion of CCL's in L1L's, and the inclusion of TAL's in CCL's. It is know that TAG and LIG are equivalent \[13\], and that TAG and HG are equivalent \[15\]. Thus, the two inclusions shown here imply the weak equivalence of all four systems. We have not included complete details of the proofs which can be found in \[16\].</Paragraph> <Section position="1" start_page="279" end_page="280" type="sub_section"> <SectionTitle> 3.1 CCL's C LIL's </SectionTitle> <Paragraph position="0"> We describe how to construct a LIG, G', from an arbitrary CCG, G such that G and G' are equivalent. Let us assume that categories m-e written without parentheses, tmless they are needed to override the left associativity of the slashes.</Paragraph> <Paragraph position="1"> A category c is minimally parenthesized if and only if one of the following holds.</Paragraph> <Paragraph position="3"> where Co E VN and each c~ is minimally parenthesize~ It will be useful to be able to refer to the components of a category, c. We first define the immediate components of c.</Paragraph> <Paragraph position="4"> when c = A the immediate component is A, when c = (col:xh...I.c.) the immediate components are co, cl,. * *, e.,,. The components of a category c are its immediate components, as well as the components of its immediate components. null Although in CCG's there is no bound on the number of categories that are derivable during a derivation (categories resulting from the use of a combinatory rule), there is a bound on the number of components that derivable categories may have. This would no longer hold if unrestricted type-raising were allowed during a derivation. Let the set Dc(G) he defined as follows.</Paragraph> <Paragraph position="5"> c E De(G) if c is a component of d where c' E f(a) for some a E VT U {e}.</Paragraph> <Paragraph position="6"> Clearly for any CCG, G, Dc(G) is a finite set. Dc(G) contains the set of all derivable components because for every category e that can appear in a sentential form of a derivation in some CCG, G, each component of c is in Dc(G). This can be shown, since, for each combinatory rule, ff it holds of the categories on the left of the rule then it will hold of the category on the right. Each of the combinatory rules in a CCG can be viewed as a statement about how a pair of categories can be combined. For the sake of this discussion, let us name the members of the pair according to their role in the rule. The first of the pair in forward rules and the second of the pair in backward rules will be named the primary category. The second of the parr in forward rules and the first of the pair in backward rules will be named the secondary category.</Paragraph> <Paragraph position="7"> As a resuit of the form that combinatory rules can take in a CCG, they have the following property. When a combinatory rule is used, there is a bound on the number of immediate components that the secondary categories of that rule may have. Thus, because immediate constituents must belong to De(G) (a finite set), there is a bound on the number of categories that can fill the role of secondary categories in the use of a combinatory rule. Thus, theae is a bound on the number of instantiations of the variables y and zi in the combinatory rules in Section 2.1. The only variable that can be instantiated to an unbounded number of categories is z. Thus, by enumerating each of the finite number of variable bindings for y and each z~, the number of combinatory rules in R can be increased in such a way that only x is needed. Notice that z will appears only once on each side of the rules (Le, they are linear).</Paragraph> <Paragraph position="8"> We are now in a position m describe how to represent each of the combinatory rules by a production in the LIG, G'. In the combinatory rules, categories can be viewed as stacks since symbols need only be added and removed from the right. The secondary category of each rule will be a ground category: either A, or (AIlcl\[2... \[ncn), for some n >__ I. These can be represented in a LIG as A\[\] or A\[hCl\[2... InCh\], respectively. The primary category in a combinatory rule will be unspecified except for the identity of its left and rightmost immediate components. Its leftmost component is a nonterminal, A, and its right-most component is a member of De(G), c. This can be represented in a LIG by A\[.. el.</Paragraph> <Paragraph position="9"> In addition to mapping combinatory rules onto productions we must include productions in G' for the mappings from lexical items.</Paragraph> <Paragraph position="10"> If c E f ( a ) where a E VT U {e} then if e = A then A\[\] ...* a E P if c-'(ahcll2...I,c,) then A\[llC112..&quot; \]nOn \] .-o, a e P We are assuming an extension of the notation for productions that is given in Section 2.2. Rather than adding or removing a single symbol from the stack, a fixed number of symbols can be removed and added in one production. Furthermore, any of the nonterminals on the right of productions can be given stacks of some fixed size.</Paragraph> </Section> <Section position="2" start_page="280" end_page="282" type="sub_section"> <SectionTitle> 3.2 TAL's C CCL's </SectionTitle> <Paragraph position="0"> We briefly describe the construction of a CCG, G' from a TAG, G, such that G and G' are equivalent.</Paragraph> <Paragraph position="1"> For each nonterminal, A of G there will be two nonterminals A deg and A c in G'. The nonterminal of G' will also include a nonterminal Ai for each terminal ai of the TAG.</Paragraph> <Paragraph position="2"> The terminal alphabets will be the same. The combinatory rules of G' are as follows.</Paragraph> <Paragraph position="3"> Forward and backward application are restricted to cases where the secondary category is some X ~, and the left immediate component of the primary category is some Ydeg.</Paragraph> <Paragraph position="4"> Forward and backward composition are restricted to cases where the secondary category has the form ((XChcl)\[2c2), and the left immediate component of the primary category is some Y% An effect of the restrictions on the use of combinatory rules is that only categories that can fill the secondary role during composition are categories assigned to terminals by f. Notice that the combinatory rules of G' depend only on the terminal and nonterminal alphabet of the TAG, and are independent of the elementary trees.</Paragraph> <Paragraph position="5"> f is defined on the basis of the auxiliary trees in G. Without loss of generality we assume that the TAG, G, has trees of the following form.</Paragraph> <Paragraph position="6"> I contains one initial tree:</Paragraph> <Paragraph position="8"> Thus, in considering the language derived by G, we need only be concerned with trees derived from auxiliary trees whose root and foot are labeled by S.</Paragraph> <Paragraph position="9"> There are 5 kinds of auxiliary trees in A.</Paragraph> <Paragraph position="10"> The CCG, G', in deriving a string, can be understood as mimicking a derivation in G of that suing in which trees are adjoined in a particular order, that we now describe. We define this order by describing the set, 2~(G), of all trees produced in i or fewer steps, for i >_ 0.</Paragraph> <Paragraph position="11"> To(G) is the set of auxiliary trees of G.</Paragraph> <Paragraph position="12"> TI(G) is the union of T~_x(G) with the set of all trees 7 produced in one of the following two ways.</Paragraph> <Paragraph position="14"> Let 3 / and 7&quot; be trees in T~-I(G) such that there is a unique lowest OA node, I?, in 7' that does not dominate the foot node, and 3/' has no OA nodes. 7 is produced by adjoining 7&quot; at in 7'.</Paragraph> <Paragraph position="15"> Let 7' be trees in T~-I(G) such that there is OA node, 7, in 7' that dominates the foot node and has no lower OA nodes. 7 is pmduceA by adjoining an auxiliary tree ~ at 17 in 7'-Each tree 7 E 2~(G) with frontier wiAw2 has tbe prop-erty that it has a single spine from the root to a node that dominates the entire string wlAw2. All of the OA nodes remaining in the tree fall on this spine, or hang immediately to its right or left. For each such tree 7 there will be a derivation tree in a', whose root is labeled by a ca~gory c and with frontier to 1W2, wher~ c encodes the remaining obligatory adjunctions on this spine in 7.</Paragraph> <Paragraph position="16"> Each OA nodes on the spine is encoded in c by a slash and nonterminal symbol in the appropriate position. Suppose the OA node is labeled by some A. When the OA node falls on the spine c will contain /.4 C/ (in this case the direction of the slash was arbiwarfly chosen to be forward). When the OA node faUs to the left of the spine c will contain \A% and when the OA node fall~ to the right of the spine c will contain/A deg. For example, the following tree is encoded by the category A\A~/AI/A~\A ~</Paragraph> </Section> </Section> <Section position="6" start_page="282" end_page="283" type="metho"> <SectionTitle> 4 Derivations Trees </SectionTitle> <Paragraph position="0"> Vijay-Shanker, Weir and Joshi \[14\] described several properties that were common to various conswained grammatical systems, and defined a class of such systems called Linear Context-Free Rewriting Systems (LCFRS's). LCFRS's are constrained to have linear nonerasing composition operations and derivation trees that are structurally identical to those of context-free grammars. The intuition behind the latter restriction is that the rewriting (whether it be of strings, trees or graphs) be performed in a context-free way; i.e., choices about how to rewrite a structure should not be dependent on an unbounded amount of the previous or future context of the derivation. Several wen-known formalisms fall into this class including Context-Free Grammars, Generalized Phrase Structure Grammars (GPSG), Head Grammars, Tree Adjoining Grammars, and Multicomponent Tree Adjoining Grammars. In \[14\] it is shown that each formalism in the class generates scmilinear languages that can be recognized in polynomial time.</Paragraph> <Paragraph position="1"> In this section, we examine derivation trees of CCG's and compare them with respect to those of formalisms that are known to be LCFRS's. In order to compare CCG's with other systems we must choose a suitable method for the representation of derivations in a CCG. In the case of CFG, TAG, HG, for example, it is fairly clear what the elementary structures and composition operations should be, and as a result, in the case of these formalisms, it is apparent how to represent derivations.</Paragraph> <Paragraph position="2"> The traditional way in which derivations of a CCG have been represented has involved a binary tree whose nodes are labeled by categories with annotations indicating which combinatory rule was used at each stage. These derivation trees are different from those systems in the class of LCFRS's in two ways. They have context-free path sets, and the set of categories labeling nodes may be infinite. A property that they share with LCFRS's is that there is no dependence between unbounded paths. In fact, the derivation trees sets produced by CCG's have the same properties as those produced by LIG's (this is apparent from the construction in Section 3A).</Paragraph> <Paragraph position="3"> Although the derivation trees that are traditionally associated with CCG's differ from those of LCFRS's, this does not preclude the possibility that there may be an alternative way of representing derivations. What appears to be needed is some characterization of CCG's that identities a finite set of elementary structures and a finite set of composition operations.</Paragraph> <Paragraph position="4"> The equivalence of TAG's and CCG's suggests one way of doing this. The construction that we gave from TAG's to CCG's produced CCG's having a specific form which can be thought of as a normal form for CCG's. We can represent the derivations of grammars in this form with the same tree sets as the derivation tree sets of the TAG from which they were constructed. Hence CCG's in this normal form can be classified as LCFRS's.</Paragraph> <Paragraph position="5"> TAG derivation trees encode the adjanction of specified elementary trees at specified nodes of other elementary trees. Thus, the nodes of the derivation trees are labeled by the names of elementary trees and tree addresses. In the construction used in Section 3.2, each auxiliary tree produces assignments of elementary categories to lexicai items. CCG derivations can be represented .with trees whose nodes identify elementary categories and specify which combinatory rule was used to combine it.</Paragraph> <Paragraph position="6"> For grammars in this normal form, a unique derivation can be recovered from these trees, but this is not true of arbitrary CCG's where different orders of combination of the elementary categories can result in derivations that must be distinguished. In this normal form, the combinatory rules are so restrictive that there is only one order in which elementary categories can be combined. Without such restrictions, this style of derivation tree must encode the order of derivation.</Paragraph> </Section> <Section position="7" start_page="283" end_page="283" type="metho"> <SectionTitle> 5 Additions to CCG's </SectionTitle> <Paragraph position="0"> CCG's have not always been defined in the same way.</Paragraph> <Paragraph position="1"> Although TAG's, HG's, and CCG's, can produce the crossing dependencies appearing in Dutch, two additions to CCG's have been considered by Steedman in \[12\] to describe certain coordination phenomena occurring in Dutch. For each addition, we discuss its effect on the power of the system.</Paragraph> <Section position="1" start_page="283" end_page="283" type="sub_section"> <SectionTitle> 5.1 Unbounded Dependent Structures </SectionTitle> <Paragraph position="0"> A characteristic feature of LCFRS's is that they are unable to produce two structures exhibiting an unbounded dependence. It has been suggested that this capability may be needed in the analysis of coordination in Dutch, and an extension of CCG's has been proposed by Steedman \[12\] in which this is possible. The following schema is included.</Paragraph> <Paragraph position="2"> where, in the analysis given of Dutch, z is allowed to match categories of arbitrary size. Two arbitrarily large structures can be encoded with two arbitrarily large categories. This schema has the effect of checking that the encodings are identical The addition of rules such as this increases the generative power of CCG's, e.g., the following language can be generated.</Paragraph> <Paragraph position="3"> {(wc)&quot; I w e {a,b} deg} In giving analysis of coordination in languages other than Dutch, only a finite number of instances of this schema are required since only bounded categories are involved.</Paragraph> <Paragraph position="4"> This form of coordination does not cause problems for LCFRS's.</Paragraph> </Section> <Section position="2" start_page="283" end_page="283" type="sub_section"> <SectionTitle> 5.2 Generalized Composition </SectionTitle> <Paragraph position="0"> Steedman \[12\] considers a CCG in which there are an inf~te number of composition rules for each n _> 1 of the form</Paragraph> <Paragraph position="2"> (... (~1:012... I,z,) This form of composition is permitted in Parenthesis-free Categorial Grammars which have been studied in \[5,4\], and the results of this section als0 apply to this system. With this addition, the generative power of CCG's increases. We show this by giving a grammar for a language that is known not to be a Tree Adjoining language. Consider the following CCG. We allow um~stricted use of arbitrarily many combinatory rules for forward or backwards generalized composition and application.</Paragraph> <Paragraph position="4"> When the language, L, generated by this grammar is intersected with the regular language we get the following language.</Paragraph> <Paragraph position="5"> nl ~3 ~1 ftl ft2 ft 3 2 1 {a I G 2 b I C 1 b 2 C 2 d~2 d~l I nl,n 2 0} The pumping lemma for Tree Adjoining Grammars \[13\] can be used to show that this is not a Tree Adjoining Language. Since Tree Adjoining Languages are closed under intersection with Regular Languages, L can not be a Tree Adjoining Language either.</Paragraph> </Section> </Section> class="xml-element"></Paper>