File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/w94-0327_metho.xml
Size: 19,043 bytes
Last Modified: 2025-10-06 14:14:01
<?xml version="1.0" standalone="yes"?> <Paper uid="W94-0327"> <Title>Bidirectional Incremental Generation and Analysis with Categorial Grammar and Indexed Quasi-Logical Form.</Title> <Section position="3" start_page="0" end_page="225" type="metho"> <SectionTitle> 2 Generation:from indexed QLF </SectionTitle> <Paragraph position="0"> A working hypothesis of the PLUS project was that strict compositionality provides too man3: meanings lbr efficient interpretation. The alternative is to rely on defeasible reasoning over an underspecified (w.r.t lexical, referential, quantificational and attachment ambiguities) representation.</Paragraph> <Paragraph position="1"> On the generation', side, we adopt a 3-way split between content (i.e. application dictated) planning with output expressed in terms of standard logical forms (LF), linguistic planning (i.e. &quot;how to say it&quot;), with output expressed in QLF, and realisation. Here, we only discuss the last (Jokinen, 1993 describes the second). The two planning components between them need to be able to exercise full control of the linguistic choices, and do so through the QLF, which includes linguistic features as well as predicate-argument structures derived from the LF via the lexical choice process.</Paragraph> <Paragraph position="2"> We might conclude from this reasoning that what we really need as surface generator input is the level of description found in a typical feature structure analysis assigned by a formalism like LFG/HPSG/FUG. Many systems in the NLG literature have adopted this kind of initial description language in preference to logical languages. Our QLF contains the same kind of information as this, encoded in a &quot;flat&quot; representation comprising a set of first order Prolog terms. The flat QLF notation means that the planner need not 'know' about the syntactic form of feature structures as defined by a particular grammar, but simply decide which grammatical constraints hold of each logical element's realisation. That QLF is a quasi logical form can be seen from two properties: (a) It is less expressive in that it lacks scope constructs. (b) It contains &quot;non-semantic&quot; information, such as grammatical or pragmatic properties of linguistic elements corresl:xmding to logical individuals and variables.</Paragraph> <Paragraph position="3"> The latter distinguishes our QLF from the bettcr known one of Alshawi. The non-semantic predicates comprise a closed class and are filtered from the QLF during lexicon lookup. In the example below, pasttime ( 94 ) and nun~sing ( 96 ) are examples of non-semantic annotations.</Paragraph> <Paragraph position="4"> \[def ( 95 ) ,name( 95 ,bill) ,book( 96 ), numsing ( 96 ), long(s (96), 96) ,very(s (96) ) ,indef (96), past time(94),write(94), sub~(94,95),ob~(94,96)\] The generator is also constrained by a syntactic description of the target phrase, but only at the top level.</Paragraph> <Paragraph position="5"> The only properties of QLF relevant to the generation algorithm are that it should be a conjunction of literals, with instantiated arguments, and that each word in the lexicon has at least one QLF term associated with it. 2 From the perspective of the inferential comlxments in the dialogue system, this is a proto-logical form and the relationship between it and LF is beyond the scope of this paper. A benefit of this formalism in relation to our generation algorithm is the simplicity of its manipulation. Since QLF statements are unordered sets, set theoretic operators (e.g. membership, union) suffice for information extraction.</Paragraph> <Paragraph position="6"> Fedder (1991) used a similar algorithm to generate from IpI.I'S: A Pragmatics-based Language Understanding System. Part-funded by the (k~mmission of the European Commu,itics. Project N deg 528.4. See l~lack et al (ITS)3) for an overview.</Paragraph> <Paragraph position="7"> 2 This is a defect of the notation, rexluiring that particles have a &quot;semantics'. This can be remedied prapqmaticallv bv either a procedural attachment to the lexical ent~' of the subcategorising~itcm (w\[aich sacrifices bidircctionality) or by a dummy semantics which can be inserted at the what to ~y stage.</Paragraph> <Section position="1" start_page="225" end_page="225" type="sub_section"> <SectionTitle> 2.1 Lexicon lookup </SectionTitle> <Paragraph position="0"> Lexical lookup from QLF begins by filtering out the predicates that do not correspond to lexemes.</Paragraph> <Paragraph position="2"> In (1), the non-lexical elements argO(1,2) and past_time ( 1 ), are ignored in accessing the lexicon, but after retrieval of the relevant lexical entries, play their part in filtering out inappropriate forms. The functors of the remaining predications are used to index into the lexicon: sleep (1) 's functor sleep corresponds to the lexeme or citation form for the lexical entD', and the non-logical annotation pasttime ( 1 ) will after lookup select the correct form slept. The indexes (1,2 in the example) are coinstantiated between the semantics and the syntax in the individual lexical and phrasal categories, so as to produce a string corresponding to the correct argument bindings. (This does not happen correctly if the indexes are uninstantiated variables, as in a parse result.)</Paragraph> </Section> <Section position="2" start_page="225" end_page="225" type="sub_section"> <SectionTitle> 2.2 Categorial Grammar </SectionTitle> <Paragraph position="0"> The generation algorithm discussed in the next section is not tied to a particular linguistic formalism, but favours a lexicalist formalism with as few rules as possible. This is especially true of CG in which most constituent structure is captured by the two rules of function application. The CG rules of forward and backward function application can be stated as lbllows in the parsing grammar: %%%% Forward application</Paragraph> <Paragraph position="2"> (Their statement in the generation grammar is slightly longer). In either case, the rules are matched by categories recursively defined over the basic categories s,np and n and the directional slash operators / and \. Briefly, an expression of category A/B combines with an expression of category B to form a phrase of category A. An instance is a determiner, category np/n combining with a common noun, category n, to its right, forming a noun phrase, category rip. All expressions in the lexicon belong to either basic or derived categories. To take a complex example, the verb &quot;bet&quot; requires a subject, two object nps and a further sentential object~ and hence has catego~' s\np/s/np/np.</Paragraph> </Section> </Section> <Section position="4" start_page="225" end_page="225" type="metho"> <SectionTitle> 3 The surface generation algorithm </SectionTitle> <Paragraph position="0"> Initial edges are asserted into a chart, for each word in the lexicon whose semantics is subsumed by the target expression's semantics.. As each edge is added to the chart, combinations are made with existing edges, as licensed by the rulcs, and new spanning edges added. * While this description may make the algorithm to appear something of a blind search, it is in fact strongly directed by the elements present in the QLF, supported by an inversion of the indexing used in parsing.</Paragraph> <Paragraph position="1"> The lexicon match is not based on direct unification of the target phrase's semantics with that of its head, a fundamental requirement of the bottom-up head-driven algorithm of Shieber et al (1989) and Van Noord (1990).</Paragraph> <Paragraph position="2"> Relaxing this requirement enables semantically equivalent QLFs (arising from commutativity of &) to be handled directly without any special mechanism. The top-level procedure is stated as follows in Prolog: generate (--Syntax: Semantics, Text ) : abolish(edge, i), generate lex lookup(Semantics,Word,Syn,Sem), acceptable ( Sem, Semantics, Compl ), add_edge ( Syn: \[ Word I R \] : R: Sem: Compl, Word). generate ( Syntax :-semantics, Text ) :ed~e(S~ntax:Text: \[ \] :_: \[ \] ).</Paragraph> <Paragraph position="3"> The lookup procedure retrieves a word whose semantics is a subset of that in Semantics, returning the wordand ts syntactic and semantic description, acceptable/3 ensures that the semantics of the word is a subset of the target semantics, and also returns the &quot;unused&quot; part of the semantics in Compl. Subsequent recursive calls work on Compl, ensuring that constituents are not generated more times (perhaps infinitely) than specified in the target semantics. The second generate/2 clause requires that all elements in the target semantics are consumed. Add edge is a recursiveprocedure that does the main work.</Paragraph> <Paragraph position="4"> addedge/l operates just as it would in a parser: alier adding edges to the chart, any combinations permitted with it are applied recursively, xapply/3 applies the grammatical rules, in this case the rules of categorial function application.</Paragraph> </Section> <Section position="5" start_page="225" end_page="226" type="metho"> <SectionTitle> 4 Type raising and Composition </SectionTitle> <Paragraph position="0"> A forward composition rule and a type raising rule have been added to those of function application, both to the parser and to the generator. Also, topicalization has been added to the parser. In the parsing grammar (Ibr brevity) these rules are stated as follows:</Paragraph> <Section position="1" start_page="226" end_page="226" type="sub_section"> <SectionTitle> 7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 </SectionTitle> <Paragraph position="0"> There have been two motivations for adding these rules to a CG. Firstly, without them, certain co-ordinate and gapping constructs cannot be described neatly. Secondly, they permit incremental interpretation, said to be motivated on psychological grounds. Examples in Section 4.1 illustrate the co-ordinate and gapping constructions that can be treated. With respect to generation, we also find incremental processing well-motivated for interactive systems. Firstly, in the context Of the PLUS project, corpus studies (particularly in French) revealed a great deal of overlap between the turns of the two parties in human-simulated machine dialogues~ and hence the generator needs to be able to begin realisation before the content is fully planned.</Paragraph> <Paragraph position="1"> Secondly, this enables the generator to be incorporated into a distributed or multi-agent architecture, since partial results are available to external evaluation. Thirdly, interleaving interpretation and planning with generation may create in the user a more l:avourable impression of response time.</Paragraph> <Paragraph position="2"> However, the benefits of incrementality are not without their costs. In using rules of function composition, we * I encounter a spurzoza&quot; ambiguity problem. This refers to the multiplicity of derivation paths that are semantically equivalent (and therefore spurious), for the same string, and was first discussed by Wittenburg (1987). This causes multiple generation of identical strings with the same analysis, and an exponential increase in the search space.</Paragraph> <Paragraph position="3"> Fortunately, this problem is already known in the domain of parsing and what we have discovered is that its solution carries over to generation more or less unaltered.</Paragraph> <Paragraph position="4"> The method of Hepple and Morrill (1989) has been used, in both parser and generator, to cope with spurious ambiguity. The main idea is to enforce normal form proofs by cutting the current branch in the search space when a sequence of rule invocations known to lead to non-normal lbrm derivations is about to be made.</Paragraph> </Section> <Section position="2" start_page="226" end_page="226" type="sub_section"> <SectionTitle> 4.1 Coverage of'the Grammar </SectionTitle> <Paragraph position="0"> We begin this section with some illustrative constructs and their representation in the lexicon and in QLF, concluding with an illustration of the non-constituent co-ordination and gapping constructs that specifically motivate the rules of function composition. Intensifier adverbs such as very, quite, really enable sentences like (2) to be parsed or generated. The QLF corresponding to adjectives is a two-place predicate where the first argument is a state-variable. The connection between a state and an object X in that state is denoted using a skolem function s applied to X. Thus, long(X) in a classical logic translation becomes Iong(s(X),X) in the new representation. The full lexical entry for adjectives is given as (3), and (4) is the lexical entD' for intensifier adverbs. (5) shows one of the definitions for &quot;and&quot; which enables sentences like (6) and (7) to be parsed and generated.</Paragraph> <Paragraph position="1"> (2) Bill wrote a very long book.</Paragraph> <Paragraph position="2"> (3) non_infl_lex(Word,n(Agr)#X/n(Agr)#X,\[QLF\]) :- adj(Word), QLF=.. \[Word, s(X),X\].</Paragraph> <Paragraph position="3"> (4) non_infl_lex(Word,(n(Agr)#X/n(Agr)#X)/ (n(Agr)#X/n(Agr)#X),\[QLF\]) :adverb(Word, grad), QLF =.. \[Word,s(X) \].</Paragraph> <Paragraph position="4"> (5) non_infl_lex ( and, C#Res \ C#Le ft/C#Right, \[conj (Res,Left,Right) \] ).</Paragraph> <Paragraph position="5"> (6) Bill and Kristiina wrote a very short book and a long letter today.</Paragraph> <Paragraph position="6"> (7) Bill saw and heard Kristiina.</Paragraph> <Paragraph position="7"> (8) Bill heard and Nancy saw Kristiina.</Paragraph> <Paragraph position="8"> (9) Bill walks and Nancy runs today.</Paragraph> <Paragraph position="9"> (10) Bill saw the man who John heard.</Paragraph> <Paragraph position="10"> (ii) Bill saw the man who heard John.</Paragraph> <Paragraph position="11"> Forward composition and type raising rules cover non-constituent co-ordination as shown in (8) and (9). They also permit analysis of WH-movement as shown in (10) and (11).</Paragraph> </Section> </Section> <Section position="6" start_page="226" end_page="227" type="metho"> <SectionTitle> 5. Incremental Generation </SectionTitle> <Paragraph position="0"> Incremental generation has been introduced (Kempen and Hoenkamp 1982) on psychological grounds, and several reports of surface generators have emphasised this property (e.g. Reithinger, 1991, de Smedt and Kempen, 1991, van de Veen forthcoming). In practical terms, the idea is that we should be able to throw logical statements at the generator, one at the time, as soon as they become available (as a product of a reasoning process in a background application, perhaps), and that the generator should be able to start generating right away, without having to wait for the stream of semantic representations to end.</Paragraph> <Paragraph position="1"> Here we argue: 1) QLF is suitable for specifying the content of the target to be generated incrementally, 2) a chart-based generation algorithm is suitable for incremental generation, and 3) CG rules used can determine the level of 'talkativeness' of an incremental generation system.</Paragraph> <Paragraph position="2"> QLF is a suitable formalism for this kind of job since it is designed especially with the representation of partial information in mind. QLFs can, while still being well-formed in a syntactic sense, codify such things as a predicate-argument structure where one argument is not yet specified, or a lack of knowledge concerning the properties of another argument, and afterwards, at another time, when it becomes available, the missing information can be given.</Paragraph> <Paragraph position="3"> The main strengths of the chart-based algorithm used are that QLF terms are not required in a particular order, or all at once.</Paragraph> <Paragraph position="4"> The only addition to the original CKY generation algorithm is that when no more edges can be added to the chart, the string(s) corresponding to all the QLF given so far is printed; more QLF is requested from the background process; It is then added as 'still to be consumed', and the generation process is called recursivcly from there.</Paragraph> <Paragraph position="5"> To see the role of the CG rules for regulating the talkativeness of the generator, note that edges that have consumed all semantic input at a given point in time, and therefore deserve to be printed, must always correspond to constituents given the grammar. Now, while a CG with only forward and backward application (FA and BA), implies a standard notion of constituency, rules like type raising (TR) and functional composition (FC) give rise to a more generous notion of constituency (this is what makes 'non-</Paragraph> <Section position="1" start_page="227" end_page="227" type="sub_section"> <SectionTitle> 7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 </SectionTitle> <Paragraph position="0"> constituent co-ordination' possible). This means that an incremental generation system of the kind sketched above, employing FA, BA, TR and FC, will be 'chattier' than the very same system employing only FA and BA.</Paragraph> <Paragraph position="1"> For example, assuming only FA and BA, and QLF = {indef(x), black(s(x),x)}, no string would be generated, since np/n and n/n do not form a constituent. Assuming FA, BA, FC and TR, and the same QLF, the string a black would be generated, since np/n and n/n can be composed into the constituent np/n. The string the black cat would be generated under both circumstances, if cat(x) was added to the above set.</Paragraph> <Paragraph position="2"> As another example, consider how the incremental version of the generator, which uses FA, BA, FC and TR, interacts with a user (where the user - input in boldface plays the role of the QLF prcxlucing background process): ?- generate.</Paragraph> <Paragraph position="3"> QLF term: def(x).</Paragraph> <Paragraph position="4"> \[ the l QLF term: man(x).</Paragraph> <Paragraph position="5"> \[the,man\] QLF term: write(e).</Paragraph> <Paragraph position="6"> OLF term: subj (e,x).</Paragraph> <Paragraph position="7"> QLF term: obj (e,y).</Paragraph> <Paragraph position="8"> QLF term: pres(e).</Paragraph> <Paragraph position="9"> \[the,n~n,writes \] QLF term: long(s(x) ,x).</Paragraph> <Paragraph position="10"> \[ the, long, man, writes \] QLF term: indef(y).</Paragraph> <Paragraph position="11"> \[the, long,man,writes, a\] QLF term: short(s(y),y).</Paragraph> <Paragraph position="12"> \[ the, long, man, writes, a, short \] QLF term: letter(y).</Paragraph> <Paragraph position="13"> \[ the, long, man, writes, a, short, letter \] In the same circumstances, but given only FA and BA, neither \[the,man,writes\] nor \[the,long,man,writes,al, or \[the,long,man,writes,a,short\] would be generated.</Paragraph> </Section> </Section> class="xml-element"></Paper>