File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/abstr/88/c88-2123_abstr.xml
Size: 27,675 bytes
Last Modified: 2025-10-06 13:46:35
<?xml version="1.0" standalone="yes"?> <Paper uid="C88-2123"> <Title>A Syntactic Description of German in a Formalism Designed for Machine Translation</Title> <Section position="1" start_page="0" end_page="594" type="abstr"> <SectionTitle> Abstract: </SectionTitle> <Paragraph position="0"> This paper presents a syntactic description of a fragment of German that has been worked out within the machine translation project Eurotra. It represents the syntactic part of the German module of this multilingual translation system. The linguistic tool for the following analyses is the so-called CATframework. null In the first two sections of this paper an introduction of the formalism and a linguistic characterization of tile framework is given. The CAT formalism as a whole is a theory of machine translation, the syntactic analysis part which is the subject of this paper is an LFG-like mapping of a constituent structure onto a functional structure.</Paragraph> <Paragraph position="1"> A third section develops principles for a phrase structure and a functional structure for German and the mapping of phrase structure onto functional structure.</Paragraph> <Paragraph position="2"> In a fourth section a treatment of unbounded movement phenomena is sketched. As the CAT-framework does not provide any global mechanisms I try to give a local treatment of this problem.</Paragraph> <Paragraph position="3"> O. Introduction There are two basic givens for Eurotra: (i) Stratificational description of language.</Paragraph> <Paragraph position="4"> The description of language consists of an analysis on three levels: ECS (Eurotra-Constituent-Structure) which describes language according to part/whole relations and word order, ERS (Eurotra-Relational-Structure) which describes language in terms of syntactic functions and IS (Interface Structure) which describes language according to deep syntactic relations enriched by semantic information such as semantic features for characterizing lexical units.</Paragraph> <Paragraph position="5"> (ii) The CAT-formalism.</Paragraph> <Paragraph position="6"> The CAT-formalism is the linguistic tool for the description of language. As this formalism has no global mechanisms there are some restrictions concerning the treatment of unbounded dependencies.</Paragraph> <Paragraph position="7"> Taking thc.~;e givens into account, I would like to present the following topics: (i) An introduction to the formal language as far as necessary for the treatment of the linguistic phenomena 1 would like to describe, (ii) A characterization of the Eurotra stratifieational description of language as a functionally oriented theory, (iii) A development of principles for a syntactic description of German (iv) A sketch of the treatment of unbounded dependencies 1. The formalism I would like to introduce only those parts of the CAT-formalism &quot;which build the basis of my analyses. That is two kinds of rules: (i) soealled b-rules (structure building rules). They build structures qud transform structures into structures.</Paragraph> <Paragraph position="8"> (ii) so-called feature-rules and killer-filters. They are put together into one class as both of them operate on structures created by b-rules expressing generalizations over attributes. 1.1. b-rules (1)(a) {cat=s} I {cat=np},(cat=vp} 1.</Paragraph> <Paragraph position="9"> (b) {eat=vp}l (cat=v},{cat=np} l.</Paragraph> <Paragraph position="10"> (c) {cat=np}\[ {cat=det},{cat=.} \].</Paragraph> <Paragraph position="11"> (d) (cat=v,lu=kaufen,lex=kau ft,t.s=tensed} In (l)(a)-(d) we have b-rules, which define a small ECSgrammar. (d) is a rule for a terminal. The dominance relation is expressed by square brackets. The grammar in (1) assigns sentence (2)(a) structure (2)(b).</Paragraph> <Paragraph position="12"> (2)(a) Das Haus kauft der Mann (The house, the man buys) (b) s I np vp det n v np I __I__ det n I I das haus kauft der mann The same way as in (l)(a)-(d) an ECS-grammar was written we can write b-rules defining functional structures. (3) is a b-rule defining the functional structure for (2)(a): (3)(a) {cat=s} \[{sf=gov,cat=v,frame=subj obj}, (sf=subj,cat=np,case=nom}, (sf=obj,cat=np,case=acc), * (sf=mod}l. (sf=syntactic function) (b) {l.=kanfen,sf=gov,frame=su bj obj }.</Paragraph> <Paragraph position="13"> b-rule (3) creates the functional structure (4) for sentence (2)(a). (4) _,s</Paragraph> <Paragraph position="15"> kaufen mann der haus das The transformation will De done by tile translation b-rule (5).</Paragraph> <Paragraph position="17"> S:{cat=s} <V,NP2,NPI>.</Paragraph> <Paragraph position="18"> A translation-b-rule (t-rule) consists of a left hand side (Ihs) which defines a representation, in our case it would unify with the ECS-strueture in (2)(b), and a right hand side (rhs) which defines a dominance and precedence relationship between the items represented by the variables (capitals). If there is a b-rule on the next level, in our case ERS, which satisfies these conditions, the translation succeeds, t-rule (5) says that structure (2)(b) shall be translated into a structure which is dominated by a node of category s which dominates the three items represented by the variables in the order given in the rhs of the t-rule. As the verbal governor, in our case 'kaufen', requires a subj obj frame, expressed by the frame feature, (3)(a) is tile ERS-b-rule which would match with the rhs of t-rule (5) and create (4).</Paragraph> <Paragraph position="19"> 1.2. f-rules and killer-filters 1.2.1. f-rules f-rules and killer filters allow for the definition of a context part (those features after the slash) and an action part as example (6) shows. An f-rule applies to a representation only if the context part strictly unifies with the object.</Paragraph> <Paragraph position="21"> (6) says that for each np consisting of a det and an n case, number and gender of the n have to be percolated into the mother node.</Paragraph> <Paragraph position="22"> I would like to make two remarks: (i) the feature percolation in example (6) could be done in b-rules. Thus, it might seem that f-rules are superfluous. However, as section 4 will show, there are many cases where we need feature percolation by f-rules. (ii) I will make a special use of f-rules. I will take everything as context and action part. That means, if f-rule f unifies with representation r, r will be replaced by the result of unification, if not, r survives unchanged.</Paragraph> <Paragraph position="23"> 1.2.2. killer-filters Killer filters specify structures which are not well-formed and which therefore have to be deleted. We might imagine a rule which kills nps having a pronominal head and an np in genitive. (7) killer-filter: {cat=rip} \[{cat=detJ,{cat=n,n type=pron},{cat=np,case=gen } I. 2. CAT as a functionally oriented framework 2.1. A comparison with a configurational framework For the linguistic ctmracterization of the Eurotra framework 1 would like to make a brief comparison between two kinds of linguistic theories: (i) those which assume syntactic functions as universal primitives of language (prototypical: LFG) (ii) those which claim that syntactic functions could be reduced to configurational facts (prototypicah GB).</Paragraph> <Paragraph position="24"> Each of the two possible ways of describing language forces the linguist to describe linguistic facts as word order, binding relations, agreement, case assignment or long distance movement in a certain way.</Paragraph> <Paragraph position="25"> The configurational framework claims that there is a general schema for phrase structure rules which is the universal pattern according to which all constituent structures of all possible languages are built. It is the x-bar schema: On this structure movement rules operate creating S(urface)structures. null So, this is the kind of explanation a configurational framework gives: There is a canonical schema (the x-bar schema) and each configuration not fitting into this schema is explained as derived by universally restricted movement transformations.</Paragraph> <Paragraph position="26"> The functional alternative has to rely on syntactic functions as universal primitives. So, phrase structure does not necessarily claim a universal status, and movement rules are not even necessary. This requires a different treatment of the linguistic phenomena. How does the CAT framework fit into this? The adoption of the three level system (ECS,ERS,IS) makes Eurotra functionally oriented as it adopts the way of linguistic description a functional approach has to adopt. While a configurational desoription consists in mapping given configurations onto a canonical schema, the x-bar schema, by explaining configurations which do not fit into x-bar as having undergone movement transformations, a functional description consists in a mapping of phrase structures onto functional structures.</Paragraph> <Paragraph position="27"> 2.2. Completeness and coherence in Eurotra There is an essential which holds for all functional frameworks, namely the completeness and coherence principle.</Paragraph> <Paragraph position="28"> This principle says: A functional structure is well-formed iff it is complete and coherent. A functional structure is complete iff it contains all the syntactic functions required by the frame of the framebearing element. A functional structure is coherent iff it contains only the required syntactic functions. Enrotra allows for the expression of this principle in two different ways: (i) Enumeration of frames The ERS grammar has to enumerate all possible patterns, all frames which are possible, as b-rules, and the value of the frame feature of the gov determines that only the wanted and nothing but the wanted governors go into the structure building rule.</Paragraph> <Paragraph position="30"> In (9) completeness is expressed by the fact that both framebound syntactic functions are obligatory. So, if one of the functions is missing, the structure is not well-formed.</Paragraph> <Paragraph position="31"> Coherence is expressed by the fact that the structure building rule only allows for the two syntactic functions and nothing else.</Paragraph> <Paragraph position="32"> This prevents e.g. the creation of an oblique object.</Paragraph> <Paragraph position="33"> (ii) Completeness and coherence by f-rules and killers There is, however another way of expressing completeness and coherence which does not require the enumeration of all frames. We need the following rules: (a) One ERS b-rule which enumerates all possible syntactic functions optionally as (I0) does:</Paragraph> <Paragraph position="35"> (b) A separate encoding of the functions a verb is subcategorized for, i.e. the frame feature is given up and a feature for each syntactic function is introduced: (11) {lu=see,subj=yes,obj=yes,sf=gov} All other syntactic function feature values will have to get the default &quot;no&quot; (by default f-rules).</Paragraph> <Paragraph position="36"> (12) {lu=see,subj=yes,obj=yes,obj2=no,obl=no,scomp=no,sf=gov} We can now state completeness and coherence independently by killer filters: (13) :k: {cat=s} l {sf=gov,cat=v,subj--yes},</Paragraph> <Paragraph position="38"> (13) determines that if the feature for subj=yes then there must be a syntactle function &quot;sub j&quot; in this representation. Expressed by a killer it reads: if there is a structure whose gov has an a-vpair subj=yes and contains only functions which are not subj then this structure is not well-formed. The same which has been stated here for subj can be stated for all functions. To get coherence we use a killer filter as in (14).</Paragraph> <Paragraph position="39"> (14) :k: {cat=s) \[ {sf=gov,cat=v,subj=no}, {sf=subj,cat=np,case=nom}, &quot; 0 I (14) says: If the structure whose gov has the feature-value 'no' for the feature 'subj' contains a feature bundle containing the feature sf=subj, plus anything else, then this structure is not well-formed.</Paragraph> <Paragraph position="40"> 3. Syntactic description of German 3.1. Principles of syntactic description As we have seen above, the syntactic description of a language in Eurotra follows a functional approach. In our description this is not only reflected by the existence of a functional level but also by the uonhierarehical, nonconfigurational description of the sentence constituent we offer. As we do not use the given x-bar schema we need no empty elements on ECS and we describe German as a uonconfigurational language.</Paragraph> <Paragraph position="41"> Though in German matrix clauses we have SVO word order, German is usually considered an SOV language. Matrix clause word order is considered as derived from subordinate clause word order by movement transformations. (of. Koster 1975, Thiersch 197~ L Reis 1985..</Paragraph> <Paragraph position="42"> On this basis we would like to make another assumption concerning phrase structure which says that there is a unique structure underlying all German sentences (matrix clause and subordinate clause). This hypothesis is called &quot;Symmetry Hypothesis&quot; or &quot;Doppelkopfanalyse&quot; (ef. Reis 1985). It is shared by most of the generative syntacticiens such as H. den Besten, H. Haider, J. Lenerz and J. Koster. I will adopt some version of this &quot;Symm~try Hypothesis&quot; (SM) which will be developed in the following: 3.2. Phrase structure description (ECS) of German (i) The initial base rule is (15) (15) sbar --> co,up s (ii) There are two left peripheral positioqs compl and comp2. We would like to represent this fact by the following expansion rule'.</Paragraph> <Paragraph position="43"> (16) cutup --> COMP1 COMP2 where COMPI and COMP2 represent positions which will be described thus: (iii) The B .. position has the feature +- tnsd which specifies it as the verb/complementizer position, being filled in the base component only by lexical complementizers. This analysis yields the following structure: (that the man the woman loves) (b) liebt(i) der mann die frau e(i) (loves the man the women) (c} der mann(i) liebt(j) e(i) die frau e(j) (the man loves the woman) (d} die frau(i) liebt(j) der mann e(i) e(9 ) (the woman loves the man) (e) wet(i) liebt(j) e(i) die frau e(j) (who loves the woman) (f) der(i) e(i) die frau liebt (who the woman loves) (g) wet(i) e(i) die frau liebt (who the woman loves) (18)(a) represents the base structure description. (18)(b) V/I representation as in yes/no questions, tile finite verb having moved into COMP2 leaving behind a trace. (18)(c) represents ordinary matrix clause word order derived by the two movement rules TI and T2. (18)(d) represents matrix clause word order with a topicalized direct object. (18)(e) is a case of a matrix clause word order interrogative. (18)(f) a relative clause and (18)(g) a subordinate clause interrogative. (18)(e) and (g) represent a case of wh-movement. Untensed subordinate clauses which would not fit into this schema would be analysed as PPs: (19) pp\[pohne\[vp\[vzufragen\]\]\] (without asking ) This SH-analysis can at least make the following claims: (i) The COMP2-position as complementizer position and as lauding site for the verb-fronting transformation nicely explains the relation between occurrence of complementizer and the occurrence of the finite verb (ii) As (18)(e) and (g) show, wh-movement can be represented equally for matrix clauses and subordinate clauses, namely as movement into COMP2. (iii) The Sll-analysis is compatible with the productive traditional &quot;Stellungsfelderhypothese&quot; (c-f. Olson 1984). Another subject of phrase structure should be mentioned here: the treatment of tile verbal-complex. We adopted the following approach: Every-&quot;~'erb is a full verb. Auxiliaries are subject dass der brief von ihm zu schreiben versucht worden is~ (that tile letter by him to write tried been has) This treatment allows an easy calculation of tense, voice and aspect on ERS, as there is still structural information. As representation (20) shows, all nonfinite verbs are treated on ECS the same way, namely as the head of left recursively branching vp-constituents. This enables an easy treatment of auxiliaries as raising verbs on ERS (see section 3.3.). 3.3. Relational structure (ERS) The relational stgucture of a language is constituted by the property of lexical units (lu) to bind certain other elements. This property is usually called &quot;valency&quot;. Formally this fact is reflected in the formalism by the property of local trees. Each local tree contains just one gov(ernor), its valency-bound elements which are the comp(lements) and its non-valencybound elements which are the mod(ifier)s: mann alt dman old the The valency of a lu is its property to bind a certain number and a certain kind of syntagma. In other words: a valency theory is a theory on how many and which kind of syntagma occur given that a certain lu occurs. We consider verbs, nouns, adjectives and prepositions as having the property of being able to bind other syntagmas. A major part of every valency theory is the design of a test which is meant to determine the difference of complement and modifier. In the history of valency theory a lot of tests have been developed, among others the following: Elimination test, free addability test, adverbial clause test, verb substitution test, do-so-test, backformation test. We adopted a revised addability test (ef. Schmidt 1986).</Paragraph> <Paragraph position="44"> 3.3.2. Word order The most important aspect with the decriptiou of the relation between ECS and ERS is that the present formalism allows for the treatment of free word order languages. We consider German as having a relatively free word order. The decisive feature is that the rhs of the t-rules are able to specify only dominance relations which is expressed by the parenthesis in (22). Permutations in the German middle field can easily be treated as shown in example (22).</Paragraph> <Paragraph position="46"> 3.3.3. The verbal complex on ERS As shown in structure (20), auxiliaries are analysed as full verbs. The structural analysis in (20) makes it easy to treat auxiliaries as raising verbs on ERS, as (23) shows.</Paragraph> <Paragraph position="47"> which is in terms of surface syntactic frames that the phrase being the subject in (a), namely 'die Kommission', is the by obj in (b) and the direct object of (a) is the surface syntaetie subject of (b) (bearing all features surface syntactic subjects usually have, as e.g. nominative case). In terms of thematic roles we could say that the agent is in both eases 'die Kommission' once realized as an NP in nominative case, once realized as a PP with the preposition 'you'.</Paragraph> <Paragraph position="48"> We keep surface syntactic information and aim at the following In universal grammar passive usually is treated in a general way, as passivization is considered a universal process: - In GB passivization is considered as a movement process which is contained in the general move alpha schema. - GPSG also treats passive on the syntactic level in form of a metarule.</Paragraph> <Paragraph position="49"> - LFG being a &quot;lexicalist&quot; theory treats passive inthe lexicon by a lexical rule which is s.th. like (subj) -> zero/(by obj) (obj) -> subj In Eurotra we have neither of these devices, neither movement rules nor metarules, nor lexical rules. However, it seems as if we could simulate the lexical rules just by putting the &quot;active frame&quot; into the b-rule as in (26).</Paragraph> <Paragraph position="51"> This has the same effect as the LFG lexical rule: only one encoding of the verb with its active sub obj-frame is necessary.</Paragraph> <Paragraph position="53"> tat schisdet what) (say hans claim peter adopt council In (29) we can see that a chain was created from the topic of the matrix clause via the topic of the embedded clause to the correct syntactic function slot. We have to guarantee that it is a correct chain which I understand as a chain that is correctly coindexed with the correct function in the ERS b-rule.</Paragraph> <Paragraph position="54"> 4.1.2. The Creation of the Correct Structure The structure in (29) is created by inserting empty elements by t-rule application in a very controlled way. 1 would like to give an exemplification by NP-complements. Structure insertion by t-rules exploit the fact that movement has its landing site which is the node called eompl in representation (17). In the lhs of the t-rule this information is exploited. We also know that each phrase whk:h occupies the eompl position on ECS has to go to an ERS slot which has sf=topic. We need the four t-rules for doing the job.</Paragraph> <Paragraph position="55"> of the b-rule slots subj, obj or obj2 in (10). It will go to sf=subj,sf=obj and sf=obj2. It is up to completeness and coherence to determine that (31) is well-formed in our case. For the top of an unbounded dependency construction (29), we need t-rule (32) which puts the topicalized np into the topic slot on ERS, but without creating a corresponding empty up.</Paragraph> <Paragraph position="56"> (33) treats the middle of unbounded dependency constructions i.e. a sentence structure which has an empty topic. The middle builds the link between embedded sentences and matrix clause. It has no empty correspondent in the structure. This structure is created by a t-rule which operates on an ECS representation which has an empty topic landing site (see (28)).</Paragraph> <Paragraph position="57"> For the bottom of the structural representation we finally need a t-rule which creates an empty topic and an empty corresponding np. (34) is this rule. It is also applied only under the condition that the ECS landing site for wh-movement is empty.</Paragraph> <Paragraph position="58"> We now have all the pieces needed for creating the correct structures which can occur in unbounded dependency structures. (28) only represents a three-fold s-structure, however rule (33) eaters for all possible middles as it will be applied as many times as there are middles.</Paragraph> <Paragraph position="59"> A few comments seem to be in order on these rules: (30) and (32) on the one hand and (33) and (34) on the other hand have the same lhs which might cause overgeneration.</Paragraph> <Paragraph position="60"> Rule (31) caters for the case that the s is tile matrix-clause containing a moved NP which has to find its functional slot downwards somewhere in a functional structure of an embedded clause. For this case we need a topic which has up correspondent on the same level.</Paragraph> <Paragraph position="61"> If we take (27)(b), rule (30) as well as rule (32) will be applied, both of them putting &quot;was&quot; into the topic function, (30) creating an empty NP-slot, (32) not creating an empty NP. So, we have two rules (30) and (32) which apply to the same lhs producing two different ERS structures. The completeness and coherence test determines which t-rule (30) or (32) creates the correct structure. Both of them will be applied but only one, namely (30) creates the correct structure according to the completeness and coherence criterion. In the case of rule (33) and (34) we have the same problem. Both of them apply to the same lhs, once inserting an empty np, once not. Again, completeness and coherence has to determine whether the result of (33) or (34) is correct.</Paragraph> <Paragraph position="62"> The creation of the correct structure is only half of the story. We have not guaranteed yet that only correct structures are created and above all that only correct chains are created. This will be done by an interaction of f-rules percolating the relevant features such as gender, number, case and the index feature and by killer filters which guarantee that only correctly indexed chains survive. First of all we need f-rules which percolate the relevant features.</Paragraph> <Paragraph position="63"> from topic to scomp. Another f-rule of the same style will percolate these features from scomp to the topic node of the embedded sentence, and a third f-rule from topic to the empty functional slot. So, if we consider example (28) the pereolatiou of the relevant features follows the following path: The same kind of f-rule will percolate the case feature independently the same path. (For the reasons see below). For feature ehecking we need killer rules which kill all structures which are not correctly indexed and those which represent an empty chain. E.g. (37) is a rule which deletes all structures where the case feature of the empty topic and the corresponding empty up is not the same.</Paragraph> <Paragraph position="64"> {cat=np,type=empty,case~=C,lndex=l), *ll, {sf=topic,cat=np,case=C,index=l}l.</Paragraph> <Paragraph position="65"> Actually we need another 6 killers which cheek number and gender.</Paragraph> <Paragraph position="66"> Rule (37) makes clear what has been the sense of the separate case-feature-percolation. If we percolated the ease feature in rules like (35) we could not use the index - feature for feature checking. 1 would like to explain this with an example. We need a rule to filter out the wrong representation (39) which is the representation of the following ill-formed sentence: gov gov subj obj topi (i) sa- Hans be- Peter verab- e(i) beschluss e(i) e(i) begem hauptst schiedet den schluss den (say Hans claim Peter adopt the decision the deci- sion) According to our f-rules the index is percolated down into the empty subject slot in the lowest scomp. (It cannot go elsewhere). This subject has case=nom which is stated in the ERS b-rule. The case feature is the means to get rid of the wrong chain as there will be a clash between the &quot;arriving&quot; case=accusative and the already stated case=nominative. If the case feature had not been percolated independently we would not have any possibility of applying killer rule (37) as the f-rule would not have been applied for the reason of the impossibility of unification. My rules percolate the index into the sub j-slot and make possible the application of (37)..</Paragraph> <Section position="1" start_page="592" end_page="594" type="sub_section"> <SectionTitle> 4.2. Control </SectionTitle> <Paragraph position="0"> Let us consider the following case of subject control: (40) dass er den Beschluss zu verabschieden zu versprechen versucht that he tries to promise to adopt the decision dass er den beschluss zu verab- zu vet- rerschieden sprechen sucht The ERS representation would look like (42). In the case of control-structures it is easy to control the insertion of structure by t-rule as embedded control structures are vps in our system. As we have seen in section 3, each vp is lacking a subject np which is inserted on ERS by t-rule (43):</Paragraph> <Paragraph position="2"> In control structures feature checking works the same way as in wh-constructions. We only need a correct feature percolation which puts the relevant features to the scomp-node and from there to the sub j-slot. We only have to take care that in the scomp-node features are not confused with topic-features. This can be guaranteed by using ctl case etc. in scomp.</Paragraph> <Paragraph position="3"> (44):f: f ctll = {cat=s}\[{sf=gov,cat=v,ctl=subj}, {sf=subj,cat=np,nb=N,geud=G,index=I}, &quot;{}, {sf=scomp,cat=s,ctl_nb=N,ctl_gend=G,ctl index=I}, *1}1.</Paragraph> </Section> </Section> class="xml-element"></Paper>