File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/84/p84-1026_metho.xml
Size: 43,179 bytes
Last Modified: 2025-10-06 14:11:34
<?xml version="1.0" standalone="yes"?> <Paper uid="P84-1026"> <Title>SYNTACTIC AND SEMANTIC PARSABILITY</Title> <Section position="4" start_page="112" end_page="113" type="metho"> <SectionTitle> 2. COULD NL'S BE REGULAR SETS? </SectionTitle> <Paragraph position="0"> Chomsky's negative answer to this question was the correct one. Although his original argument in Syntactic Structures (1957) for the non-regular character of English was not given in anything like a valid form (cf. Daly 1974 for a critique), others can be given. Consider the following, patterned after a suggestion by Brandt Corstius (see Levelt 1974, 25-26). The set (1): (I) {a white male (whom a white male) n (hired) n hired another white male I n ~ 0} is the intersection of English with the regular set & whi;e male (.whom a white male)* hired* another white male. But (1) is not regular, yet the regular sets are closed under intersection; hence English is not regular. Q.E.D.</Paragraph> <Paragraph position="1"> It is perfectly possible that some NL's happen not to present the inherently self-embedding configurations that make a language non-regular.</Paragraph> <Paragraph position="2"> Languages in which parataxis is used much more than hypotaxis (i.e. languages in which separate clauses are strung out linearly rather than embedded) are not at all uncommon. However, it should not be thought that non-regular configurations will be found to be rare in languages of the world. There are likely to be many languages that furnish better arguments for non-regular character than English does; for example, according to Bag~ge (1976), center-embedding seems to be commoner and more acceptable in several Central Sudanic languages than it is in English. In Moru, we find examples such as this (slightly simplified from Ha~ege (1976, 200); Xi is the possession marker for nonhuman nouns, and ro is the equivalent for human nouns): (2) kokyE \[toko \[odrupi \[ma ro\] ro\] ri\] drate 1 2 3 3 2 1 dog wife brother me of of of is-dead &quot;My brother's chief wife's black dog is dead.&quot; The center-embedding word order here is the only one allowed; the alternative right-branching order (&quot;dog chief-wife-of brother-of me-of&quot;), which a regular grammar could handle, is ungrammatical.</Paragraph> <Paragraph position="3"> Presumably, the intersection of odrupi* ma to* drate with Moru is n n {odrupi ma ro drate \[ n > 0} (an infinite set of sentences with meanings like '~y brother's brother's brother is dead&quot; where n 3). This clearly non-regular, hence so is Moru.</Paragraph> <Paragraph position="4"> The fact that NL's are not regular does not necessarily mean that techniques for parsing regular languages are irrelevant to NL parsing.</Paragraph> <Paragraph position="5"> Langendoen (1975) and Church (1980) have both, in rather different ways, proposed that hearers process sentences as if they were finite automata (or as if they were pushdown automata with a finite stack depth limit, which is weakly equivalent) rather than showing the behavior that would be characteristic of a more powerful device. To the extent that progress along these lines casts light on the human parsing ability, the theory of regular grammars and finite automata will continue to be important in the study of natural languages even though they are not regular sets.</Paragraph> <Paragraph position="6"> The fact that NL's are not regular sets is both surprising and disappointing from the standpoint of parsahility. It is surprising because there is no simpler way to obtain infinite languages than to admit union, concatenation, and Kleene closure on finite vocabularies, and there is no apparent priori reason why humans could not have been well served by regular languages. Expressibility considerations, for example, do not appear to be relevant: there is no reason why a regular language could not express any proposition expressible by a sentence of any finite-string-length language.</Paragraph> <Paragraph position="7"> Indeed, many languages provide ways of expressing sentences with self-ambedding structure in nonself-embedding ways as well. In an SOV language like Korean, for example, sentences with the tree-structure (3a) are also expressible with left-branching tree-structure as shown in (3b).</Paragraph> <Paragraph position="9"> Clearly such structural rearrangement will not alter the capacity of a language to express propositions, any more than an optimizing compiler makes certain programs inexpressible when it irons out true recursion into tail recursion wherever possible. null If NL's were regular sets, we know we could recognize them in deterministic linear time using the fastest and simplest abstract computing devices of all, finite state machines. However, there are much larger classes of languages that have linear time recognition. One such class is the deterministic context-free languages (DCFL's). It might be reasonable, therefore~ to raise the question dealt with in the following section.</Paragraph> </Section> <Section position="5" start_page="113" end_page="114" type="metho"> <SectionTitle> 3. COULD NL'S BE DCFL'S? </SectionTitle> <Paragraph position="0"> To the best of my knowledge, this question has never previously been raised, much less answered, in the literature of linguistics or computer science. Rich (1983) is not atypical in dismissing the entire literature on DCFL's without a glance on the basis of an invalid argument which is supposed to show that English is not even a CFL, hence fortiori not a DCFL.</Paragraph> <Paragraph position="1"> I should make it clear that the DCFL's are not just those CFL's for which someone has written a parser that is in some way deterministic. They are the CFL's that are accepted by some deterministic pushdown stack automaton. The term &quot;deterministic parsing&quot; is used in many different ways (cf. Marcus (1980) for an attempt to motivate a definition of determinism specifically for the parsing of NL's).</Paragraph> <Paragraph position="2"> For example, a translator system with a post-processor to rank quantifier-scope ambiguities for plausibility and output only the highest-ranked translation might be described as deterministic, but there is no reason why the language it recognizes should he a DCFL; it might be any recursive language. The parser currently being implemented by the natural language team at HP Labs (in particular, by Derek Proudian and Dan Flickinger) introduces an interesting compromise between determinism and nondeterminism in that it ranks paths through the rule system so as to make some structural possibilities highly unlikely ones, and there is a toggle that can be set to force the output to contain only likely parses. When this option is selected, the parser runs faster, but can still show ambiguities when both readings are defined as likely. This is an intriguing development, but again is irrelevant to the language-theoretic question about DCFL status that I am raising.</Paragraph> <Paragraph position="3"> It would be an easy slip to assume that NL's cannot be DCFL's on the grounds that English is well known to be ambiguous. We need to distinguish carefully between ambiguity and inherent ambiguity.</Paragraph> <Paragraph position="4"> An inherently ambiguous language is one such that all of the gra~mrs that weakly generate it are ambiguous. LR gr-----rs are never ambiguous; but the LR grammars characterize exactly the set of DCFL's, hence no inherently ambiguous language is a DCFL. But it has never been argued, as far as I know, that English as a stringset is inherently ambiguous. Rather, it has been argued that a descriptively adequate grammar for it should, to account for semantic intuitions, be ambiguous. But obviously, a DCFL can have an ambiguous grammar.</Paragraph> <Paragraph position="5"> In fact, all languages have ambiguous grazmnars.</Paragraph> <Paragraph position="6"> (The proof is trivial. Let w be a string in a language ~ generated by a grannnar G with initial symbol S and production set P. Let B be a nonterminal not used by G. Construct a new grammar G&quot; with production set P&quot; = E U {S --> B, B --> w}.</Paragraph> <Paragraph position="7"> G&quot; is an ambiguous grannuar that assigns two structural descriptions to w.) The relevance of this becomes clear when we observe that in natural language processing applications it is often taken to be desirable that a parser or translator should yield just a single analysis of an input sentence. One can imagine an impemented natural language processing system in which the language accepted is described by an ambiguous CF-PSG but is nonetheless (weakly) a DCFL. When access to all possible analyses of an input is desired (say, in development work, or when one wants to take no risks in using a database front end), an all-paths parser~translator is used, but when quick-and-dirty responses are required, at the risk of missing certain potential parses of ambiguous strings, this is replaced by a deterministic one-path parser. Despite the difference in results, the language analyzed and the grammar used could be the same.</Paragraph> <Paragraph position="8"> The idea of a deterministic parser with an ambiguous grammar, which arises directly out of what has been done for programming languages in, for example, the Yacc system (Johnson 1978), is explored for natural languages in work by Fernando Pereira and Stuart Shieber. Shieber (1983) describes an implementation of a parser which uses an ambiguous grammar but parses deterministically.</Paragraph> <Paragraph position="9"> The parser uses shift-reduce scheduling in the manner proposed by Pereira (1984). Shieber (1983, i16) gives two rules for resolving conflicts between parsing actions: (I) Resolve shift-reduce conflicts by shifting.</Paragraph> <Paragraph position="10"> (If) Resolve reduce-reduce conflicts by performing the longer reduction.</Paragraph> <Paragraph position="11"> The first of these is exactly the same as the one given for Yacc by Johnson (1978, 13). The second is more principled than the corresponding Yacc rule, which simply says that a rule listed earlier in the grammar should take precedence over a rule listed later to resolve a reduce-reduce conflict: But it is particularly interesting that the two are in practice equivalent in all sensible cases, for reasons I will briefly explain.</Paragraph> <Paragraph position="12"> A reduce-reduce conflict arises when a string of categories on the stack appears on the right hand side of two different rules in the grammar. If one of the reducible sequences is longer than the other, it must properly include the other. But in that case the prior application of the properly including rule is mandated by an extension into parsing theory of the familiar rule interaction principle of Proper Inclusion Precedence, due originally to the ancient Indian grammarian Panini (see Pullum 1979, 81-86 for discussion and references). Thus, if a rule N_~P --> NP PP were ordered before a rule VP --> V NP PP in the list accessed by the parser, it would be impossible for the sequence &quot;NP PP&quot; ever to appear in a VP, since it would always be reduced to NP by the earlier rule; the VP rule is useless, and could have been left out of the grammar. But if the rule with the properly including expansion &quot;V NP PP&quot; is ordered first, the NP rule is not useless. A string &quot;V NP PP PP&quot;, for example, could in principle be reduced to &quot;V NP PP&quot; by the NP rule and then to &quot;VP&quot; by the VP rule. Under a principle of rule interaction made explicit in the practice of linguists, therefore, the proposal made by Pereira and Shieber can be seen to be largely equivalent to the cruder Yacc resolution procedure for deterministic parsing with ambiguous grammars.</Paragraph> <Paragraph position="13"> Techniques straight out of programming language and compiler design may, therefore, be of considerable interest in the context of natural language processing applications. Indeed, Shieber goes so far as to suggest psycholinguistic implications.</Paragraph> <Paragraph position="14"> He considers the,class of &quot;garden-path sentences&quot; such as those in (4).</Paragraph> <Paragraph position="15"> (4) The diners hurried through their meal were annoyed.</Paragraph> <Paragraph position="16"> That shaggy-looking sheep should be sheared is important.</Paragraph> <Paragraph position="17"> On these, his parser fails. Strictly speaking, therefore, they indicate that the language parsed is not the same under the one-path and the all-paths parsers. But interestingly, human beings are prone to fail just as badly as Shieber's parser on sentences such as these. The trouble with these cases is that they lack the prefix property---that is, they have an initial proper substring which is a sentence. (From this we know that English does not have an LR(0) grammar, incidentally.) English speakers tend to mis-parse the prefix as a sentence, and baulk at the remaining portion of the string. We might think of characterizing the notion &quot;garden-path sentence&quot; in a rigorous and non-psychological way in terms of an all-paths parser and a deterministic one-path parser for the given language: the garden path sentences are just those that parse under the former but fail under the latter.</Paragraph> <Paragraph position="18"> To say that there might be an appropriate deterministic parser for English that fails on certain sentences, thus defining them as garden-path sentences, is not to deny the existence of a deterministic pushdown automaton that accepts the whole of English, garden-path sentences included, it is an open question, as far as I can see, whether English as a whole is weakly a DCFL. The likelihood that the answer is positive is increased by the results of Bermudez (1984) concerning the remarkable power and richness of many classes of deterministic parsers for subsets of the CFL's.</Paragraph> <Paragraph position="19"> If the answer were indeed positive, we would have some interesting corollaries. To take just one example, the intersection between two dialects of English that were both DCFL's would itself be a DCFL (since the DCFL's are closed under intersection). This seems right: if your dialect and mine share enough for us to communicate without hindrance, and both our dialects are DCFL's, it would be peculiar indeed if our shared set of mutually agreed-upon sentences was not a DCFL. Yet with the CFL's in general we do not have such a result.</Paragraph> <Paragraph position="20"> Claiming merely that English dialects are CFL's would not rule out the strange situation of having a pair of dialects, both CFL's, such that the intersection is not a CFL.</Paragraph> </Section> <Section position="6" start_page="114" end_page="115" type="metho"> <SectionTitle> 4. ARE ALL NL'S CFL'S? </SectionTitle> <Paragraph position="0"> More than a quarter-century of mistaken efforts have attempted to show that not all NL's are CFL's.</Paragraph> <Paragraph position="1"> This history is carefully reviewed by Pullum and Gazdar (1982). But there is no reason why future attempts should continue this record Of failure.</Paragraph> <Paragraph position="2"> It is perfectly clear what sorts of data from a NL would show it to be outside the class of CFL's. For example, an infinite intersection with a regular set having the form of a triple-counting language or a string matching language (Pullum 1983) would suffice. However, the new arguments for non- null context-freeness of English that have appeared between 1982 and the present all seem to be quite wide of the mark.</Paragraph> <Paragraph position="3"> Manaster-Ramer (1983) points to the contemptuous reduplication pattern of Yiddish-influenced English, and suggests that it instantiates an infinite string matching language. But does our ability to construct phrases like Manaster-Ramer Schmanaster-Ramer (and analogously for any other word or phrase) really indicate that the syntax of English constrains the process? I do not think so. Manaster-Ramer is missihg the distinction between the structure of a language and the culture of verbal play associated with it. I can speak in rhyming couplets, or with adjacent word-pairs deliberately Spoonerized, or solely in sentences having an even number of words, if I wish. The structure of my language allows for such games, but does not legislate regarding them.</Paragraph> <Paragraph position="4"> Higginbotham (1984) presents a complex pumpinglemma argument on the basis of the alleged fact that sentences containing the construction a N such that S always contain an anaphoric pronoun within the clause S that is in syntactic agreement with the noun N. But his claim is false. Consider a phrase like any society such that more people get divorced than get married in an average 7ear. This is perfectly grammmtical, but has no overt anaphoric pronoun in the such that clause. (A similar ex-mple is concealed elsewhere in the text of this paper.) Langendoen and Postal (1984) consider sentences like Joe was talking about some bourbon-lover, but WHICH bourbon-lover is unknown, and argue that a compound noun of any length can replace the first occurrence of bourbon-lover provided the same string is substituted for the second occurrence as well. They claim that this yields an infinite string matching language extractable from English through intersection with a regular set. gut this argument presupposes that the ellipsis in WHICH bourbon-lover \[Joe was talking about\] must find its antecedent in the current sentence. This is not so. Linguistic accounts of anaphora have often been overly fixated on the intrasentential syntactic conditions on antecedent-anaphor pairings. Artificial intelligence researchers, on the other hand, have concentrated more on the resolution of anaphora within the larger context of the discourse. The latter emphasis is more likely to bring to our attention that ellipsis in one sentence can have its resolution through material in a preceding one. Consider the following exchange: (5)A: It looks like they're going to appoint another bourbon-hater as Chair of the Liquor Purchasing Committee.</Paragraph> <Paragraph position="5"> B: Yes--even though Joe nominated some bourbon-lovers; but WHICH bourbon-hater is still unknown.</Paragraph> <Paragraph position="6"> It is possible for the expression WHICH bourbon-hater in B's utterance to be understood as WHICH bourbon-hater \[they're ~oinR to appoint\] despite the presence in the same sentence of a mention of bourbon-lovers. There is thus no reason to believe that Langendoen and Postal's crucial example type is syntactically constrained to take its antecedent from within its own sentence, even though that is the only interpretation that would occur to the reader when judging the sentence in isolation.</Paragraph> <Paragraph position="7"> Nothing known to me so far, therefore, suggests that English is syntactically other than a CFL; indeed, I know of no reason to think it is not a deterministic CFL. As far as engineering is concerned, this means that workers in natural language processing and artificial intelligence should not overlook (as they generally do at the moment) the possibilities inherent in the technology that has been independently developed for the computer processing of CFL's, or the mathematical results concerning their structures and properties.</Paragraph> <Paragraph position="8"> From the theoretical standpoint, however, a different issue arises: is the oontext-free-ness of English just an accident, much like the accident it would be if we found that Chinese was regular? Are there other languages that genuinely show non-context-free properties? I devote the next section to this question, because some very important results bearing on it have been reported recently.</Paragraph> <Paragraph position="9"> Since these results have not yet been published, l will have to st~anarize them rather abstractly, and cite forthcoming or in-preparation papers- for further details.</Paragraph> </Section> <Section position="7" start_page="115" end_page="116" type="metho"> <SectionTitle> 5. NON-CONTEXT-FREENESS IN NATURAL LANGUAGES </SectionTitle> <Paragraph position="0"> Some remarkable facts recently reported by Christopher Culy suggest that the African language Bambara (Mande family, spoken in Senegal, Mali, and Upper Volta by over a million speakers) may be a non-CYL. Culy notes that Bambara forms from noun stems compound words of the form '~oun-~-Noun&quot; with the meaning &quot;whatever N&quot;. Thus, given that wulu means &quot;dog&quot;, wulu-o-wulu means &quot;whatever dog.&quot; He then observes that Bambara also forms compound noun stems of arbitrary length; wulu-filela means &quot;dogwatcher,&quot; wulu-nyinila means &quot;dog-hunter,&quot; wulufilela-nyinila means &quot;dog-watcher-hunter,&quot; and so on. From this it is clear that arbitrarily long words like wulu-filela-nyinila-o-wulu-filelanyinila &quot;whatever dog-watcher-hunter '~ will be in the language. This is a realization of a hypothetical situation sketched by Langendoen (1981), in which reduplication applies to a class of stems that have no upper length bound. Culy (forthcoming) attempts to provide a formal demonstration that this phenomenon renders Bambara non-contextfree. null If gambara turns out to have a reduplication rule defined on strings of potentially unbounded length, then so might other languages. It would be reasonable, therefore, to investigate the case of Engenni (another African language, in the Kwa family, spoken in Rivers State, Nigeria by about 12,000 people). Carlson (1983), citing Thomas (1978), notes that Engenni is reported to have a phrasal reduplication construction: the final phrase of the clause is reduplicated to indicate &quot;secondary aspect.&quot; Carlson is correct in noting that if there is no grammatical upper bound to the length of a phrase that may be reduplicated, there is a strong possibility that Engenni could be shown to be a non-CFL.</Paragraph> <Paragraph position="1"> But it is not only African languages in which relevant evidence is being turned up. Swiss German may be another case. In Swiss German, there is evidence of a pattern of word order in subordinate infinitival clauses that is very similar to that observed in Dutch. Dutch shows a pattern in which an arbitrary number of noun phrases (NP's) may be followed by a finite verb and an arbitrary number of nonfinite verbs, and the semantic relations between them exhibit a crossed serial pattern--i.e. verbs further to the right in the string of verbs take as their objects NP's further to the right in the string of NP's. Bresnan et al. (1982) have shown that a CF-PSG could not assign such a set of dependencies syntactically, but as Pull,-, and Gazdar (1982, section 5) show, this does not make the stringset non-context-free. It is a semantic problem rather than a syntactic one. In Swiss German, however, there is a wrinkle that renders the phenomenon syntactic: certain verbs demand dative rather than accusative case on their objects, as a matter of pure syntax. This pattern will in general not be one that a CF-PSC can describe. For example, if there are two verbs and ~&quot; and two nouns ~ and n', the set {xv I ~ is in (n, n')* and ~ is in (v, v')* and for all ~, if the i'th member of x is n the i'th member of y is ~} is not a CFL. Shieber (1984) has gathered data from Swiss German to support a rigorously formulated argument along these lines that the language is indeed not a CFL because of this construction.</Paragraph> <Paragraph position="2"> It is possible that other languages will have properties that render them non-context-free. One case discussed in 1981 in unpublished work by Elisabet Eugdahl and Annie Zaenen concerns Swedish.</Paragraph> <Paragraph position="3"> In Swedish, there are three grammatical genders, and adjectives agree in gender with the noun they describe. Consider the possibility of a &quot;respectively&quot;-sentence with a meaning like '~he NI, N2, and N3 are respectively AI, A2, and A3,&quot; where NI, N2, and N3 have different genders end AI, A2, and A3 are required to agree with their corresponding nouns in gender. If the gender agreement were truly a syntactic matter (c~ntra Pullum and Gazdar (1982, 500-501, note 12)), there could be an argument to be made that Swedish (or any language with these sort of facts) was not a CFL.</Paragraph> <Paragraph position="4"> It is worth noting that arguments based on the above sets of facts have not yet been published for general scholarly scrutiny. Nonetheless, what I have seen convinces me that it is now very likely that we shall soon see a sound published demonstration that some natural language is non-contextfree. It is time to consider carefully what the implications are if this is true.</Paragraph> </Section> <Section position="8" start_page="116" end_page="118" type="metho"> <SectionTitle> 6. CONTEXT-FREE GRAMMARS AND SEMANTIC FILTERING </SectionTitle> <Paragraph position="0"> What sort of expressive power do we obtain by allowing the definition of a language to be given jointly by the syntax and the semantics rather than just by the syntax, so that the syntactic rules can generate strings judged ill-formed by native speakers provided that the semantic rules are unable to assign interpretations to them? This idea may seem to have a long history, in view of the fact that generative gr-mmsrians engaged in much feuding in the seventies over the rival merits of gr--,-ars that let &quot;semantic&quot; factors constrain syntactic rules and grammars that disallowed this but allowed &quot;interpretive rules&quot; to filter the output of the syntax. But in fact, the sterile disputes of those days were based on a use of the term &quot;semantic&quot; that bore little relation to its original or current senses. Rules that operated purely on representations of sentence structure were called &quot;semantic&quot; virtually at whim, despite matching perfectly the normal definition of &quot;syntactic&quot; in that they concerned relations holding among linguistic signs. The disputes were really about differently ornamented models of syntax. null What I mean by semantic filtering my be illustrated by reference to the analysis of expletive NP's like there in Sag (1982). It is generally taken to be a matter of syntax that the dummy pronoun subject there can appear as the subject in sentences like There are some knives in the drawer but not in strings like *There broke all existin2 records. Sag simply allows the syntax to generate structures for strings like the latter. He characterizes them as deviant by assigning to there a denotation (namely, an identity function on propositions) that does not allow it to combine with the translation of ordinary VP's like broke all ex~stin~ records. The VP aye ~pme knives in the drawer is assigned by the semantic rules a denotation the same as that of the sentence Some ~pives are in the drawer, so there combines with it and a sentence meaning is obtained. But br~ke all existinz records translates as a property, and no sentence meaning is obtained if it is given ~here as its subject. This is the sort of move that I will refer to as semantic filtering.</Paragraph> <Paragraph position="1"> A question that seems never to have been considered carefully before is what kind of languages can be defined by providing a CF-PSG plus a set of semantic rules that leave some syntactically generated sentences without a sentence meaning as their denotation. For instance, in a system with a CF-PSG and a denotational semantics, can the set of sentences that get assigned sentence denotations be non-CF? I am grateful to Len Schubert for pointing out to me that the answer is yes, and providing the following example; Consider the following gra~mmr, composed of syntactic rules paired with semantic translation schemata.</Paragraph> <Paragraph position="3"> Assume that there are two basic semantic types, and B, and that 4&quot; andS&quot; are constants denoting entities of types A and B respectively. ~, ~, and are cross-categorlal operators. ~(~) has the category of functions from X-type things to !-type things, ~(~) has the cate~ry of functions from A_type things to ~-type things, and H(X) has the category of functions from B-type things to X-type things. Given the semantic translation schemata, every different X constituent has a unique semantic category; the structure of the string is coded into the structure of its translation. But the first rule only yields a meaning for the S constituent if L&quot; and ~&quot; are of the same category. Whatever semantic category may have been built up for an instance of ~', the F operator applies to produce a function from things of that type to things of type B, and the rule says that this function must be applied to the translation of ~'. Clearly, if R&quot; has exactly the same semantic category as L&quot; this will succeed in yielding a B-type denotation for S, and under all other circumstances S will fail to be assigned a denotation.</Paragraph> <Paragraph position="4"> The set of strings of category S that are assigned denotations under these rules is thus {xx I ~ in (A, ~)+} which is a non-CF language. We know, therefore, that it is possible for semantic filtering of a set of syutactic rules to alter expressive power significantly. We know, in fact, that it would be possible to handle Bambara noun stems in this way and design a set of translation principles that would only allow a string '~oun-~-Noun&quot; to be assigned a denotation if the two instances of N were stringwise identical. What we do not know is how to formulate with clarity a principle of linguistic theory that adjudicates on the question of whether the resultant description, with its infinite number of distinct semantic categories, is permissible.</Paragraph> <Paragraph position="5"> Despite the efforts of Barbara Hall Partee and other scholars who have written on constraining the Moutague semantics framework over the past ten years, questions about permissible power in semantic apparatus are still not very well explored.</Paragraph> <Paragraph position="6"> One thing that is clear is that Gazdar and others who have claimed or assumed that NL's are context-free never intended to suggest that the entire mechanism of associating a sentence with a meaning could be carried out by a system equivalent to a pushdown automaton. Even if we take the notion &quot;associating a sentence with a meaning&quot; to be fully clear, which is granting a lot in the way of separating out pragmatic and discourse-related factors, it is obvious that operations beyond the power of a CY-PSG to define are involved. Things like identifying representations to which lambdaconversion can apply, determining whether ali variables are bound, checking that every indexed anaphoric element has an antecedent with the same index, verifying that a structure contains no vacuous quantification, and so on, are obviously of non-CF character when regarded as language recognition problems. Indeed, in one case, that of disallowing vacuous quantifiers, it has been conjectured (Partee and Marsh 1984), though not yet proved, that even an indexed grammar does not have the requisite power.</Paragraph> <Paragraph position="7"> It therefore should not be regarded as surprising that mechanisms devised to handle the sort of tasks involved in assigning meanings to sentences can come to the rescue in cases where a given syntactic framework has insufficient expressive power. Nor should it be surprising that those syntactic theories that build into the syntax a power that amply suffices to achieve a suitable syntax-tosemantics mapping have no trouble accommodating all new sets of facts that turn up. The moment we adopt any mechanisms with greater than, say, context-free power, our problem is that we are faced with a multiplicity of ways to handle almost any descriptive problem.</Paragraph> <Paragraph position="8"> 7. GRAMMARS WITH INFINITE NONTERMINAL VOCABULARIES Suppose we decide we want to reject the idea of allowing a souped-up semantic rule system do part of the job of defining the membership of the language. What syntactic options are reasonable ones, given the kind of non-context-free languages we think we might have to describe? There is a large range of theories of grammar definable if we relax the standard requirement that the set N of nonterminal vocabulary of the gr---,=r should be finite. Since a finite parser for such a gr,mm-r cannot contain an infinite list of nonterminals, if the infinite majority of the nonterminaIs are not to be useless symbols, the parser must be equipped with some way of parsing representations of nonterminals, i.e. to test arbitrary objects for membership in N. If the tests do not guarantee results in finite time, then clearly the device may be of Turing-machine power, and may define an undecidable language. Two particularly interesting types of grammar that do not have this property are the following: Indexed 2rammars. If members of N are built up using sequences of indices affixed to a members of a finite set of basic nonterminals, and rules in P are able to add or remove sequence-initial indices, attached to a given basic nonterminal, the expressive power achieved is that of the indexed grammars of Aho (1968). These have an automata-theoretic characterization in terms of a stack automaton that can build stacks inside other stacks but can only empty a stack after all the stacks within it have been emptied. The time complexity of the parsing problem is exponential.</Paragraph> <Paragraph position="9"> Unification Krannaars. If members of N have internal hierarchical structure and parsing operations are permitted to match hierarchical representations one with another globally to determine whether they unify (roughly, whether there is a minimal consistent representation that includes the distinctive properties of both), and if the number of parses for a given sentence is kept to a finite number by requiring that we do not have A ==> A for any A, then the expressive power seems to be weakly equivalent to the grammars that Joan Bresnan and Ron Kaplan have developed under the name lexical-functional ~ramar (LF___GG; se___fie Bresnan, e__dd., 1982; c__ff, also the work of Martin Kay on unification grammars). The LFG languages include some non-indexed languages (Kelly Roach, unpublished work), and apparently have an NP-complete parsing problem (Ron Kaplan, personal communication).</Paragraph> <Paragraph position="10"> Systems of this sort have an undeniable interest in connection with the study of natural language.</Paragraph> <Paragraph position="11"> Both theories of language structure and computational implementations of grammars can be usefully explored in such terms. My criticism of them would be that it seems to me that the expressive power of these systems is too extreme. Linguistically they are insufficiently restrictive, and computationally they are implausibly wasteful of resources. However, rather than attempt to support this vague prejudice with specific criticisms, I would prefer to use my space here to outline an alternative that seems to me extremely promising.</Paragraph> <Paragraph position="12"> 8. READ GRAMMARS AND NATURAL LANGUAGES In his recent doctoral dissertation, Carl Pollard (1984) has given a detailed exposition and motivation for a class of grammars he terms head K~ammars. Roach (1984) has proved that the languages generated by head grammars constitute a full AFL, showing all the significant closure properties that characterize the class of CFL's. Head grammars have a greater expressive power, in terms of weak and strong generative capacity, than the CF-PSG's, but only to a very limited extent, as shown by some subtle and suprising results due to Roach (1984). For example, there is a head grammar for {anbncna n I n 2 0} but not for {anbncndna n \[ n 2 O} and there is a head grammar for {ww I w is in (a, b)*} but not for {ww J w is in (a, b)*}.</Paragraph> <Paragraph position="13"> The time complexity of the recognition problem for head grammars is also known: a time bound proportional to the seventh power of the length of the input is sufficient to allow for recognition in the worst case on a deterministic Turing machine (Pollard 1984). This clearly places head grammars in the realm of tractable linguistic formalisms.</Paragraph> <Paragraph position="14"> The extension Pollard makes in CF-PSG to obtain the head gra--,ars is in essence fairly simple.</Paragraph> <Paragraph position="15"> First, he treats the notion '*head&quot; as a primitive. The strings of terminals his syntactic rules define are headed s~rings, which means they are associated with an indication of a designated element to be known as the head. Second, he adds eight new '~rapping&quot; operations to the standard concatenation operation on strings that a CF-PSG can define. For a given ordered pair <B,C> of headed strings there are twelve ways in which strings B and C can be combined to make a constituent A. I give here the descriptions of just two of them which I will use below: LCI(B,C): concatenate C onto end of B; first argument (B) is head of the result.</Paragraph> <Paragraph position="16"> Mnemonic: Left Concatenation with ist as new head.</Paragraph> <Paragraph position="17"> LL2(B,C): wrap B around C, with head of B to the left of C; C is head of the result.</Paragraph> <Paragraph position="18"> Mnemonic: Left wrapping with head to the Right and ~nd as new head.</Paragraph> <Paragraph position="19"> The full set of operations is given in the chart in figure I.</Paragraph> <Paragraph position="20"> A simple and linguistically motivated head grammar can be given for the Swiss German situation mentioned earlier. I will not deal with it here, because in the first place it would take considerable space, and in the second place it is very simple to read off the needed account from Pollard's (1984) treatment of the corresponding situation in Dutch, making the required change in the syntax of case-marking.</Paragraph> <Paragraph position="21"> In the next section I apply head grammar to cases like that of Bambara noun reduplication.</Paragraph> </Section> <Section position="9" start_page="118" end_page="118" type="metho"> <SectionTitle> 9. THE RIDDLE OF REDUPLICATION </SectionTitle> <Paragraph position="0"> I have shown in section 6 that the set of Bambara complex nouns of the form '~oun--~-Noun&quot; could be described using semantic filtering of a context-free grammar. Consider now how a head grammar could achieve a description of the same facts.</Paragraph> <Paragraph position="1"> Assume, to simplify the situation, just two noun stems in Bambara, represented here as ~ and b. The following head grammar generates the language {x The structure this gr---,-r assigns to the string ba-o-ba is shown in figure 2 in the form of a tree with crossing branches, using asterisks to indicate heads (or strictly, nodes through which the path from a label to the head of its terminal string We know, therefore, that there are at least two options available to us when we consider how a case like Bambara may be described in rigorous and computationally tractable terms: semantic filtering of a CF-PSG, or the use of head gr-----rs. However, I would like to point to certain considerations suggesting that although both of these options are useful as existence proofs and mathematical benchmarks, neither is the right answer for the Bembara case. The semantic filtering account of Bembara complex nouns would imply that every complex noun stem in Bambara was of a different semantic category, for the encoding of the exact repetition of the terminal string of the noun stem would have to be in terms of a unique compositional structure. This seems inherent implausible; &quot;dog-catchercatcher-catcher&quot; should have the same semantic category as &quot;dog-catcher-catcher&quot; (both should denote properties, I would assume). And the head grammar account of the same facts has two peculiarities. First, it predicts a peculiar structure of word-internal crossing syntactic dependencies (for example, that in dog-catcher-~-dog-catcher, one constituent is dog-dog and another is doK-catcher~-doR) that seem unmotivated and counter-intuitive. Second, the grammar for the set of complex nouns is profligate in the sense of Pullmn (1983): there are inherently and necessarily more nonterminals involved than terminals---and thus more different ad hoc syntactic categories than there are noun stems. Again, this seems abhorrent.</Paragraph> <Paragraph position="2"> What is the correct description? My analytical intuition (which of course, I do not ask others to accept unquestioningly) is that we need a direct reference to the reduplication of the surface string, and this is missing in both accounts.</Paragraph> <Paragraph position="3"> Somehow I think the grammatical rules should reflect the notion &quot;repeat the morpheme-string&quot; directly, and by the same token the parsing process should directly recognize the reduplication of the noun stem rather than happen indirectly to guarantee it.</Paragraph> <Paragraph position="4"> I even think there is evidence from English that offers support for such an idea. There is a construction illustrated by phrases like Trac 7 hit it and hit it and hit it. that was discussed by Browne (1964), an unpublished paper that is summarized by Lakoff and Peters (1969, 121-122, note 8). It involves reduplication of a constituent (here, a verb phrase). One of the curious features of this construction is that if the reduplicated phrase is an adjective phrase in the comparative degree, the expression of the comparative degree must be identical throughout, down to the morphological and phonological level: (8)a. Kimgot lonelier and lonelier and lonelier.</Paragraph> <Paragraph position="5"> b. Kim got more and more and more lonely.</Paragraph> <Paragraph position="6"> c. *Kim got lonelier and more lonely and lonelier.</Paragraph> <Paragraph position="7"> This is a problem even under transformational conceptions of gr-----r, since at the levels where syntactic transformations apply, lonelier and more lonely are generally agreed to be indistinguishable. The symmetry must be preserved at the phonological level. I suggest that again a primitive syntactic operation &quot;repeat the morpheme-string&quot; is called for. I have no idea at this stage how it would be appropriate to formalize such an operation and give it a place in syntactic theory.</Paragraph> </Section> class="xml-element"></Paper>