TRACE & UNIFICATION GRAMMAR 
liana Ulrich Block and Stefanie Schachtl 
Siemens AG, Corporate Research, ZFE ST SN 74 
Otto Ilalm-Ring 6 
D-8000 Mih*chen 83 
Germany 
block@ztiv~x.zfe.siemens.de 
Abstract 
This paper presents Trace & Unification Grammar 
(TUG), a declarative and reversible grammar formal- 
ism that brings together Unification Grammar (UG) 
and ideas of Government & Binding Theory (on)~ 
The main part of the paper consists in a description 
of many free word order phenomena of German syn- 
tax. It will be shown that the use of traces allows 
for an easier and more elegant way of description 
than competing approaches like ID/LP-format rules 
as used e.g. in GPSG and HPSG. Grammkrs writ- 
ten in the TUG-formallsm can be compiled to a very 
efficient parser. The occurrence of head movement, 
wh-movement and scrambling in one sentence does 
not lead to any decrease in parsing efficiency. 
1 Introduction 
The basic motivation in creating TUG formalism was 
to capture the empiric knowledge that represents the 
outcome of at least 15 years' linguistic discussion of 
German word order, while maintaining the efficiency 
that is required from a grammar formalism of today's 
standard. With reference to the basic work of \[Len77\] 
on marked and unmarked word order linguists in the 
generative tradition such as \[Thi82\], \[dB84\], \[Fau87\] 
and \[Cze87\] revealed an impreesive list of descrip- 
tive phenomena that can be appropriately handled by 
the assumption of s conflgurational, i.e. VP contain- 
ing description of German word order. Among these 
phenomena count asymetries in the serialization be- 
haviour of nominatively marked NPs in passive and 
the so called "psych" constructio*m*, the account for 
scrambling phenomena in Acl/ECM Constructions 2 
and the observations of \[Cze87\] wrt. to the voices of 
double accusative verbs in German. What all these 
works intend is to pronounce the danger a description 
of German is likely to run iuto, that "draws immedi- 
ate conclusions as to the surface position of argumet, t 
NPH on the basis of their surface cases" (\[dB84\]:59). 
t \[L~n77\] 0howl that v~rlm like w~ndcrn, ge.falle., #din#ca, 
etc. with non &gentive subjects pat*era with the paaalvized 
formu of "agentive" verb~ in prefering the unmarked word order 
indlr¢ct object - subject. \[dB84\] extends thit ob~scrvation on 
copular corm*ruction with NP governin~ a4jectivee 
~t~e also \[dB84\] for a thor* outline of tltis problematic ittstte 
2 The TUG formalism 
The basis of TUG is formed by a context free gram- 
mar that is augmented by PATK ll-styie feature equa- 
tions. Besides this basis, the main features of TUG 
are feature typing, mixing of attribute-value-pair and 
(PROLOG-) ternr unilication, flexible macros, unre~ 
stricted disjunction attd special rule type~ for argu- 
ment and head movement, 
2.1 Basic rule types 
As a very simple example we will look at the TUG 
version of the example grammar in {Shi84\] z. 
9', type daflnition 
s => ~. 
np => f (agr :agrmnt). 
vp => f (agr : ag'~mt) . 
v => f (agr : agrmnt). 
agrmut => f(numbar:number,perton:ptrton). 
number => {alnguler,plural). 
permon => {first,necond,thlrd}. 
% rules 
• ---> up, vp I 
ap:agr - vp:agr. 
vp -~-> v, np I 
vp:agr - v:agr, 
The two main differences to PATR II in the basic 
framwork are that first, TUG i~ less flexible in that it 
has a "hard" contextfree backbone, whereas in PArR 
lI categories of the context frcc part are placeholders 
for feature structures, their names beeing taken as 
the value of the cat feature in the structure. Second, 
TUG has a strict typing. For a feature path to be 
well defined, each of its attributes has to be declared 
in tile type definition. 
3 For a Jlightly mor~ dctMled d~*eription of the bmdc fea- 
tures of TU(I refer to (\[Blo91\]). 
ACq'ES DE COLING-92, NAmXS. 23-28 Ao{rr 1992 8 7 I'R(XL OV COLING-92, NA~s, Auo. 23-28, 1992 
2.2 Movement rules 
Further to these more standard UG-features, TUG 
provides special rule formats for the description of dis- 
continuous dependencies, so called "movement rules". 
Two main types of movement are distinguished: ar- 
gument movement and head movement. The format 
and processing of argument movement rules is greatly 
inspired by \[CLW88\] and \[Che90\], the processing of 
head movement is based on GPSG like slash features. 
2.2.1 Head Movement 
A head movement rule defines a relation between two 
positions in a parse tree, one is the landing site, the 
other the trace position. Itead movement is con- 
strained by the condition that the trace is the head of 
a specified sister (the root node) of the landing site 4. 
Trace and antecedent are identical with the excep- 
tion that the landing site contains overt material, the 
trace does'nt. 
To formulate head movement in TUG the following 
format is used. First, a head definition defines which 
category is the head of which other, e.g. for the V- 
projection line of the above grammar: 
v is_head_of vp. 
vp is_head of s. 
Second, the landing site is defined by a rule like 
s' ---> v+s I ... 
where landing site and root node are linked by a +. 
To inclnde recursive rules in the head path, heads are 
defined by the following head definitions. In a struc- 
ture \[~\[ D1 ... D,\] Di is the head ell(if either Di 
is_head_of M is defined or Di has the same category 
as M and eitt*er D~ is_head_of X or X is_head_of Di 
is defined for any category X. 
2.2.2 Argument Movement 
Argument movenmnt rules describe a relation be- 
tween a landing site and a trace. The trace is always 
e-commanded by the landing site, its antecedent. 
Two different traces are distinguished, anaphoric 
traces and variable traces. Anaphoric traces must 
find their antecedent within the same bounding node, 
variable trace binding is constrained by sut~jaceney, 
c.a. the binding of tim trace to its antecedent must 
not cross two bounding nodes. Anaphoric traces are 
found for example in English passive constructions 
\[s \[np Tim book of this author\]i was read t~\] 
wbereas variable traces are usually found in wh- 
constructions and topicalization. Similar to the pro- 
posal in \[CLW88\], argument movement is coded in 
TUG by a rule that describes tim landing site, as for 
example in 
4}Iere, "head of" is a transitive relation ~.t. if x is head of 
y and y is head of z then x is head of z. 
s2 ---> np:ante<trace(var,np:trace), sl I 
ante:fx = traee:fx, 
This rule states that rip:ante 5 is tile antecedent of 
an np-trace that is dominated by sl. 
The first argument in the trace-term indicates 
whether the landing site is for a variable (vat) or 
for an anaphoric (aaa) trace. Other than head move- 
ment, where trace and antecedent are by definition 
identical, the feature sharing of argument traces with 
their antecedents has to be defined in the grammar by 
feature equations (ante : fx = trace : fx, . . .). Fur- 
thermore, it is not necessary that the antecedent and 
the trace have the same syntactic category. 
The current version of the formalisms requires that 
the grammar contains a declaration on which cate- 
gories are possible traces. In such a declaration it 
is possible to assign features to a trace, for example 
marking it as empty: 
trace(np) I rip:empty = yee. 
Bounding nodes have to be declared as such in the 
grammar by statements of the form 
bounding_node (rip). 
bounding_node(s) ~ s:tense = yes. 
As in the second case, bounding nodes may be de- 
fined in terms of category symbols and features. 
The main difference of argument movement to 
other approaches for the description of discontinuities 
like extraposition grammars (\[Per81\]) is that argu- 
ment movement is not restricted to nested rule appli- 
cation. This makes the approach especially atractive 
for a scrambling analysis of the relative free word or- 
der in the German Mi~telfeld as explained in more 
detail below. 
3 Some facts on German syn- 
tax 
3.1 Basic assumptions 
In the following we will sketch tile basic structures 
of German syntax. According to the position of the 
finite verb, we distinguish sentences with the verb in 
the second (a), the first(b) and the last position (c). 6 
(1) a. Karl fdhrl nach Hamburg. 
Karl goes to Hamburg 
b. Fiihrt Karl nach tlamburg? 
Goes Karl to Hamburg 
c .... daft Karl nach llamburg f~ihrt. 
... that Karl to Hamburg goes 
5The notation Ca*~ : Index is used to distinguish two or more 
occurrences of the same category in the same rule in tile equa- 
tion part. : antQ arid : t race are arbitrary names used as index 
to refer to the two different nps. 
eEnglish literal translations are given in typewriter font. 
AcqT/s Dh" C()LING-92, NArqrFS, 23-28 AO~Zf 1992 8 8 PRec. el: COLING-92, NANTES, AUG. 23-28, 1992 
We call the sentence types V2-S (a), V1-S (b) and 
Ve-S (c) respectively. In V1-S and V2-S, compound 
verbforms build a so called Satzklammer between the 
finite and the nonfinite parts. 
(2) Karl ist ,nit dam tug nach Hamburg 
Karl has with the train to Haatburg 
gefahren. 
golle 
"Karl has gone to Hamburg by train." 
Traditionally, according to the position of the ver- 
bal elements, we call the position in front of the finite 
verb the Vorfeld and the positions within the Satzk. 
laminar the Millelfeld, 
In accordance with the mentioned eonfigurational 
analysis of the german sentence, we suppose an un- 
marked "underlying" structure, that is similar to the 
order in tim german subordinate clause. This struc- 
ture contains four different positions for verb argu- 
ments, as exemplified by the following sentences. 
(3) ~'. \[S~ \[X' daft \[s \[NI-" de," Man,\] 
\[vv \[N~ der .~n\] \[v, \[NP de, Bnch\] 
\[v. geg.b~,* hat\] \] 11 \] \] 
that the man the woman the book given 
hae 
"that the man lies given the woman the book" 
b. \[sz \[s* daft \[s \[NP der Mann\] 
rye \[v, \[,-' d., ~.~h\] \[v~ \[~ t,, d~n 
Seh,~.k\] re. g~leyt hat\] l\]l \] \] \] 
that the man the book into the 
bookshelf put has 
"that the man has put the book into the book- 
shelf' 
A sentence always contains the v-projection line 
VK, V t, VP, S, even if the sentence contains less than 
three arguments. 
(4) a. \[s= \[s' daft \[s \[~'v der Man,\] 
\[vv re, \[vK t.n*t\]\]\]\]\]\] 
that the ~an\[nom\] dance9 
b. \[s s \[s'* daft \[s \[v." \[Nv dam Mann\] 
re* \[VK gehol\[e, wild\]Jill\] 
that the aan\[dat\] helped is 
"that the nian is helped" 
c. \[s~ b' daft \[s \[vp Iv' \['~v das ~ud,\] 
iv. gel~,~n ~i~d\]\]\]\]\]\] 
that the book\[send read is 
"that the book is read" 
d. \[s= \[s, daf\[s \[vv \[v, \[vie geta,,zt wird\]\]\]\]\]\] 
that danced is 
"that there is dancing" 
As is shown in (4a.) vs. (4c.) the nominative 
may be assigned to a VP-external or a VP-internal 
position. Adverbials are ehomksy-adjoind to S, VP, 
V t and VK. An adjunction is only possible, if the 
right daughter is binarily brandfing. 
(~) a. \[s~ \[s, daf \[s h~,,te (s \[m. d~," Mann\] 
\[w' re' \[w¢ ta,m\]\]\]\]\]\] 
M.n,,\] \[v' \[v. get, orion wi,q\]\]\]\]\]\] 
c. Cs~ Is, da~ \[s \[vP \[~, h~ut~ iv, \[~ des 
Un~a\] Iv. ~l ..... i,~I\]\]11\] 
d. \[s~ \[s' daft \[s \[vP \[v' \[vK heute 
\[v J; get.,,~t ~i,4\]\]\]\]\] 
For a word order that differs from the underlying 
structure the movement rules of TUG are used. V1-S 
are formed by head-movement of the finite verb to 
the position of the complementizer. 
(6) a, \[ss \[s, daft \[s Peter \[VP \[V 1 die 13ilder 
\[w .b,,,.u\]\]\]\]\]\] 
that Peter the pictures copies 
"that Peter copies the pictures" 
b. \[s. \[s, ,,,,,tt~ \[s P~te; \[vp \[v, di~ ~il&r 
\[v. a6 t,\]\]\]\]\]l 
copies Peter the pictures 
"Does Peter copy the pictures" 
This can be formulated ia a rather compact way in 
TUG by definition of the head relation and rules for 
the introduction of the landing site of the linite verb. 
(7) v is~ead_of vk. 
vk is~head_of v:t. 
vi isA~ead_of vp. 
vp isJxead_of s. 
(8) sl ---> comp, s 
sl ---> v+s 
V2-S are formed by occupying the Vorfeld, i.e. the 
position immediately dominated by S 2 with either a 
verb argument by argument movement (a), by an ad- 
verbial (b) or by a Vorfeld-es (c). 
(9) a. \[S'~ Pete,', \[S' malt, \[s ti \[Vl' \[V' die Bilder 
\[w, ab tAll\]l\] 
Peter copies the pictures 
b \[: Heet~ \[ ....... ~t, b' ~ete; re,, re, die 
~ila~ \[v. ab t,\]\]\]\]\]\] 
Today copies Pete~: the pictures 
c. \[s~ Es \[s' ,,,alt~ \[s ,°ete,'\[v,~ \[v' die raider 
Iv.,," ,,b td\]\]\]\]\] 
It copies Peter the picture~ 
"Peter copies tire pictures" 
This facts can be described by tim following rules: 
(10) s2 ---> es, sl 
s2 ---> pp, sl 
s2 ---> advp, st 
s2 ---> rip<trace(vat,rip), sl 
trace(vat,rip). 
Free word order in £1xe Miltclfchl is described by 
"moving" an argument to a chomsky-adjoined posi- 
tion on the V-projection. llere it obeys the same 
conditions a.s an adverbial and leaves a trace in the 
original argument position. 
Acr\[~s DF, COLING 92, NANrJiS, 23-28 ao(rr 1992 8 9 PRec. OF CO\[,ING-92, NAN'H~S, AUd. 23-28, 1992 
(11) a. \[sa Is, daft Is der M .... \[vP \[N~" da,g 
Buch\]i \[vn. der bYan \[v, tl \[v~ gegebe, 
aatllllll\] 
that the tan the book the soaan 
given has 
b. is, \[s, d~ \[s (uP d~r -~"b \[s \[uP da, 
~atl\]\]\]\]\]\]\] that the uoaan\[dat\] the book\[anal 
the aan\[noa\] given has 
~that the man has given the woman the 
book" 
So, for scrambling, we basically need the following 
rules: 
a ---> np<trace(a~a,np), s 
vp ---> up<trace(ann,up)0 vp 
3.2 Alternative approaches 
Whereas meet concurrent theories adopt the view 
that an argument phrase in the Vorfeld is linked to 
the argument position by a trace - be it by movement 
or by the slash-feature - the relative free word order in 
the Miffelfeld is often accounted for by the distinction 
of phrase structure rules into immediate dominance 
(ID) rules and linear precedence (LP) rules. ID rules 
define the hierarchical structure of constituents, LP 
rules the linear ordering of daughters constituents. 
In this paradigm the german Miftelfeld inelu'ding the 
finite verb typically is supposed to form a fiat struc- 
ture, generated by an ID rule like a ---> np\[no~, 
np\[akk\], np\[dat\], v\[~ln\], vk\[infin\] r. The ele- 
ments on the right hand side can then be (partially) 
ordered by LP statements of the form vf:fi~ < up0 
np< vk (a finite verb precedes an NP, a VK follows 
an NP). As no LP statement is made for the NPs, the 
rules generate all possible permutations of NPs. 
(13) Is hat dot Mann der Frau d~ Buch gegeben\] 
\[8 hat dex Frau der Mann d~ Bach gegeben\] 
Is hat das Buch der Mann der Fran gegeben\] 
Is hat der Fran du Buch der Mann gegeben\] 
\[s hat der Mann du Buch der Fran gegeben\] 
So, where TUG supposes a fixed unmarked word 
order, from which marked orders are derived by move- 
ment rules (scrambling), GPSG and ItPSG suppose 
unordered ID rules and express constraints on order 
explicitly by LP Statements. The same holds for the 
position of the finite verb in the different german sen- 
tence types. As for movement to the Vorfeld the 
GPSG/HPSG approach using the slash feature and 
the TUG approach are rather similar, as in fact move- 
ment is implemented in TUG by structure sharing. 
3.3 Some more facts on German syn- 
tax 
In the following paragraph we will outline a few de- 
scriptive phenomena where we think the use of traces 
?But ace \[Reag9\] for an alternative approach using LP- 
itatementa that do¢4 not have to Msume a fiat etructure. 
as in TUG allows for more elegant formulations of the 
facts. 
3.3,1 Preposition stranding 
In many German dialects, prepositional proforrrm like 
damd or dsoegsn and question forms like vaornit or 
wofefen can be used discontinuously. Both the slash 
analysis and the movement to the Vorfeld are able 
to describe appropriately sentences like (14b.), where 
the pronoun part of a prepositional preform is de- 
tached from its preposition. 
(14) a. Damit kann er diese Theorie wider- 
legen. 
Therewith can he thin theory 
de~ea~ 
b. Da kann er dices Theorie mit wide~ 
legen. 
There can he this theory with 
defeat 
"With th~ he ean de~at the theory" 
Whereas the preceding data give strong evidence 
for movement into the Vet.reid, preposition stranding 
in German is not restricted to that position. The 
da of a discontinuous preform can also occur in the 
Miftelfeld. 
(15) a. Er kann diese Theorie damit wider- 
legen. 
He can thin theory therewith 
defeat. 
b. Er kann da dieseTheoriemit wider 
legen. 
He can there this theory with 
defeat. 
Furthermore, it kazan be combined in one sentence 
with another discontinuity, e. g. discontinuous was 
//r. 
(16) a. Was kann er da fiir sine Theorie mit 
widerlegen. 
What can he there for a theory 
with defeat 
"What a theory can he defeat with 
this" 
b. Was~ kann erda/ t/ fdr eine Theorie 
t/mit widerlegen. 
Evidently, a formalism who~ slash feature allows 
only for one discontinuous constituent has to describe 
da ... mit by LP rules. Therefore, da and mit would 
have to be of the same constituent as diese Theorie. 
But in 14 (a) damit obviously forms one constituent. 
It is unclear how in a flat structure the obligatoriness 
of da can be expressed, an effect that naturally falls 
out of the trace approach. 
AcrEs DE COLING-92. NANTES. 23-28 AOUT 1992 9 0 PROC. ON COLING-92. NANTES. AUo. 23-28. 1992 
3.3.2 The position of pronouns 
The word order of personal pronouns in the German 
sentence is rather restricted as exemplified by the fol- 
lowing sentences. 
(17) a. dai\] der Chef ihn ihr vorstellt 
that ths boss him\[ace\] hsr\[dat\] 
presents 
"that the boss presents her to him" 
b. daft der Chefihr den neuen Mitarbeiter 
vorstellt 
that the boss hsr\[dat\] the nsu 
collsgue presants 
"that the boss presents the new col- 
legue to her" 
c. *dab der Chef den neuen Mitarbeiter 
ihr vorstellt 
*that the boss the new 
colle~le\[ace\] her\[dat\] presents 
d. *daft der Chef ihr ihn vorstellt 
*that the boss hor\[dat\] him\[ace\] 
presents 
e. daft ihn ihr der Chef vorstellt 
that him\[ace\] her\[dat\] the boss 
presents 
£ *7daft ihnder Chef ihr vorstellt 
*that his\[ace\] the boss her\[dat\] 
presents 
g. *daft ihn ihr er vorstellt 
*that hi=\[ace\] her\[dat\] he\[sol\] 
presants 
Even though it might be possible to describe these 
restrictions by a set of LP statements, our impression 
is that this kind of analysis obscurs the rather simple 
pronoun word order. Furthermore, we cannot see how 
LP rules could allow for (a) and (e) while excluding 
(f)a. Ihn may preceed the nominativ NP, but only if 
there is no dativ pronoun following the latter, s 
In our analysis, personal pronouns have a fixed po- 
sition in the sentence either between S 1 and S or be- 
tween S und VP. 
(is) a. \[s a \[st daft esl ihrj \[s tier Mann \[vl, tj 
\[v, t, aibt\]\]\]\]\] 
b. \[s a \[st daft \[s tier Mann esl ihr i \[vP tj 
\[Va t, gibt\]\]ll \] 
3.4 ACI-constructions 
In ACI-coustructious however a personal pronoun 
may very well follow a non-nominativ NP. Compare 
(17e.) and (19). 
s (f) seems to be somehow acceptable in some dialects but 
completely agrammatical in others. 
°As far as we understand it, also a solution by sequence 
union \[Rea89\] could not account for these facts. 
(19) a. Gestern hat Karl den Jungen ihr helfen 
la88en. 
¥sstsrday has Karl the boy 
her\[dat\] help let 
"Yesterday Karl has made the boy help 
her" 
b. Gestern hat Karl den Jungen \[ve ihr 
helfen\] lassen. 
To save an LP analysis as indicated above we would 
have to say that ihr helfen is part of another con- 
stituent as den Jungen, therefore the LP statements 
do not hold between den Jangen and ihr. The struc- 
ture of the sentence (19a) might then be sketched as 
in (19b). 
But now consider the following sentence: 
(20) Gestern hat sieh Karl eine L~sung sin- 
fallen l~sen. 
Yesterday has Karl himself a 
solution come_to~ind let 
This sentence, although it repreeents a regular ACI- 
construction and consequently must have an embed- 
ded VP constituent, shows the same serialization wrt. 
the pronoun as (17), where all NPs belonged to the 
same constituent. ID and LP rules therefore lead to a 
contradiction in handling (19) and (20) t°. A TUG de- 
scription on the other hand can make use of a scram- 
bling analysis in (20). The pronoun sich leaves a trace 
in the pronoun position of the embedded construc- 
tion and appears in the appropriate pronoun position 
of the matrix constituent, thereby maintaining the 
relevant serialization conditions in both constituents. 
And a further restriction may be implemented when 
configurational relations are available: the scrambling 
of a dative pronoun out of an ACI-coustruction as 
shown in (20) is sensitive as to whether the accusative 
of the ACI-verb replaces an external or internal nom- 
inative of the embedded verb. Finite cinfallen marks 
its nominative internally, finite half ca marks it ex- 
ternally. Only internal nominative-accusatives allow 
being scrambled over, cf. (20) with (21). 
(21) *Gestern hat ihr Karl den Jungen helfen 
lassen. 
*Yesterday has hsr\[dat\] Karl the 
boy help let 
3.4.1 Scrambling and Thematic Structure 
Another aspect of the configurational differentiation 
between external and internal arguments can be made 
use of in analysing the thematic structure of a sen- 
tence. So e.g. the thematic differentiation between 
wide and narrow scope of a verbal argument depends 
on its appearing in marked or unmarked position. 
Compare the readings of (22) and (23) vs. (26), where 
boldface marks the posodically prominent syllable: 
10A description like the one in \[Rea89\] would presumably 
have to state that verbs like h¢lfen are combined in ACI- 
constructions by concatenation~ verbs like tin\]allen by se- 
quence union. 
A~q'ES DE COL1NG-92. NA~'ES, 23-28 AoIYr 1992 9 1 PRec. OF COLING-92, NANTES, AUG. 23-28, 1992 
(22) \[ch glaube daft der Kollege dem Vorstand 
widersprochen hat. 
I think that the collegue has the 
board contradicted 
"I think that the collegue has contradicted 
to the board." 
(23) Ich glaube dab dem Vorstand der Kollege 
widersprochen hat. 
I think that the board the 
collegue contradicted has 
"1 think that the colluege has contradicted 
to the board." 
(22) is a possible answer to the question what hap- 
pened on the meelmg yesterday whereas (23) is only 
allowed in specific contexts such as Wet hat dem 
Vorstand widersprochen?. Widersprcchen like helfea 
marks its nominative externally, therefore the dative 
NP follows the nominative NP in the umnarked word 
order as shown in (22). If scrambling applies as in 
(23) this correlates with a change in the thematic and 
prosodic structure. In the TUG framework this would 
be achieved in combining the scrambling mechanism 
with a feature structure that indicates the desired 
thematic interpretation of the sentence: 
(24) vp ---> np<tracs(ana,np:trace), 
vp:h 
h : S C ope=narro~;, 
(25) vp ---> np, v~. 
np: scope=vp : scope. 
A straightforward implementation of this observa- 
tion in the ID/LP format however would come to a 
halt in the case of (26): 
(26) Ich glaube dab dem Vorstand die LSsung 
eingefallen ist. 
I think that the board the 
solution come_to_mind has 
"I think that the solution has come to 
mind to the board." 
Although the nominative follows the dative in (26) 
the sentence has wide scope interpretation and un- 
marked prosodic structure 11 , (26) again is a likely an- 
swer to the general question Was ist gestern passiert? 
This clearly contradicts (22)-(23) if only the surface 
case marking of the arguments can be referred to in 
the generalization. On the other hand this result is 
exactly what would be expected by a TUG analy- 
sis: Since einfallen marks its nominative internally, 
no scrambling is involved in (26) vs. (23). 
11 John Pheby in \[HFM81\] poatulates the distinction between 
maxked and unmarked prosodic structure in Gennaaa. \[vSU86\] 
combine this with a configurational syntax. See also \[Uhm91\] 
for a reformulation of the relevant obeservations in the frame- 
work of \[Pie80\]. 
4 Parsing with TUG 
TUG can be processed by a parser and a generator. 
Before parsing and generation, the grammar is coat- 
piled to a more efficient form. The first compilation 
step that is common to generation and parsing trans- 
forms the attribute-value-pair structure to (PRO- 
LOG) term structure. This transformation makes use 
of the type definitions. For parsing, TUG is processed 
by a Tomita parser \[Tom86\]. For usage in that parser 
the result of the transformation to PROLOG further 
undergoes several transformations (expansion of head 
nlovement rules, transformation of argument move- 
ment rules, elimination of empty productions, conver- 
sion to LR.(K) format and computation of LR tables). 
This compilation leads to a considerable increase in 
processing speed that makes it possible to use TUG 
for the syntactic description in real-timesystems. Es- 
pecially the seperatc compilation of head movement 
aml argument movement leads to run time grammars 
that do not show the usual decrease in efficicncy due 
to empty productions (traces). In fact, a compiled 
TUG does not contain empty productions any longer. 
Parsing time for simple sentences of about 10 words 
using a grammar of German with rather broad cov- 
erage is between 1 and 2 sees. on the average on a 
SUN SPARC I workstation running Quintus Prolog, 
even if the sentence contains verb fronting, argument 
movement to the Vorfeld and scrambling in tile Mit- 
telfeld, 
5 Conclusion 
We have presented Trace & Unification Grammar, a 
grammar formalism that tries to bridge the gap be- 
tween UG and GB theory with the aim of adopting 
many of the linguistic descriptions of German found 
in the linguistics literature. Besides German, the pre- 
sented grammar formalism has also been used suc- 
cessfully to describe a smaller subset of Chinese. We 
have compared TUG descriptions of some phenomena 
in german syntax to approaches that do not make use 
of movement rules but use \[D/LP rules instead and 
shown that in all these cases TUG provides a sim- 
ple and elegant description whereas the ID/LP ap- 
proach in most cases even fails to describe the data at 
all. Furthermore we have briefly mentioned tile way 
TUG can be compiled to an efficient parser. Em- 
pirical tests have shown that using movement rules 
and traces does not lead to a considerable decrease in 
parsing speed. 
Acknowledgements 
We would like to thank Ping Peng, Manfred Gehrke, 
l~udi Iiunze and Ludwig Schmid for many discus- 
sions Oil tile TUG-formalism. The work described 
here is partly sponsored by the Federal Ministry of 
Research and Technology in tile ASL project under 
no. 01IV102AO. 
AcrEs DE COLING-92, NANTES, 23-28 ho(rr 1992 9 2 PROC. OV COLING-92, NAh°i'I..'s, AUO. 23-28, 1992 
9 3 PRO(:. OV COLIN(;~92. NANIES, At;G. 23-28, 1992 

References 

\[Blo91\] Hans Ulrich Block, Compiling Trace 
Unification Grammar for Parsing and Gen- 
eration. In Proceedings of the I~versible 
Grammar Workshop, ACL, Berkeley, 1991. 

\[Cheg0\] It.-11. Chen. A logic-based government- 
binding parser for mandarin chinese. In 
I3th International Conference on Compu- 
tational Linguistics (COLING.90), l)age~ 
1 6, 1990. 

\[CLW88\] H.-tI. Chen, l.-P. Lin, and C.-P. Wu. A 
new design of prolog-based bottom-up pars- 
ing system with government-binding the- 
ory. In 12th International Conference on 
Computational Linguistics (COLING-88), 
pages 112-116, 1988. 

\[Cze87\] tlartnmt Czephtch. Lexikalische Argu- 
mentstruktur und syntaktisehe Projektio. 
neu: zur Beschreibung grammatiseher \[{e- 
lationen. ZS, 6(1):3~a6, 1987. 

\[dB84\] Hans den Besten. The ergative hypothesis 
and free word order in Dutch and German. 
In Jindrieh Toman, editor, Studies in Ger- 
man Grammar, pages 23 65. 1984. 

\[Fan87\] Gisbert l"anselow. Konfiguratioaalitiit. 
Narr, Tiibingen, 1987. 

\[IIFM81\] Karl E. Heidolph, Walter Flgmig, and 
Wolfgang Motsch. Grundziige einer 
deutscheu Grammatik. Akademie Verlag, 
Berlin, 1981. 

fLea77\] Jiirgen Lenerz. Zur Abfolge nomiaaler 
Satzglieder im Deutsehen. Narr,'l?flbingen, 
1977. 

\[Per81\] Fernando Pereira. Extraposition gram- 
mar. Computational Linguistics, 7:243- 
256, 1981. 

\[Pie80\] Janet Pierrehumbert. The Phonology and 
Phonetics of English Intonation. PhD the- 
sis, M.1,T., 1980. 

\[Rea89\] Mike Reape. A logical treatment of semi- 
free word order and bounded discontinu- 
ous constituency. In Proceedings of the 4th 
Conference of the European Chapter of the 
ACL (EACL-89), pages 1(13-110, 1989. 

\[Shi84\] Stuart M. Shieber. The Design of a Com- 
puter Language for Linguistic info}mation. 
In lOlh International Conference on Com- 
putational Linguistics (COLING-84), pages 
362-366, 1984. 

\[Thi82\] Craig Thierseh. A note on scrambling and 
the existence of vp. Wiener Lin!luistische 
Gazette, (27-28):83 95, 1982. 

\[Tom86\] M. Tomita. ElJieient Parsing for Natu- 
ral Lauguage: A fast Al.qorithm for l'ract~- 
cal Systems. Kluwer Academic Publishers, 
Boston, 1986. 
Arn'lis I)E COLING-92, Na~"n!s, 23 28 horn' 1992 

\[Uhm91\] Susanne Uhnmnn. bbkusphonologie. 
Niemeyer, Tiibingen, 1991. 

\[vSU86\] Arnim yon Steehow and Susanne Uhnmnn. 
Some remarks on focus projection. In 
Werner Abraham and Sjaak de Meij, edi- 
tors, Topic, Focus, and ConJigurationalily, 
pages 295-320. john benjamins, 1986. 
