<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-2103">
  <Title>GENERATION FROM UNDER- AND OVERSPECIFIED STRUCTURES*</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Motivation
</SectionTitle>
    <Paragraph position="0"> The algorithm given ill \[14\] allows to generate fi'om a fltlly specified feature structure, e.g. tile input structure is equal to a structure that would be derived during parsing. For ai)plications otlter than testing a granllnar for overgeneration the equality-condition is too restrictive.</Paragraph>
    <Paragraph position="1"> The algorithm given in \[15\] and \[16\] then Mlows to generate frolu all uuderspceified structure, if there is a fully specified (semantic) predicatc-argontentstructure which is nnt ~dlowed to be extended during generation, e.g. tile l)redicate-argunlent structure must be conqllete and coherent with respect to the target grammar, One of the disadvantages of this algorithm is, that it must be marked for tile generator, which substructure is not allowed to be changed during generation. Further, in certain applications, the condition that there is a partiM feature structure which is complete and coherent with respect to the target grammar might be ,also too restrictive.</Paragraph>
    <Paragraph position="2"> The generator described in this paper had been deycleped for projects whielt are involved in machine translation. While one of the projects makes use only of syntactic information encoded in a feature structure the other in'eject uses semantic information ~s well. In I)oth cases the inI)ut feature structure for tile generator is at least undersl)eeified with respect to *The work reported here is part of the Sonderforschungsbereich 340 Sp~chtheo,'etische G~ltndlagen der ('omputerlinguPS~tik l For details of the LFe, formalism see (1 b the target grammar, not only for al;omic attribute value pail's but also fro' complex pairs. This means tile gencrator has to introduce information into the given feature structure to get a structure which is valid with l-espect to tile target grtunmm~r.</Paragraph>
    <Paragraph position="3"> In both projects a similar architecture is used: 2  1. parse a sentellCe and return the feature structure Fp 2. extrat:t tile inforlnation for the translation from Fp and build F,j 3. generate fronl F 9 a sentence  In such an architecture the creation of Fg is usually independent of the target grammar, in the sense that the creation is not automatically coutroUed by tile target gralnular.</Paragraph>
    <Paragraph position="4"> In machiuc traaslation the grammars used for parsing and for generation are basically spccilic for tile two single languages one wants to translate between. It is usually desirable to sl)eeify F~ only in ,~s rudimentary and ms general lnauller ;L~; possible. This lueans tile details of how to generate a wdid surface string of tim target language are only known in the target grammar, rather than spelled out ill th&amp;quot; translation relation. Ill other words, a single grammar G describes only the relation of a surface string of a language L and a feature structure valid for tile grammar G of L. ~trther, a valid feature structure for G will represent only information necessary for L, but not neeessarily information necessary for the lauguage to translate into. For example, a gramlaar fro' German will describe a fl~atttre structure which h,'us information for the tenses past, present, and future, but no information about progressive ms it is required for English. Therefore, ill tile translation German to English the generator has to generate froln a feature structure which might be underspecified with respect to tense information, while ill the translation Englislt to German the generator has to generate from a feature structure which might be overspecified with respect to tense information.</Paragraph>
    <Paragraph position="5"> ht general, in describing the translation relation between two languages one lta.s to face tile probleuts of interfaces: * Infornmtion is missing and must be derived froin tim target gralnmar, e.g. tile input structure is uuder,~pecified.</Paragraph>
    <Paragraph position="6"> 2For the re~ons of this architecture see for example \[4\]. There are also other MT projects like GRADE (see \[9\], \[10\] and \[8\]) which nl~tke use of a similar architecture. ACRES DE COLING-92. NANTES. 23-28 AOt3T 1992 6 8 6 Prec. oF COLING-92. NANTES, AUG. 23-28, 1992 * There is more information than defined by tile target grammar, e.g. there is no string of the target language for which the grammar describes a feature structure which contains all attributevahle pairs given ill the iuput structure FS 9. The input structure is overspccifled and the overspceif)cation could be ignored duriug geueration.</Paragraph>
    <Paragraph position="7"> * There is informatiou which is incousisteut with the target grammar, e.g. the input structure is illforrned with respect to the target gramnlar.</Paragraph>
    <Paragraph position="8"> This requires some error treatment.</Paragraph>
    <Paragraph position="9"> All algorithm for generation then h~s to provide uwchanisms which allow geueration from underspecifled structures as well as from overspecilicd ones. This will allow to deal with certain types of trauslation mismatches as they are described for example in \[2\].</Paragraph>
    <Paragraph position="10"> Further, the treatmcut of fllformed structures shouhl be such. that the invldid elements of the input structure could he made visible for debugging purposes, illstead of just failing to generate anything. As it turned ont, even for ulediuul sized grallllnars it Call beconle quite dill)cult for a linguist to debug the grammar if there is only a debugger available which had been develolled for the generM l)urpnse programming language the system is inq)lemented ill, e.g. prolog.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Terminology
</SectionTitle>
    <Paragraph position="0"> The alger)tirol has been devehlped for grammars written in the Ll.'c;-formalism. This uleal!s, it works on a eoutext-frec grammar G with annotated fcatm'e descriptions. Given a feature structure FSi, as input the algorithm has to generate all those surface strings, for which G ;Lssociates a feature structure FS,j, with FSI~ coutpatihle to FS,~.</Paragraph>
    <Paragraph position="1"> V~rhat co're, pal)hie means depends on tile kind of application the generator is used iu: * If the application is to test a grammar for overgeu(,ration, FSin lnust lie equal to FSu, e.g. lie iuformation is introduced into or deleted from FSi,, during geueration, and \]i~Si,, unifies ill terms of feature unification with FS,j.</Paragraph>
    <Paragraph position="2"> * If the alll)licatiou is to test whether a structure of a certain attribute might be sufficient for generalieu, i.e. whether the senlautic structure does not m'ergenerate, FSI,~ must I)e subsumed by FS,~, e.g. all information of FSI,, nlust be required for generation, and it is only allowed to introduce iMonnation lute FSin.</Paragraph>
    <Paragraph position="3"> * If the application is machiue trauslation, FSi,, and FSI~ must unify, e.g. FSI,, might contain nlore inlorulation and ,also less iuforluatiou th~t.u FS u .</Paragraph>
    <Paragraph position="4"> Del)endiug on tile al)l)licati(m the algorithm is i)arametrized as to whether it allows the introduction of information into FSi, and whether it allows FSI, to be overspecified.</Paragraph>
    <Paragraph position="5"> For those not familiar with LFG I will give a short overview of tile elements of the feature descriptious as I will use them afterwards. In general a feature deseril)tiou consists of a coujuuction of equations or a disjunction of feature descriptions. In this paper I will only cousider feature descriptions without disjunctious. The equations are distinguished into * defining equations indicated by tile operator = * inequatimts indicated by the operator # * constraining equations indicated by the operator =e All equation consists of a reference to a structure, tile el)era)or , and ,'L,~ second argulueut of the operatiou oue of * all atomic v~due like raas * a semantic form, indicated by double quotes, with an atou)ic uaule aud all optional arguuleut list,, i.e. &amp;quot;man&amp;quot;, &amp;quot;give (SuuJ,ot~J}&amp;quot; * a referellee to a structure A reference to a structure is either a mete-variable or a path applied to a mete-variable. Examl)les are * the meta-wtriable 1, which stands for the structure assnciated with tile nlother llode, e.g. the category given on tile left hand side of a rule.</Paragraph>
    <Paragraph position="6"> * ttw meta-variMilc 1, which stands fur tile structure a.ssociate(1 with a (laughter uode of a rule, e.g. the nolle on the right hand side of a rule where tile feature description is an annotation of.</Paragraph>
    <Paragraph position="7"> * (~ GENI)ER), which refers to a structure under the attrillute (;\[.;NDI.~R ill tile feature structure associated with tile mother node.</Paragraph>
    <Paragraph position="8"> Equations, which have references on both sides of a equatiou arc called ree~ttr(trtey equations.</Paragraph>
    <Paragraph position="9"> Semantic forms describe unique vMues, e.g. while two atoufic values unify if they are described by the same fern), two semantic forms will not. The arguments of a semantic form of at) attribute A are paths which are members of the governable f~mctions of A. This set will be named as gf (A). %) alh)w semantic forms )~s possil)le values tilt ally attribute is a generalization of the Ilse of sltnlantic forlllS a,s they are given in \[1\] where semantic forms are only values of the attrihute PRED. Semantic forms contain all information ueeessary to test the conditiolm of COml)leteness and coherence.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Coherence and Completeness
</SectionTitle>
      <Paragraph position="0"> Using the generalization tile conditious of completeness and coherence ms given in \[3, pp. 211/212\] are reformulated ~s * A feature structure 5' is locMly complete iff for each attribute A in S where gf(A) is non-empty tile governable functions defined by tile vMue of A exist ill S with a value for the attribute A, and if all values required are defined. A structure is conq)lcte if all of its substructures are locally complete.</Paragraph>
      <Paragraph position="1"> * A feature structure S is loeMly coherent, iff for each attribute G of S which is member of gf (A) G is governed by the value of A, e.g. the argulueut list of the vMue of A contains G, and if ,all attributes of S are given by tile grammar. A structure is coherent if ,all of its substructures are locally coherent.</Paragraph>
      <Paragraph position="2"> Ac~rEs DE COLING-92. NnivrI,kS. 23-28 ^olyr 1992 6 8 7 PRec. OF COLING-92. NANTES. AUG. 23-28. 1992 The struettn'e FS derived in the generation process must at least fttllfiqll these contlitions of completeness and coherence, e.g. ally violation of one of these conditions is treated as an error. Since the input structure FSi,, should be part of the derived structure, the conditions for attribate-valae pairs of the input structure are modified to be able to use the input structure to control the generation process and to bc able to allow overspecification, a * If an attribute A of FSi, is licensed by a defining equation or inequation in the rules of tile grammar which are not explicitly excluded by FSi,, it shouhl be checked that A is actually constnned daring generation. This condition extends the condition of coml)leteness.</Paragraph>
      <Paragraph position="3"> * If an attribute A of FSi, does not occur in any equation of the graulmar, tim input structure is ovcrspecified. It depends on the application, whether this type of overspeeification is allowed, e.g. whethcr it should be considercd a.s a violation of the coherence condition or shoultl be ignored.</Paragraph>
      <Paragraph position="4"> * If an attribute A of FSi, is not lieeased by a defining eqnation or an inequation in the rules of the granunar which are not explicitly excluded by FSi, the input structurc is overspecified. It depcnds on tbc allplication whether this type of overspecifieatiml is allowed. In ease overspecification is allowed, A and its value are ignored, otherwise it is treated ,as a violation of the coherence condition.</Paragraph>
      <Paragraph position="5"> As indicated by tile last extension to the coherence and completeness conditions, it depends on the application what kind of input structure is considered to be a valid one for the target gralonlar. Ill case a grammar should he tested for overgeneration a valid input structure is not allowed to be extended tlnriug generation and is not anowed to be ow~rspecifictl.</Paragraph>
      <Paragraph position="6"> In the case of machine translation the input structure can be considered as a valid one, even it is underspecified. Del)ending on the language pair it might be also apl)ropriate to consider an overspeeified input structure ms valid.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The Algorithm
</SectionTitle>
    <Paragraph position="0"> The algorithm works on a granmmr tlescription and an input feature structure. The grammar description cuasists of context free rules with annotated feature descriptions.</Paragraph>
    <Paragraph position="1"> For siml)licity it is assumed that the annotated feature descriptions do not contain disjunctions. A disjunction in a feature description can always be transformed into a disjunction of nodes on the c-structure level. Furthernmre, a siugle ode is a concatenation of terminal and uon-termiual nodes, and for each category C of a grammar the rules for C are treated as one disjunction.</Paragraph>
    <Paragraph position="2"> aThis mealm, it is not sufficient to require, that the inptlt structure has to Ilnify with a structure derived from the grammar to get a generatim~, since this would allow to produce sentences which do not contain all of the semantics given in the inptll structure as well ms to produce sentences with any kind of possible modifiers the grammar could derive, that is infinile many.</Paragraph>
    <Paragraph position="3"> Tim algorithm starts witb a current category C~, initialized with the gual category, and a current feature structure FS~, initialized with the input feature structure FSin.</Paragraph>
    <Paragraph position="4"> The algorithm proceeds as follows: * Match the current feature strncture FS~ with the current category C~ by matehiug FS~ with the feature descriptions FDi of the nodes Ni on the right hand side of tile rule for Cc, where FSc is bound to the mata variable T which deaotates the structure associated with the nlother node C,, on the left hand side. The matching works  top-down I)readth-first. During tile match FS~ will lint be nmdified.</Paragraph>
    <Paragraph position="5"> * Eztend FS,. by the application of a feature description FD.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Matching
</SectionTitle>
      <Paragraph position="0"> The matching of the current feature structure FSe with the current category C~ will always te,'minate.</Paragraph>
      <Paragraph position="1"> During the matching a structure which is used as a chart and an agenda is built which keeps track of * which structures are already matched with which categories.</Paragraph>
      <Paragraph position="2"> * whether there occurs a trivial recursion, e.g.</Paragraph>
      <Paragraph position="3"> given a structure and a category there is a recursion on tile c-structure level which uses the salne structure.</Paragraph>
      <Paragraph position="4"> * tim use of whicb nodes can be constrained by tim input strncture, and what is tile result, e.g. is the usage of the node excluded or licened by tile input structure.</Paragraph>
      <Paragraph position="5"> * which nodes are lmrely eontroUed on tile e-structure level, e.g. there it~ no equation for a node which dcnotates the structure of the mother node. Such nodes bare to produce only finite many snhstrings.</Paragraph>
      <Paragraph position="6"> For each category C ~fll its rules arc considered in parallel, which avoids ally dependency almut the ordering of the single rules for C.</Paragraph>
      <Paragraph position="7"> For each node N on the right hand side of C~ the input feature structure is matched with its feature description FD. This match results ill at least one of the following descriptions: Exclusion: FSc is not coml)atil)le with FD. Therefore the node N will be excluded. Other results of the matching are of no relevance. The exclusion of N excludes those nodes which are part of the same rule as N.</Paragraph>
      <Paragraph position="8"> Activation: FD defines a path-value-pair which is already part of FS~, or FD defines a reentrency which already exists ill FSc.</Paragraph>
      <Paragraph position="9"> Examination: In FD occurs a reentranee equation where only one of the paths exists ill FS~. The result ezamination contains the category CN named by the node N and tile associated sub-structure FS..</Paragraph>
      <Paragraph position="10"> Tile folh)wing cases are distinguished: Amw~s DE COL1NG-92. NANTES, 23-28 AOt~q&amp;quot; 1992 6 8 8 PROC. Or COLING-92, NANTES, AUO. 23-28, 1992 trivial equation: N is a non-terminal node.</Paragraph>
      <Paragraph position="11"> The catgories C,: and CN are associated with tile same (sub)structure. Beside 1&amp;quot; - .\[ equations uf the form (1 X) = (1 X) are also considered ,as triviM equations.</Paragraph>
      <Paragraph position="13"> gory CA&amp;quot; will be matched with the structure denotated by (~ X).</Paragraph>
      <Paragraph position="15"> category CN will be matched for (.\[ Y) with the structure denotated by (1 X). This ease covers the treatment of multiple ro~)ted structnres a-s they nlight occur in gralnnlars written in all IIPSt; style 4.</Paragraph>
      <Paragraph position="17"> with tile structure denotated by (1 X).</Paragraph>
      <Paragraph position="18"> Uncontrolled: FD does not contain any equation which can be applied oil FSc. In this case FS~ does not eontroll the oceurcnce of tile substring associated with the node N, and it depends on tile partial c-structure alone given I W the category C~, whether there are tinite ninny sub-strings described.</Paragraph>
      <Paragraph position="19"> Suspension: FD contains equations which allow controll of generation by FS,., but FS,, does not contain enough information to make a (tecision al)out exc\[usiolL activation or exatninatiolt. Therefore, the matching of N with FS~ has to he decided later. In case the application forbids introduction of infornmtion into FS~. during gem eration the conditions of suspension will lead to immediate exclusion.</Paragraph>
      <Paragraph position="20"> Only tile results activation and examination may occure in parallel The result examination causes a further exanfination of the category CN with tile selected (sub)-structure, if they have (lot Mready ()(!eli eXalllined and are not already under exaluillalion. Thus tho matching of a category with a (sub)structure is performed only once during the matching of the input feature structure with the goal category. This guarantuecs the termination of the matching and is efficient.</Paragraph>
      <Paragraph position="21"> Since the matching works top-down breadth-first it is llOSSible to detect inconsistencies between the iupttt feature structure and parts of the rules fairly early. From the complete match it is possible to deter: mine the set of these attribute=value pairs, which are part of tile original input structure and which could I)e used either by a defining equation or all incquation. These attribute-value pairs are marked that they have to be used which is an equivalent of adding temporarely constraining equations to the grammar, which guarantee that a maxinmm of illformation from the input structure is used for generation. It should be noted, that this step is only necessary, if overspecification of the input structure is allowed. Otherwise all attribute value pairs of the input structure could be marked at star(up that they have to be used during generation.</Paragraph>
      <Paragraph position="22"> The matching produces a set of IIossible solutions.</Paragraph>
      <Paragraph position="23"> This makes it possible to distinguish a failure caused by an illegal input structure from the generate-and-test Iiehaviour of the backtracking *nechanism. Since 4 For a description of Ites(\] se~ 11 l\].</Paragraph>
      <Paragraph position="24"> there is enough illfornlation of the current goal in tile generation process, it is possil)le to produce an error message which descril)es * the c-structure build so far * the node and its ammtated feature description which is inconsistent with the input structure * the part of tile input structure which caused the failure ~l(ch all error luessage vC/onld lie in tern(s of the grammar rather than in terms of the iinplenlention lan= guage of the algorithm. An error message eouhl be I couldn't yenemtc aTt NP for the structure \[ PRH) (ua((\] spt:c idef J because SPEC' : idef is ille.rlal for the grammar.</Paragraph>
      <Paragraph position="25"> Since it is distinguishc&lt;I which parts of the structure are intruduccd during generation it is possible to show tufty those faihu'es which are caused by the original input structtu'e. This would also allow one to ignore illegal parts of the inliut structure emnpletely alld t\[) ev~211 ~Cllcl';ttc fr()lll illformcd structures. In con(flint to the cmue of overspccification this would require repairing either tile input structure or extending tile target gr~.(nlllar.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Extension
</SectionTitle>
      <Paragraph position="0"> Tile extension of FS~ by a feature description FD means, that all information fi'om FD is incorporated into FS,,. Since only non-disjuuctiw~, feature descriptions are cmtsideretl it is not necessary to describe tile treatment of disjunctive information. The only source of alternatives are the rules. These alternatives are treated by backtracking. The selection of alternatives starts with those disjuncts, which do not lead to reeursion. This guarantees that recurs(on is applied oaiy in those ca.ses, where it could be part of tile c-structure to generate.</Paragraph>
      <Paragraph position="1"> The extension h~t~ several aspects. First, it is made explicit in tile feature structure which attrilmte-value pairs are defined by the grammar, and how often a definition h~u oceured during tile generation. The latter information is used to stop the generation from infinite loolis I)y giving a maximum amonnt of repeated definitions of the same l)ieee of information. Reasonable limits are values between 10 and 20. It should be noted that the semantic foT~ns of LFG reduce this linfit to 1 for attributes which take a semantic for((( as value 5.</Paragraph>
      <Paragraph position="2"> Second, a partial representation of the e-strncture is built in parallel to the feature structure, which allows at the end of the generation process to extract the surface string by a traversal of the complete c-structure.</Paragraph>
      <Paragraph position="3"> Third, it can be deternfined which attribute-value pairs have been introduced into the original structure. Only these attrilmte-value pairs are relevant to reexamine suspended nodes.</Paragraph>
      <Paragraph position="4"> SFor LFG grammars this aspect of semantic forms is the main reason that tile generation will terminate without the superficial limltati!m of repeated definitions.</Paragraph>
      <Paragraph position="5"> ACRES DE COLING-92, NANTES. 23-28 AOI3T 1992 6 8 9 PROC. oF COLING-92, NANTEs, Auo. 23-28, 1992</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.3 The main loop
</SectionTitle>
      <Paragraph position="0"> 1. For each node Nj of the right hand side of the rule of the current category Cc match tlle annotated feature description FDj with the current feature structure FS~. The matching ternfiuates always, and during the matching no new information is introduced into FSc. The match determiues, whether the node Nj might be excluded, activated, suspeuded, and whether the category N should be examined for some part of FSc.</Paragraph>
      <Paragraph position="1"> 2. If there are uo nodes left which can be activated, nodes which are still suspended axe excluded attd tile filial coherence and completeness tests are performed on the input structure FSI,. In case of success the surface string can be extracted from the c-structure which is built in parallel to the derivation of the input feature structure.</Paragraph>
      <Paragraph position="2"> Ill case of failure, other solutions are tried by backtracking.</Paragraph>
      <Paragraph position="3">  3. Select only these nodes which can be activated which will not lead to a recursion. Extend the partial feature structures associated with these nodes by applying the annotated feature descriptions. null 4. Compaxe those nodes again which have been suspended ms in step 1.</Paragraph>
      <Paragraph position="4"> 5. Repeat the steps 3 aud 4 until there are no nodes left which can be activated aud which do not lead to it recursion.</Paragraph>
      <Paragraph position="5"> 6. Nodes which could be activated but lead to recursiou axe activated only in case there is ltO indication that the recursion conld be applied infinite many times s.</Paragraph>
      <Paragraph position="6"> 7. Contimte with step 2.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 Example
</SectionTitle>
    <Paragraph position="0"> In order to illustrate how tbe algorithm works, I will oaly give a very simple and somewhat superficial example. For more detailed examples especially on the treatment of recursion see \[5\]. 7 The exantple makes nse of the grammax in figure 1 to generate a German sentence with a simple NP and all intransitive verb. The grammar is written ill a usual LFG notation. The input feature structure for generatiun is given in figure 2. For the example it is assumed that the feature stucture contains the semantic representation of the analysis of the Englisb sentence the man is running which should be translated into German, The goal category for generation is S.</Paragraph>
    <Paragraph position="1"> The generation starts with the matching of S with FSo. The NP node of the right haud side of the S rule is suspended, since there is no attribute SUBJ in the input structure. The trivial equation of the V1 a node immediately leads to the matching of FSo with the category VP. The trivial equation on the V node leads in turn to the matching of the category V with FSo. The existence of (SEM REL) = r~n in FSo</Paragraph>
    <Paragraph position="3"> would allow to activate botb verbs of the example lexicon, but the equation (T SEM TIME END) = past excludes the eutry for rannte.</Paragraph>
    <Paragraph position="4"> The resulting partial c-structure of the match is  Since tile solution set of the match does not require to use (SEM TIME END) tiffs information can be ignored for the further generation, although it had been used to exclude an entry. This shows a case of overspecifieatiou, where an attribute is in the set of possible attributes of a gramntax but is not always determined by the grammax.</Paragraph>
    <Paragraph position="5"> The extension of FSo then leads to the structure in figure 3. It should be noted that the algorithm autontatically selected the semantic head, although ACTES DE COLING-92. NANTES, 23-28 AOt~q&amp;quot; 1992 6 9 0 PROC. OF COLING-92, NAh'TES, AUG. 23-28, 1992 feature structure c-strnctnre \] &amp;quot;rau&amp;quot; I +L +,,L altG1 53/S,','C/ clef / / / \[\] s,+,\[\] sg '//  the bead is eml)edded in at substructure. Tiffs means the algorithnl is implicit head-driven without any assunq)tions which part of an inj)ut structure the head should be. As it is shown ill \[5\], this allows to generate in cases of head-switching, where syntactic att(l semantic head differ.</Paragraph>
    <Paragraph position="6">  Tit(&amp;quot; introduction of SUBJ leads to tim matching of the suspended NP imde with FSo. The equation</Paragraph>
    <Paragraph position="8"> For the NP rule there are three nodes to be matched with FS4. Siucc on all three nodes a trivial equation is atmotated, the categories D and N have to be matched with FS4. The equations (l SEM REL) = man and (T SEM NUM) = sg activates tile noun curry, and requires that (SEM ltEL) and (SEM NUM) of FS4 nlust be nsed for geueratiou.</Paragraph>
    <Paragraph position="9"> The equation (1&amp;quot; SEM SPEC) = dcf activates the determiner entry and requires to use (SEM SPEC) of FS 4 .</Paragraph>
    <Paragraph position="10"> The two alternatives of the NP rule &amp;quot;allow to consider two lmssible extension shown in table 1.</Paragraph>
    <Paragraph position="11"> Since (SEM SPEC) of FS4 must be used, the second alternative will be rejected by tile final constraint test. Therefore, the only solution is tile first alternative. This results in tile e-structnrc</Paragraph>
  </Section>
  <Section position="8" start_page="0" end_page="0" type="metho">
    <SectionTitle>
6 Comparison with Shiebers approach
</SectionTitle>
    <Paragraph position="0"> The semantic-head driven ,algorithm giveu in \[13\] also starts with a tol)-down initalizatiou with a I)ottom-u l) generation. In Shieber ct al the nodes whicll eoutam the semantic head arc determined during tile couq)ilation of the grammar. This seems to be a bit problenmtic fur gramluars which describe head-switching t)henomcnons, ~ in 100 l~tres of wine, where a possibh~ ananlysis is that 100 litres syntactically governs ultn.e, but semantically is a moditicr of wine. The algorithm llreseuted here does not require to llrecomlmte tile nodes which contain tile semantic head, but finds the head relewmt for the giveu input structure automatically.</Paragraph>
    <Paragraph position="1"> Tile problem with free variables for the coherence constraint given in Slficbcr ct al does not occur for the alguritbm l/reseuted in this paper, since it &amp;quot;always distinguishes between the struetnre and the descril)tiun of the structurc, and keeps track of which parts of the structure are already derived during generatiun. Since the a\[gorithln I)resented here always hmq infurmatiml at)out wlfi(:h parts are from the original input structure and which ones have been added, it is possible to check the coherence couditiuu at any step of the generation process. In addition, the so lution in Slfieber et al with binding variables seems somewhat llroblematic, since it requires to know for sure, that the variable part of the semantics shouht uot lie exteuded.</Paragraph>
    <Paragraph position="2"> The augmentation of the generator described ill Shiet)er et al with a chart to avoid rccomputation att(l elinfinate redtmdaucies is an integral part of the algorithut presented here.</Paragraph>
  </Section>
class="xml-element"></Paper>