A Syntactic Description of German in a Formalism Designed for Machine Translation

Transcription

1 A Syntactic Decription of German in a Formalim Deigned for Machine Tranlation Paul Schmldt A-Eurotra-D Martln-Luther-Str. 14 D-6600 Saarbrlickcn Wet-Germany Abtract: Thi paper preent a yntactic decription of a fragment of German that ha been worked out within the machine tranlation project Eurotra. t repreent the yntactic part of the German module of thi multilingual tranlation ytem. The linguitic tool for the following analye i the o-called CATframework. n the firt two ection of thi paper an introduction of the formalim and a linguitic characterization of tile framework i given. The CAT formalim a a whole i a theory of machine tranlation, the yntactic analyi part which i the ubject of thi paper i an LFG-like mapping of a contituent tructure onto a functional tructure. A third ection develop principle for a phrae tructure and a functional tructure for German and the mapping of phrae tructure onto functional tructure. n a fourth ection a treatment of unbounded movement phenomena i ketched. A the CAT-framework doe not provide any global mechanim try to give a local treatment of thi problem. O. ntroduction There are two baic given for Eurotra: (i) Stratificational decription of language. The decription of language conit of an analyi on three level: ECS (Eurotra-Contituent-Structure) which decribe language according to part/whole relation and word order, ERS (Eurotra-Relational-Structure) which decribe language in term of yntactic function and S (nterface Structure) which decribe language according to deep yntactic relation enriched by emantic information uch a emantic feature for characterizing lexical unit. (ii) The CAT-formalim. The CAT-formalim i the linguitic tool for the decription of language. A thi formalim ha no global mechanim there are ome retriction concerning the treatment of unbounded dependencie. Taking thc.~;e given into account, would like to preent the following topic: (i) An introduction to the formal language a far a neceary for the treatment of the linguitic phenomena 1 would like to decribe, (ii) A characterization of the Eurotra tratifieational decription of language a a functionally oriented theory, (iii) A development of principle for a yntactic decription of German (iv) A ketch of the treatment of unbounded dependencie 1. The formalim would like to introduce only thoe part of the CAT-formalim "which build the bai of my analye. That i two kind of rule: (i) oealled b-rule (tructure building rule). They build tructure qud tranform tructure into tructure. (ii) o-called feature-rule and killer-filter. They are put together into one cla a both of them operate on tructure created by b-rule expreing generalization over attribute b-rule (1)(a) {cat=} {cat=},(cat=} 1. (b) {eat=}l (cat=v},{cat=} l. (c) {cat=}[ {cat=det},{cat=.} ]. (d) (cat=v,lu=kaufen,lex=kau ft,t.=tened} n (l)(a)-(d) we have b-rule, which define a mall ECSgrammar. (d) i a rule for a terminal. The dominance relation i expreed by quare bracket. The grammar in (1) aign entence (2)(a) tructure (2)(b). (2)(a) Da Hau kauft der Mann (The houe, the man buy) (b) det n v det n da hau kauft der mann The ame way a in (l)(a)-(d) an ECS-grammar wa written we can write b-rule defining functional tructure. (3) i a b-rule defining the functional tructure for (2)(a): (3)(a) {cat=} [{f=,cat=v,frame=ubj obj}, (f=ubj,cat=,cae=nom}, (f=obj,cat=,cae=acc), * (f=mod}l. (f=yntactic function) (b) {l.=kanfen,f=,frame=u bj obj }. b-rule (3) create the functional tructure (4) for entence (2)(a). (4) _, f,v ubj, obj, L L L kaufen mann der hau da The tranformation will De done by tile tranlation b-rule (5). (5) tl = S:(cat=} NPl:{cat=}, ~:{cat=} V:{cat=v},NP2:{cat=}ll S:{cat=} <V,NP2,NP>. A tranlation-b-rule (t-rule) conit of a left hand ide (h) which define a repreentation, in our cae it would unify with the ECS-trueture in (2)(b), and a right hand ide (rh) which define a dominance and precedence relationhip between the item repreented by the variable (capital). f there i a b-rule on the next level, in our cae ERS, which atifie thee condition, the tranlation ucceed, t-rule (5) ay that tructure (2)(b) hall be tranlated into a tructure which i dominated by a node of category which dominate the three item repreented by the variable in the order given in the rh of the t-rule. A the verbal ernor, in our cae 'kaufen', require a ubj obj frame, expreed by the frame feature, (3)(a) i tile ERS-b-rule which would match with the rh of t-rule (5) and create (4). 589

2 1.2. f-rule and killer-filter f-rule f-rule and killer filter allow for the definition of a context part (thoe feature after the lah) and an action part a example (6) how. An f-rule applie to a repreentation only if the context part trictly unifie with the object. (6) { cae=c,nb=n,gead=g,/cat=} {/cat=det},{cae=c,nb=n,gend=g,/eat=n} (6) ay that for each coniting of a det and an n cae, number and gender of the n have to be percolated into the mother node. would like to make two remark: (i) the feature percolation in example (6) could be done in b-rule. Thu, it might eem that f-rule are uperfluou. However, a ection 4 will how, there are many cae where we need feature percolation by f-rule. (ii) will make a pecial ue of f-rule. will take everything a context and action part. That mean, if f-rule f unifie with repreentation r, r will be replaced by the reult of unification, if not, r urvive unchanged killer-filter Killer filter pecify tructure which are not well-formed and which therefore have to be deleted. We might imagine a rule which kill having a pronominal head and an in genitive. (7) killer-filter: {cat=rip} [{cat=detj,{cat=n,n type=pron},{cat=,cae=gen }. 2. CAT a a functionally oriented framework 2.1. A arion with a configurational framework For the linguitic ctmracterization of the Eurotra framework 1 would like to make a brief arion between two kind of linguitic theorie: (i) thoe which aume yntactic function a univeral primitive of language (prototypical: LFG) (ii) thoe which claim that yntactic function could be reduced to configurational fact (prototypicah GB). Each of the two poible way of decribing language force the linguit to decribe linguitic fact a word order, binding relation, agreement, cae aignment or long ditance movement in a certain way. The configurational framework claim that there i a general chema for phrae tructure rule which i the univeral pattern according to which all contituent tructure of all poible language are built. t i the x-bar chema: undergone movement tranformation, a functional decription conit in a mapping of phrae tructure onto functional tructure Completene and coherence in Eurotra There i an eential which hold for all functional framework, namely the letene and coherence principle. Thi principle ay: A functional tructure i well-formed iff it i lete and coherent. A functional tructure i lete iff it contain all the yntactic function required by the frame of the framebearing element. A functional tructure i coherent iff it contain only the required yntactic function. Enrotra allow for the expreion of thi principle in two different way: (i) Enumeration of frame The ERS grammar ha to enumerate all poible pattern, all frame which are poible, a b-rule, and the value of the frame feature of the determine that only the wanted and nothing but the wanted ernor go into the tructure building rule. (9) {cat=} [ {f=,cat=v,frame=ubj obj}, {f=nhj,cat=,cae=nom}, {f=obj,cat=,cae=acc}, *{f=mod} ] n (9) letene i expreed by the fact that both framebound yntactic function are obligatory. So, if one of the function i miing, the tructure i not well-formed. Coherence i expreed by the fact that the tructure building rule only allow for the two yntactic function and nothing ele. Thi prevent e.g. the creation of an oblique object. (ii) Completene and coherence by f-rule and killer There i, however another way of expreing letene and coherence which doe not require the enumeration of all frame. We need the following rule: (a) One ERS b-rule which enumerate all poible yntactic function optionally a (0) doe: (10) :b: {cat=} [ {f=,cat=v}, ^ {f=nbj,cnt=,cae=nom}, ^ {f=obj,cat=,cae=acc}, ^ {f=obj2,cat=}, ^ {f=obl,cat=pp}, ^ {f=,cat=}, * {f=mod,cat=pp} ^ {f=topic} ] (b) A eparate encoding of the function a verb i ubcategorized for, i.e. the frame feature i given up and a feature for each yntactic function i introduced: () pec Xdoublebar xbar xbar beta gamma (11) {lu=ee,ubj=ye,obj=ye,f=} All other yntactic function feature value will have to get the default "no" (by default f-rule). (12) {lu=ee,ubj=ye,obj=ye,obj2=no,obl=no,=no,f=} (8) repreent the x-bar chema, alo D(eep)-Structure in GB. On thi tructure movement rule operate creating S(urface)- tructure. So, thi i the kind of explanation a configurational framework give: There i a canonical chema (the x-bar chema) and each configuration not fitting into thi chema i explained a derived by univerally retricted movement tranformation. The functional alternative ha to rely on yntactic function a univeral primitive. So, phrae tructure doe not necearily claim a univeral tatu, and movement rule are not even neceary. Thi require a different treatment of the linguitic phenomena. How doe the CAT framework fit into thi? The adoption of the three level ytem (ECS,ERS,S) make Eurotra functionally oriented a it adopt the way of linguitic decription a functional approach ha to adopt. While a configurational deoription conit in mapping given configuration onto a canonical chema, the x-bar chema, by explaining configuration which do not fit into x-bar a having We can now tate letene and coherence independently by killer filter: (13) :k: {cat=} l {f=,cat=v,ubj--ye}, ^ {f=obj,cat=,cae=acc}, ^ {f=obj2,cat=}, ^ {f=obl,cat=pp}, ^ {f=,cat=}, * {f=mod,cat=pp} ^ {f=topic) (13) determine that if the feature for ubj=ye then there mut be a yntactle function "ub j" in thi repreentation. Expreed by a killer it read: if there i a tructure whoe ha an a-air ubj=ye and contain only function which are not ubj then thi tructure i not well-formed. The ame which ha been tated here for ubj can be tated for all function. To get coherence we ue a killer filter a in (14). 590

3 (14) :k: {cat=) [ {f=,cat=v,ubj=no}, {f=ubj,cat=,cae=nom}, " 0 (14) ay: f the tructure whoe ha the feature-value 'no' for the feature 'ubj' contain a feature bundle containing the feature f=ubj, plu anything ele, then thi tructure i not well-formed. 3. Syntactic decription of German 3.1. Principle of yntactic decription A we have een above, the yntactic decription of a language in Eurotra follow a functional approach. n our decription thi i not only reflected by the exitence of a functional level but alo by the uonhierarehical, nonconfigurational decription of the entence contituent we offer. A we do not ue the given x- bar chema we need no empty element on ECS and we decribe German a a uonconfigurational language. Though in German matrix claue we have SVO word order, German i uually conidered an SOV language. Matrix claue word order i conidered a derived from ubordinate claue word order by movement tranformation. (of. Koter 1975, Thierch 197~ L Rei On thi bai we would like to make another aumption concerning phrae tructure which ay that there i a unique tructure underlying all German entence (matrix claue and ubordinate claue). Thi hypothei i called "Symmetry Hypothei" or "Doppelkopfanalye" (ef. Rei 1985). t i hared by mot of the generative yntacticien uch a H. den Beten, H. Haider, J. Lenerz and J. Koter. will adopt ome verion of thi "Symm~try Hypothei" (SM) which will be developed in the following: 3.2. Phrae tructure decription (ECS) of German (i) The initial bae rule i (15) (15) bar --> co,up (ii) There are two left peripheral poitioq l and 2. We would like to repreent thi fact by the following expanion rule'. (16) cutup --> COMP1 COMP2 where COMP and COMP2 repreent poition which will be decribed thu: (iii) The B.. poition ha the feature +- tnd which pecifie it a the verb/lementizer poition, being filled in the bae onent only by lexical lementizer. Thi analyi yield the following tructure: (17) bar COMP COMP2 vfin (B) bar ] [ vfin (a) da der mann die frau liebt (that the man the woman love) (b) liebt(i) der mann die frau e(i) (love the man the women) (c} der mann(i) liebt(j) e(i) die frau e(j) (the man love the woman) (d} die frau(i) liebt(j) der mann e(i) e(9 ) (the woman love the man) (e) wet(i) liebt(j) e(i) die frau e(j) (who love the woman) (f) der(i) e(i) die frau liebt (who the woman love) (g) wet(i) e(i) die frau liebt (who the woman love) (18)(a) repreent the bae tructure decription. (18)(b) V/ repreentation a in ye/no quetion, tile finite verb having moved into COMP2 leaving behind a trace. (18)(c) repreent ordinary matrix claue word order derived by the two movement rule T and T2. (18)(d) repreent matrix claue word order with a topicalized direct object. (18)(e) i a cae of a matrix claue word order interrogative. (18)(f) a relative claue and (18)(g) a ubordinate claue interrogative. (18)(e) and (g) repreent a cae of wh-movement. Untened ubordinate claue which would not fit into thi chema would be analyed a PP: (19) pp[pohne[[vzufragen]]] (without aking ) Thi SH-analyi can at leat make the following claim: (i) The COMP2-poition a lementizer poition and a lauding ite for the verb-fronting tranformation nicely explain the relation between occurrence of lementizer and the occurrence of the finite verb (ii) A (18)(e) and (g) how, wh-movement can be repreented equally for matrix claue and ubordinate claue, namely a movement into COMP2. (iii) The Sll-- analyi i atible with the productive traditional "Stellungfelderhypothee" (c-f. Olon 1984). Another ubject of phrae tructure hould be mentioned here: the treatment of tile verbal-lex. We adopted the following approach: Every-"~'erb i a full verb. Auxiliarie are ubject control verw (of Netter1986, 1988, and Brenan 1982). ~20) bar... vfin vinfj.n vinfin vinfin da der brief von ihm zu chreiben verucht worden i~ (that tile letter by him to write tried been ha) (iv) Two movement rule operate on thi tructure, deriving all non SOV tructure. Thee two rule are: T : Verb fronting and T2 : Topicalization where COMP2 i the landing ite for the finite verb and COMP1 thb landing ite for X-double-bar. We will how now in (18) how poible German entence tructure can be derived according to SM. Thi treatment allow an eay calculation of tene, voice and apect on ERS, a there i till tructural information. A repreentation (20) how, all nonfinite verb are treated on ECS the ame way, namely a the head of left recurively branching -contituent. Thi enable an eay treatment of auxiliarie a raiing verb on ERS (ee ection 3.3.) Relational tructure (ERS) Principle The relational tgucture of a language i contituted by the property of lexical unit (lu) to bind certain other element. Thi property i uually called "valency". Formally thi fact i reflected in the formalim by the property of local tree. Each local tree contain jut one (ernor), it valency-bound element which are the (lement) and it non-valencybound element which are the mod(ifier): 591

4 (21) ubj,,n mod,ap mod,detp,adj,det mann alt d- man old the The valency of a lu i it property to bind a certain number and a certain kind of yntagma. n other word: a valency theory i a theory on how many and which kind of yntagma occur given that a certain lu occur. We conider verb, noun, adjective and prepoition a having the property of being able to bind other yntagma. A major part of every valency theory i the deign of a tet which i meant to determine the difference of lement and modifier. n the hitory of valency theory a lot of tet have been developed, among other the following: Elimination tet, free addability tet, adverbial claue tet, verb ubtitution tet, do-o-tet, backformation tet. We adopted a revied addability tet (ef. Schmidt 1986) Word order The mot important apect with the decriptiou of the relation between ECS and ERS i that the preent formalim allow for the treatment of free word order language. We conider German a having a relatively free word order. The deciive feature i that the rh of the t-rule are able to pecify only dominance relation which i expreed by the parenthei in (22). Permutation in the German middle field can eaily be treated a hown in example (22). (22) S:{eat=har} -:{eat=} TOPC,V:{cat=v}, ~:{cat=}[advl:* {cat--adv2}, NPh(cat=}, ADV2:* {cat=adv2}, NP2:{cat=}, ADV3:* (cat=adv2}, NP3:{cat=}, ADV4:* {cat=adv2}, VP:{cat=}l] S:<(TOPC,V,ADV1,NP1,ADV2,NP2,ADV3,NP3,ADV4,VP)> The verbal lex on ERS A hown in tructure (20), auxiliarie are analyed a full verb. The tructural analyi in (20) make it eay to treat auxiliarie a raiing verb on ERS, a (23) how. (23) 7.-- ~ov OOV,V ubj, l up(i) ubj,, (i).l gi v'v ~ev,v ubj,, da ein brief werder den n~(i>~ov <~j, by_phi,up ver- a chrei- e uehen ben yon ihm Paive The problem with paive i the following: There i a relation between the two entence in (24) (24) Die Kommiion verabchiedet den Beehlu (The Commiion adopt the deciion) Der Beehlu wurde yon der Kommiion verabchiedet (The deciion wa adopted by the council) which i in term of urface yntactic frame that the phrae being the ubject in (a), namely 'die Kommiion', i the by obj in (b) and the direct object of (a) i the urface yntaetie ubject of (b) (bearing all feature urface yntactic ubject uually have, a e.g. nominative cae). n term of thematic role we could ay that the agent i in both eae 'die Kommiion' once realized a an NP in nominative cae, once realized a a PP with the prepoition 'you'. We keep urface yntactic information and aim at the following tructure: (25) undef ubj ~v ubj.by_?bj warden bechlu verabchieden e von kommiion n univeral grammar paive uually i treated in a general way, a paivization i conidered a univeral proce: - n GB paivization i conidered a a movement proce which i contained in the general move alpha chema. - GPSG alo treat paive on the yntactic level in form of a metarule. - LFG being a "lexicalit" theory treat paive inthe lexicon by a lexical rule which i.th. like (ubj) -> zero/(by obj) (obj) -> ubj n Eurotra we have neither of thee device, neither movement rule nor metarule, nor lexical rule. However, it eem a if we could imulate the lexical rule jut by putting the "active frame" into the b-rule a in (26). (26) {cat=,voice=palve} l {f=,cat=v,frame=ubj_obj}, (f=ubj,eat=,cae=nom}, {f=by ohj,eat=pp,pform=von), '1} {lufverabchleden,f=,cat= v,framefu bj obj } Thi ha the ame effect a the LFG lexical rule: only one encoding of the verb with it active ub obj-frame i neceary. 4. Treatment of Unbounded Movement Phenomena 4.1. Wh-movement The Repreentation would like to explain my approach with an example: (27)(a) wa agt Han, behauptet Peter, verabchiedet der Rat (what ay Han claim Peter adopt the council) what doe Han ay that Peter claim the council adopt (b) wa agt Han According to our ECS grammar the following ECS tree i created: (28) bar l topic v u l-- i bar v bar _ v wa agt Han behaup- Peter verab- der rat tet chiedet (what ay Han claim Peter adopt the council) We imagine a functional repreentation like (29), (29) under ( ~ ~ gov gov ubj JSCOmp t--~(i) %... i agen han behaup- Peter verab- Rat e(i) e(i) e(i) wa tat chidet what) (ay han claim peter adopt council n (29) we can ee that a chain wa created from the topic of the 592

5 matrix claue via the topic of the embedded claue to the correct yntactic function lot. We have to guarantee that it i a correct chain which undertand a a chain that i correctly coindexed with the correct function in the ERS b-rule The Creation of the Correct Structure The tructure in (29) i created by inerting empty element by t-rule application in a very controlled way. 1 would like to give an exemplification by NP-lement. Structure inertion by t-rule exploit the fact that movement ha it landing ite which i the node called eompl in repreentation (17). n the lh of the t-rule thi information i exploited. We alo know that each phrae whk:h occupie the eompl poition on ECS ha to go to an ERS lot which ha f=topic. We need the four t-rule for doing the job. (30) tl= S:{cat=har) ~:(cat=,tn= tened} [TOPC:{cat=},V:(cat=v} h ~: {cat=,tn=untened} [NP2;{cat=),~:^{cat=pUnCrt},SBAR:^ {cat=har)l] =). S:{cat=} <V,NP2,{cat=,n type=empty},sbar,topc:{f=toplc} >. The t-rule in (30) treat" local wh-movenrent a in (2)(a) and create tructure (31 ). ( 31 ) unde f ubj obj topic ( i ) kaufen mann e(i) hau (30) create an empty -lot which ha to be interpreted a one of the b-rule lot ubj, obj or obj2 in (10). t will go to f=ubj,f=obj and f=obj2. t i up to letene and coherence to determine that (31) i well-formed in our cae. For the top of an unbounded dependency contruction (29), we need t-rule (32) which put the topicalized into the topic lot on ERS, but without creating a correponding empty up. (32) t2= $:{cat=bar} [~:{cat=,tn=tencd} [TOPC: {cat=up}, V: {cat=v}l, ~:{cat=,tn=untened} [NP2: ^{cat=up}, ~: ^{cat=punct}, SBAR: ^(cat=bar}ll S:{cat=} < V, NP2, SBAR, TOPC:{f=toplc,cat=} >. (33) treat the middle of unbounded dependency contruction i.e. a entence tructure which ha an empty topic. The middle build the link between embedded entence and matrix claue. t ha no empty correpondent in the tructure. Thi tructure i created by a t-rule which operate on an ECS repreentation which ha an empty topic landing ite (ee (28)). (33) i3= S:{cat=bar} ~: {cat=,tn=tened} lv:{cat=v}l, ~: {cat=,tn=untened} [Np2: ^{cat=up}, -: ^{cat=puuct}, SBAR: ^{cat=:bar}l] S:{cat=} < V,NP2,SBAR,{cat=,n type=empty,f=topic} >. For the bottom of the tructural repreentation we finally need a t-rule which create an empty topic and an empty correponding. (34) i thi rule. t i alo applied only under the condition that the ECS landing ite for wh-movement i empty. (34) t4= S:(cat=bar) ~:{cat=,tn=tened) V:{cat=v}l, ~:{cat=,tn=untened) [ NP2: ^{cat=,tp), -: ^{cat=punct}, SBAR: ^{cat=har}l] S:{cat=} <V,NP2,{cat=,n type=empty), SBAR,(cat=,n type=empty,f=topic } >. We now have all the piece needed for creating the correct tructure which can occur in unbounded dependency tructure. (28) only repreent a three-fold -tructure, however rule (33) eater for all poible middle a it will be applied a many time a there are middle. A few comment eem to be in order on thee rule: (30) and (32) on the one hand and (33) and (34) on the other hand have the ame lh which might caue overgeneration. Rule (31) cater for the cae that the i tile matrix-claue containing a moved NP which ha to find it functional lot downward omewhere in a functional tructure of an embedded claue. For thi cae we need a topic which ha up correpondent on the ame level. f we take (27)(b), rule (30) a well a rule (32) will be applied, both of them putting "wa" into the topic function, (30) creating an empty NP-lot, (32) not creating an empty NP. So, we have two rule (30) and (32) which apply to the ame lh producing two different ERS tructure. The letene and coherence tet determine which t-rule (30) or (32) create the correct tructure. Both of them will be applied but only one, namely (30) create the correct tructure according to the letene and coherence criterion. n the cae of rule (33) and (34) we have the ame problem. Both of them apply to the ame lh, once inerting an empty, once not. Again, letene and coherence ha to determine whether the reult of (33) or (34) i correct Feature Checking The creation of the correct tructure i only half of the tory. We have not guaranteed yet that only correct tructure are created and above all that only correct chain are created. Thi will be done by an interaction of f-rule percolating the relevant feature uch a gender, number, cae and the index feature and by killer filter which guarantee that only correctly indexed chain urvive. Firt of all we need f-rule which percolate the relevant feature. (35) :f: a_top to = {cat=z} [ {f=}, ^{f=ubj}, ^{f=obj}, ^{f=obj2}, {f=,top index=l,top.b=n,top _gend=g}, *{f=mod}, {f=topic,iudex=l,.b=n,gend=g}. (35) i an example which percolate number, gender and index from topic to. Another f-rule of the ame tyle will percolate thee feature from to the topic node of the embedded entence, and a third f-rule from topic to the empty functional lot. So, if we conider example (28) the pereolatiou of the relevant feature follow the following path: (36) obj e topic e topic topic The ame kind of f-rule will percolate the cae feature independently the ame path. (For the reaon ee below). For feature ehecking we need killer rule which kill all tructure which are not correctly indexed and thoe which repreent an empty chain. E.g. (37) i a rule which delete all tructure where the cae feature of the empty topic and the correponding empty up i not the ame. 59)

6 (37) :k: ktopic3= {cat=} [{f=,cat=v}, "0, {cat=,type=empty,cae~=c,lndex=l), *ll, {f=topic,cat=,cae=c,index=l}l. Actually we need another 6 killer which cheek number and gender. Rule (37) make clear what ha been the ene of the eparate cae-feature-percolation. f we percolated the eae feature in rule like (35) we could not ue the index - feature for feature checking. 1 would like to explain thi with an example. We need a rule to filter out the wrong repreentation (39) which i the repreentation of the following ill-formed entence: (38) * Den Bechlu agt Han, behauptet Peter, verabchiedet den Bechlu (The deciion ay Han, that Peter claim, adopt the deciion) (39) under ub3 topic(i) topic(i) ubj ubj obj topi (i) a- Han be- Peter verab- e(i) bechlu e(i) e(i) begem hauptt chiedet den chlu den (ay Han claim Peter adopt the deciion the deciion) According to our f-rule the index i percolated down into the empty ubject lot in the lowet. (t cannot go elewhere). Thi ubject ha cae=nom which i tated in the ERS b-rule. The cae feature i the mean to get rid of the wrong chain a there will be a clah between the "arriving" cae=accuative and the already tated cae=nominative. f the cae feature had not been percolated independently we would not have any poibility of applying killer rule (37) a the f-rule would not have been applied for the reaon of the impoibility of unification. My rule percolate the index into the ub j-lot and make poible the application of (37) Control Let u conider the following cae of ubject control: (40) da er den Bechlu zu verabchieden zu verprechen verucht that he trie to promie to adopt the deciion Our ECS-grammar would aign the ECS-trueture (41): (41) bar ubcon mp n prep v prep da er den bechlu zu verab- zu vet- rerchieden prechen ucht The ERS repreentation would look like (42). n the cae of control-tructure it i eay to control the inertion of tructure by t-rule a embedded control tructure are in our ytem. A we have een in ection 3, each i lacking a ubject which i inerted on ERS by t-rule (43): (42) under da veruchen ubj(i) er vetprechen ubj o(i) ubj obj verab- e(±) bechieden chlu (43) tl = VP:{cat=} [ NPl:{cat=},VP:(cat=},~:{cat=prep}, V:{cat=v,tn=untened}] VP:{cat=} V,{cat=,type=empty,f=ubj},NPl,VP >. n control tructure feature checking work the ame way a in wh-contruction. We only need a correct feature percolation which put the relevant feature to the -node and from there to the ub j-lot. We only have to take care that in the -node feature are not confued with topic-feature. Thi can be guaranteed by uing ctl cae etc. in. (44):f: f ctll = {cat=}[{f=,cat=v,ctl=ubj}, {f=ubj,cat=,nb=n,geud=g,index=}, "{}, {f=,cat=,ctl_nb=n,ctl_gend=g,ctl index=}, *1}1. 5, Summary The decription of a ignificant fragment of German above eem to be a good bai for a tranlation ytem. The functional tructure created in our ytem can eaily be mapped onto deep yntactic predicate-argument-tructure which are enriched by emantic information. From there tranfer hould happen. A far a the treatment of unbounded dependencie i concerned there might be ome problem in tranfer. Certain pied piping phenomena and multiple wh-movement might make neceary a more powerful mechanim. 6. Literature: Abraham,W.(ed)(1985) Erkl~irende Syntax de Deutchen, Tiibingen, (=Studien zur deutehen Grammatik 25). Arnold,D. et a1.(1987) The Eurotra Reference Manual, Releae 2.1., m. Utrecht. Brenan,J.(1982) The Paive in Lexical Thoery, in: Brenan,J The Mental Repreentation of Grammatical Relation Cambridge, Ma./London Engl. Koter,J.(1975) Dutch a an SOV Language. Linguitic Analyi l,pp.l Lenerz,J.(1984) Diachronic Syntax: Verb Poition and COMP in German, in: Toman (1984). Netter,K.(]986) Getting Thing out of Order. An LFG Propoal for the Treatment of German Word Order, Coling Proceeding (1986),p Olon,S.(1984) On Deriving V-1 and V-2 Structure in German, in: Toman (1984). Rei,M.(1985) Satzeinleitende Strukturen. Ueber COMP, Hauptund Nebenaetze, w- Bewegung und Doppelkopfanalye, in: Abraham (1985). Steiner,E., Sehmidt, P,, Zelinky, C. (1988) (forthcoming) from Syntax to Semantic. (New night from Machine Tranlation). London Schmidt,P.(1986) Valency Theory in a Stratificational MT Sytem,in: Coling Proceeding (1986). mod den Thierch, C.: Topic in German Syntax, uub. Di