Semantic-based Transfer* 
Michael Dorna and Martin C. Emele 
Institut fiir Maschinelle Sprachverarbeitung 
Azenbergstralle 12 
D-70174 Stuttgart 
{dorna, emele} @ims. uni-stuttgart, de 
Abstract 
This article presents a new semantic- 
based transfer approach developed and 
applied within the Verbmobil Machine 
Translation project. We give an overview 
of the declarative transfer formalism to- 
gether with its procedural realization. 
Our approach is discussed and compared 
with several other approaches from the 
MT literature. The results presented in 
this article have been implemented and 
integrated into the Verbmobil system. 
1 Introduction 
The work presented in this article was developed 
within the Verbmobil project (Kay et al., 1994; 
Wahlster, 1993). This is one of the largest projects 
dealing with Machine Translation (MT) of spo- 
ken language. Approxinmtely 100 researchers in 
29 public aad industrial institutions are involved. 
The application domain is spontaneous spoken 
language in face-to-face dialogs. The current sce- 
nario is restricted to the task of appointment 
scheduling and the languages involved are English, 
German and Japanese. 
This article describes the realization of a trans- 
fer approach based on the proposals of (Abb and 
Buschbeck-Wolf, 1995; Caspari and Schmid, 1994) 
and (Copestake, 1995). Transfer-based MT 1, see 
e.g. (Vauquois and Boi~et, 1985; Nagao et al., 
1985), is based on con~rastive bilingual corpus 
analyses from which a bilingual lexicon of trans- 
fer equivalences is derived. In contrast to a purely 
*This work was funded by the German Federal 
Ministry of Education, Science, Research and Tech- 
nology (BMBF) in the framework of the Verbmobil 
project under grant 01 IV 101 U. We would like 
to thank our colleagues of the Verbmobil subproject 
Transfer, our IMS colleagues Ulrich Heid and C.J. 
Rupp and our anonymous reviewers for usefltl feed- 
back a~ld discussions on earlier drafts of the paper. 
The responsibility for the contents of this paper lies 
with the authors. 
1For a more detailed overview of different ap- 
proaches to MT, see e.g. (Hutchins and Solners, 1992). 
lexicalist approach which relates bags of lexical 
signs, as in Shake-and-Bake MT (Beaven, 1992; 
Whitelock, 1992), our transfer approach operates 
on the level of semantic representations produced 
by various analysis steps. The output of transfer is 
a semantic representation for the target language 
which is input to the generator and speech synthe- 
sis to produce the target language utterance. Our 
transfer equivalences abstract away from morpho- 
logical and syntactic idiosyncracies of source and 
target languages. The bilingual equivalences are 
described on the basis of semantic representations. 
Since the Verbmobil domain is related to dis- 
course rather than isolated sentences the model 
theoretic semantics is based on Kamp's Discourse 
Representation Theory, DRT (Kamp and Reyle, 
1993). In order to allow for underspecification, 
variants of Underspecified Discourse Representa- 
tion Structures (UDRS) (Reyle, 1993) are em- 
ployed as semantic formalisms in the different 
analysis components (Bos et al., 1996; Egg and 
Lebeth, 1995; Copestake et al., 1995). 
Together with other kinds of information, such 
as tense, aspect, prosody and morpho-syntax, 
the different semantic representations are mapped 
into a single multi-dimensional representation 
called Verbmobil Interface Term (VIT) (Dorna, 
1996). This single information structure serves as 
input to semantic evaluation and transfer. The 
transfer output is also a VIT which is based 
on the semantics of the English grammar (el. 
Copestake et al. (1995)) and used for generation 
(see Kilger and Finkler (1995) for a description of 
the generation component). 
Section 2 of this paper sketches the seman- 
tic representations we have used for transfer. In 
section 3 we introduce transfer rules and dis- 
cuss examples. In section 4 we compare our 
approach with other MT approaches. In sec- 
tion 5 we present a summary of the implemen- 
tation aspects. For a more detailed discussion of 
the implementation of the transfer formalism see 
Dorna and Emele (1996). Finally, section 6 sum- 
marizes the results. 
316 
2 Semantic Representations 
The different Verbmobil semanti(: construction 
components use variants of UDRS as their semas> 
tic formalisms, el. (Bos et al., 1996; Egg and Le- 
beth, 1995; Copestake et al., 1995). The ability 
to underspecify quantifier and operator scope to- 
gether with certain lexical ambiguities is impof 
rant for a practical machine translation system 
like Verbmobil because it supports ambiguity pre- 
serving translations. The (lisambiguation of dif- 
ferent readings couht require an m'bitrau'y amount 
of reasoning on real-world knowledge and thus 
should be avoided whenever possible. 
In the following examples we assume an ex- 
plicit event-based semantics (Dowry, 1989; Par- 
sons, 1991) with a Neo-Davidsonian representa- 
tion of semantic argument relations. All seman- 
tic entities in UDRS are uniquely labeled. A la- 
bel is a pointer to a semantic predicate nmking it 
easy t,o refer to. The labeling of all scntantic enti- 
ties allows a fiat ret)resentation of the hierarchical 
structure of arguinent an(1 ot)erator and quantifier 
scope embeddings as a set of labeled conditions. 
The recursive embedding is expressed via addi- 
tional subordination constraints on labels which 
occur as arguments of su(:h olmrators. 
Example (la) shows one of the classical 
Verbmobil examples and its possible English 
translation (lb). 
(1) a. Das paflt echt scMecht bei 'mir. 
b. That really doesn't suit me well. 
The corresponding semat)l,ic representations are 
given in (2a) and (2b), respectively. 2 
(\[2) a. \[ll:echt(12), 12:schlecht(il), 
13:passen(il), 13:arg3(il,i2), 
14:pron(i2), 15:bei(il,i3), 16:ich(i3)\] 
b. \[ll:real(12), 12:neg(17), 17:good(il), 
13:suit(il), 13:arg3(il,i2), 
14:pron(i2), 15:arg2(il,i3), 16:ego(i3)\] 
Scnlantic entities in (2) are reprc, scnted as a Pro- 
log list of laImlcd conditions. After the unification- 
t)ased s(~illantic construction, the logical wu'iables 
for labels and nmrkers, such as events, states and 
individuals, are skolemize(l with special constant 
symt)ols, e.g. 11 for a label att(t ±1 for a state. Ev- 
ery condition is prefixed with a lal)el serving as a 
unique identifer. Labels are also useflll for group- 
ing sets of conditions, e.g. for i)artitions whidt be- 
long to the restriction of a qmmtifier or which 
are part of a specific sub-DRS. Additionally, all 
these special constants can be seen as pointers for 
adding o1" linking informal, ion within ml(t between 
multiple levels of the VIT. 
Only the set of semantic conditions is shown in 
(2); the ot;her levels of the multi-dimensional VIT 
ret)resentation , whidt contain additional senmntic, 
2For l)resentaJ;ion purposes we have simplilied the 
actmd VIT representations. 
pragmatic, morpho-syntactic mid prosodic infor- 
mation, have been left, out here. If necessary, such 
additional information can be used in transfer and 
sem;mtic evaluation for resolving ambiguities or in 
generation for guiding tile realization choices. Fur- 
thermore, it .allows traasfer to make fine-grained 
distinctions between alternatives in cases where 
the semantic representations of source mid target 
lmlguage do not match up exactly. 
Semantic operators like negation, modals or in- 
tensitier adverbials, such as really, take extra label 
arguments fi)r referring to other elements in the 
flat; list which m'e in the relative scope of these 
operators.a 
This form of semantic representation has the 
following adwmtages for transfer: 
• It; is possible to preserve the underspecifica- 
tion of quantifier amd operator scope if there 
is no divergence regarding scope ambiguity 
between Sollrce and target languages. 
• Coindexation of labels and markers in the 
source and target parts of transfer rules en- 
sures that the semantic entities are correctly 
related and hence ()bey any semantic con- 
straints which may be linked to dram. 
To produce an adequate target utterance 
additional constraints whirl1 arc important 
for generation, e.g. sortal, topic/focus con- 
straints etc., may be preserw;d. 
* There need not be a 1 : 1 relation between 
semantic entities and individual lexical items. 
Instead, lexical units may be decomposed int;o 
a set of seinantic entities, e.g. in the case of 
deriwztions arid fi)r a nmre line grained lexical 
semantics. Lexical decomposition allows us to 
express generalizations attd to apply transDr 
rules to parts of the decomposition. 
3 Our Transfer Approach 
Transfer equivalences are stated as relations be- 
tween sets of source language (SL) and sets of tar- 
get language (TL) senlantie entities. They are usu- 
ally based on individual lexical items but might 
also involve partial phrases for treating idioms and 
other collocations, e.g. verb-noun collocations (see 
example (8) below). After skolemization of the se- 
n, antic representation the input to transfer is vari- 
able free. This allows the use of logical variables 
for labels and markers in transfer rules to express 
coindexation constraints between individual enti- 
ties such as predicates, operators, quantifiers and 
aFor tim concrete exaanple at hand, the relative 
scope ha.s been fully resolved by using the explicit la- 
bds of other conditions. If the scope were underspeci- 
fled, explicit subordination constraints would be used 
in a speciM scope slot of the VIT. The exact details 
of subordination are beyond tim scope of this paper, 
of, Prank aa~d Reylc (1995) and Bos et al. (1996) for 
implementations. 
317 
(abstract) thematic roles. Hence the skolemiza- 
tion prevents unwanted unification of labels and 
markers while matching individual transfer rules 
against the semantic representation. 
The general form of a transfer rule is given by 
SLSem, SLConds Tau0p TLSem, TLConds. 
where SLSem and TLSem are sets of semantic enti- 
ties. Tau0p is an operator indicating the intended 
application direction (one of <->,->, <-). SLeonds 
and TLConds are optional sets of SL and TL con- 
ditions, respectively. All sets are written as Prolog 
lists and optional conditions caa be omitted. 
On the source language, the main difference be- 
tween the SLSem and conditions is that the for- 
mer is matched against the input and replaced by 
the TLSem, whereas conditions act as filters on the 
applicability of individual transfer rules without 
modifying the input representation. Hence condi- 
tions may be viewed as general inferences which 
yield either true or false depending on the context. 
The context might either be the local context as 
defined by the current VIT or the global context 
defined via the domain and dialog model. Those 
inferences might involve arbitrarily complex infer- 
ences like anaphora resolution or the determina- 
tion of the current dialog act. In an interactive 
system one could even imagine that conditions are 
posed as yes/no-questions to tile user to act as a 
negotiator (Kay et al, 1994) for choosing the most 
plausible translation. 
K the translation rules in (3) are applied to the 
semantic input in (2a) they yield the semantic out- 
put in (2b). We restrict the following discussion 
to the direction from German to English but the 
rules can be applied in the other direction as well. 
(3) a. \[L:echt(A)\] <-> \[L:real(h)\]. 
b. \[L: passen (E) ,L: arg3 (E,Y) ,LI : bei (E,X)\] <-> 
\[L:suit (E) ,L:arg2(E,X) ,L:arg3 (E,Y)\]. 
c. \[L:schlecht(E)\], \[Ll:passen(E)\] <-> 
\[L:neg(A),A:good(E)\] . 
d. \[L:ich(X)\] <-> \[L:ego(X)\]. 
e. \[L:pron(X)\] <-> \[L:pron(X)\]. 
The simple lexical transfer rule in (3a) relates the 
German intensifier echt with the English real 4. 
The:variables L and A ensure that the label and 
the argument of the German echt are assigned to 
the English predicate real, respectively. 
The equivalence in (3b) relates the German 
predicate passen with the English predicate suit. 
The rule not only identifies the event marker E, 
but unifies the instances X and Y of the relevant 
thematic roles. Despite the fact that the German 
bei-phrase is analysed as an adjunct, it is treated 
exactly like the argument arg3 which is syntacti- 
cally subcategorized. This rule shows how struc- 
tural divergences can easily be handled within this 
approach. 
4The semantic predicate real abstracts away from 
the adjective/adverbial distinction. 
(4) \[L:passen(E), Ll:bei(E,X)\] <-> 
\[L: suit (E), L: arg2 (E,X)\]. 
The rule in (3b) might be further abbreviated to 
(4) by leaving out the unmodified arg3, because it 
is handled by a single metarule, which passes on 
all semantic entities that are preserved between 
source and target representation. This also makes 
the rule for (3e) superfluous, since it uses an inter- 
lingua predicate for the anaphor in German and 
English. 
The rule in (3c) illustrates how an additional 
condition (ILl :passen(E)\]) might be used to 
trigger a specific translation of schleeht into not 
good in the context ofpassen. The standard trails- 
lation of schlecht to bad is blocked for verbs 
like suit, that presuppose a positive attitude 
adverbial. 5 One main advantage of having such 
conditions is the preservation of the modularity 
of transfer equivalences because we do not have to 
specify the translation of the particular verb which 
only triggers the specific translation of the adver- 
bial. Consequently, the transfer units remain small 
and independent of other elements, thus the in- 
terdependencies between different rules are vastly 
reduced. The handling of such rule interactions is 
known to be one of the major problems in scaling 
up MT systems. 
A variation on example (1) is given in (5). 
(5) a. Das paflt mir echt schleeht. 
b. That really doesn't suit me well. 
Tile translation is exactly the same, but tile Ger- 
man verb passen takes an indirect object mir in- 
stead of the adjunct be/-phrase in (1). The appro- 
priate transfer rule looks like (6a) which can be 
reduced to (6b) because no argument switching 
takes place and we can use the metarule again. 
(6) a.\[L:passen(g) ,L:arg2(E,X) ,L:arg3(E,g)\]<-> 
\[L: suit (E), L:arg2(E,X) ,L:arg3(E,Y)\] . 
b. \[L:passen(E)\] <-> \[L:suit(E)\]. 
In a purely monotonic system without overriding 
it would be possible to apply the transfer rule in 
(6b) to sentence (1) in addition to the rule in (4) 
leading to a wrong translation. Whereas in the 
underlying rule application scheme assumed here, 
the more general rule in (6b) will be blocked by 
the more specific rule in (4). 
The specificity ordering of transfer rules is 
primarily defined in terms of the cardinality of 
matching subsets and by the subsumption order 
on terms. In addition, it also depends on the 
cardinality and complexity of conditions. For the 
passen example at hand, the number of match- 
ing predicates in the two competing transfer rules 
defines the degree of specificity. 
~Instead of using a specific lexical item like passen 
the rule should be abstracted for a whole class of verbs 
with similar properties by using a type definition, e.g. 
type (de, pos_att itude_verbs, \[gehen, passen .... \] ). 
For a description of type definitions see (11) below. 
318 
The following example illustrates how condi- 
tions are used to enforce selectional restrictions 
from the domain model. For example ~rmin in 
German might either be translal;ed as appointment 
or as date, depending on the context. 
(7) a. \[L:termin(X)\] <-> It:appointment(X)\]. 
b. \[L : terrain (X) \], 
\[sort(X)=<'temp_point\] <-> \[L:date(X)\]. 
The second rule (7b) is more specific, because it 
uses an additional condition. This rule will be 
tried first by calling the external domain model 
R)r testing whe.ther the sort assigned to X is not 
suhsumed by the sort, letup_point. Here, the first 
rule (7a) serves as a kind of default with respect to 
the translation of Terrain, in cases where no spe- 
cific sort information on the marker X is awfilable 
or the condition in rule (7b) Nils. 
In (8), a light verb construction like einen Ter- 
minvorsehlag aachen is translated into su.qgest a 
date by decomposing the compound and light verb 
to a simplex verb and its modifying noun. 
(8) \[L : machen (E) , L : arg3 (g, X) , 
LI : terminvorschlag (X) \] <-> 
\[L : sugge st (E) , L : arg3 (E, X), LI : date (X) \] . 
We close this section with a support verb example 
(9) showing the treatment of head switching in our 
approa.ch. The German comparat;ive construct,ion 
lieber sei'n (lit.:bc more liked) in (9a) is t;ranslated 
by the verb prefer in (9t)). 
(9) a. Diensta9 ist air lieber. 
h. \[ would l, refer Tuesday. 
(IO) \[L : suppore (S, LI ), L2 : experioncer (S, X) 
LI :lieb(Y) ,LI: comparative(Y)\] <-> 
\[L:prefer (S) ,L: argl (S,X) ,L: arg3 (S,Y)\] . 
The tra.nsfer rule in (10) matches the decoinposi- 
lion of the comI)at'at;ivt; form lieber into its posi- 
tive forin lieb atnt an additional comt)arative pred- 
icate toget;her with l;he. support verb sei'n such t;IKtl; 
tile comparative construction lieber sein (g ist X 
liebeT) is translated as a whoh; to the English verb 
prefer (x prefers Y). 
.4 Discussion 
The main motivation for using a senmntic-based 
at)proach for transfer is the abilil;y to abstract 
aww froln morplioh)gical and syntactic idiosyn- 
crasies of individual languages. Many of the, tra- 
ditional cases of divergences discussed, e.g. by 
Dorr (1994), at'e already handled in the Verbmobil 
syntax-seniantics interface, hence they (lo not 
show up in our transfer at)proach. Examples in- 
clude cases of categorial and thematic divergences. 
These are treatt;d in tile linking between syntac- 
tic arguments and their corresponding thematic 
roles. 
Another advantage of a semantic-based t;rans- 
fer approach over a pure interlingua apt)roach, 
e.g. Dorr (1993), or a direct sl;ructural c()rrespon- 
dence approach, e.g. Slocum el; al. (1987), is the 
gain in modularity by allowing language indepen- 
dent grammar development. Translation equiva- 
lences relating semantic entities of the source and 
target grammars can be fi)mmlated in a grmnmar 
independent bilingual semantic lexicon. In cases 
where the semantic representations of source, and 
target language are not isomorphic, a nontrivial 
transfer relation between the two representations 
is needed. But it is cleat'ly much easier to niap be- 
tween fiat semantic representations than between 
either syntact;ic trees or deeply nested semantic 
re, presentations 
An inl;erlingua approadt presumes thai; a sire 
gle representation for arbitrat'y languages exists 
or can be developed. We believe fi-om a grammar 
engineering point of view it is unrealistic to come 
tip with such aai interlingua representation with- 
out a strict coordination between the monolingual 
grammars. In general, a pure interlingua approach 
results in very application and domain specific 
knowledge sources which at'e difficult to maintain 
atM extend to new languages anti domains. This 
holds especially in the Verbmobil context with its 
distributed gratnmat- development. 
Whereas our approach does not preclude the use 
of interlingua predicates. We use interlingua rep- 
resentations for time and date expressions in the 
Verbmobil domain. Sinfilarly for prepositions, cf. 
Buschbeck-Wolf and Niibel (1995), it makes sense 
1,o use inore abstract relations which express tim- 
dmnental relationships like temporal location or 
spatial location. Then it is left to the language spe- 
cific grammars to make the right lexical choices. 
(11) a. type(de ,leap lea, \[an, in,um,zu\] ). 
b. am Dienstag, im Mai, um drei, zu Ostern 
c. type(en,temp loc,\[on,in,at\]). 
(t. on Tuesday, in May, at three, at E~Lster 
'File class deiinitions in (lla) arid (llc) cluster 
together those prepositions which can be used t,o 
express a temporal location. The names de and en 
are the SL and TL modules in which the (:lass is 
deiined, temp loc is the (:lass natne and the list 
denotes the extension of the class. (11b) and (11d) 
show possible German and English lexicalizations. 
(12) \[temp_loc(E,X)\] , \[sort(X)=<time\] <-> 
\[temp loc(E,X)\]. 
The interlingua rule in (12) identifies the abstract 
teinl)oral location predicates under the condition 
that the internal argument is more specitlc than 
the sort time. This condition is necessary be- 
cause of the polysemy of those prepositions. Dur- 
ing comt)ilation the SL class definition will be au- 
tomatically expanded to the individual predicates, 
whereas the TL class dclinition will be kept unex- 
panded such that the tat'get gratnmar might be 
able to choose one of the idiosyncratic preposi- 
tions. 
Mixed approaches like Kaptan et al. (1989) can 
be characterized by mapping syntax ,as well as 
a predicate-m'gument structure (f-structure). As 
319 
already pointed out, e.g. in (Sadler and Thomp- 
son, 1991), this kind of transfer has problems with 
its own multiple level mappings, e.g. handling of 
verb-adverb head switching, and does not cleanly 
separate monolingual from contrastive knowledge, 
either. In Kaplan and Wedekind (1993) an im- 
proved treatment of head switching is presented 
but it still remains a less general solution. 
A semantic approach is much more indepen- 
dent of different syntactic analyses which axe the 
source of a lot of classical translation problems 
such as structural and categorial divergences and 
mismatches. In our approach grammars can be de- 
veloped for each language independently of the 
transfer task and can therefore be reused in other 
applications. 
At first glance, our approach is very similar 
to the semantic transfer approach presented in 
Alshawi et al. (1991). It, uses a level of underspec- 
ified senmntic representations as input and output 
of transfer. Tile main differences between out' ap- 
proach and theirs are the use of flat semantic rep- 
resentations and tile non-recursive transfer rules. 
Tile set-oriented representation allows much sim- 
pler operations in transfer for accessing individual 
entities (set membership) and for combining the 
result of individual rules (set union). Furthermore, 
because the recursive rule application is not part 
of tile rules themselves, our approach solves prob- 
lems with discontinuous translation equivalences 
which tile former approach cannot handle well. A 
transfer rule for such a case is given in (4). 
Out" current apt)roach is strongly related to 
tile Shake-and-Bake approach of Beaven (1992) 
and Whitclock (1992). But instead of using 
sets of lexical signs, i.e. morpho-syntactic lex- 
emes as in Shake-and-Bake, we specify trans- 
lation cquivalences on sets of arbitrary seman- 
tic entities. Therefore, before entering tile trans- 
fer component of our system, individual lex- 
emcs can already be decoinposcd into sets of 
such entities, e.g. for stating generalizations on 
the lexical semantics level or providing suit- 
able representations for inferences. For example, 
the wh-question word when is decomposed into 
temp loc(E,X), whq(X,R), time(R,X) (lit.: at 
which time), hence no additional transfer rules are 
required. Similarly, German composita like Ter- 
minvorschlag axe decomposed into its compounds, 
e.g. termin(i2), n n(il,±2), vorschlag(il) 
where n_n denotes a generic noun-noun relation. 
As a result a compositional translation as proposal 
for a date is possible without stating any addi- 
tional translation equivalences to the ones for the 
simplex nouns. 
Another major difference is the addition of coil- 
ditions which trigger and block the applicability of 
individual transfer rules. For instance in the spe- 
cific t,'anslation of schlecht to not good as defined 
in (3c), without conditions, one would have to add 
tile verb passen into the bag to test for such a 
specific context. As a consequence the translation 
of the verb needs to be reduplicated, whereas in 
our approach, the translation of the verb can be 
kept; totally independent of this specific transla- 
tion of tile adverbial, because the condition func- 
tions merely as a test. 
These examples also illustrates the usefulness 
of labeled conditions, because the negation op- 
erator can take such a label as an argument 
and we can use unification again to achieve 
the correct coindexation. If we would use a hi- 
erarchical semantics instead, as in the origi- 
nal Shake-and-Bake aproach, where the negation 
operator embeds the verb semantics we would 
have to translate schlecht (e), passen(e) into 
not(suit(e), well(e)) in one rule because 
there is no coindexation possible to express the 
correct embedding without the unique labeling of 
predicates. 
Finally, we have filled the lack of an adequate 
control strategy for Shake-and-Bake by develop- 
ing a nonmonotonic control strategy which orders 
more specific rules before less specific ones. This 
strategy allows the specification of powerfifl de- 
fanlt translations. Whereas without; such an or- 
dering special care is needed to prevent a compo- 
sitional translation in cases where a more specific 
noncompositional translation also exists. 
The same argument about control holds in com- 
parison to the unification-based transfer approach 
on Mimimal Recursion Semantics (MRS) (Copes- 
take et al., 1995; Copestake, 1995). In addition, we 
use matching on first order terms instead of fea- 
ture structure unification. Full unification might 
be problematic because it is possible to add ar- 
bitrary information during rule application, e.g. 
by further unifying different arguments. The other 
main difference is our nonmonotonic control com- 
ponent whereas tile MRS approach assumes a 
monotonic computation of all possible transfer 
equiv'&lences which are then filtered by the gen- 
eration grammar. It is difficult to judge the feasi- 
bility of their approach given the fact that only a 
limited coverage has been addressed so far. 
5 Implementation 
A more detailed presentation of the implementa- 
tion aspects of our transfer approach can be found 
in Dorna and Emele (1996). The current transfer 
implementation consists of a transfer rule compiler 
which takes a set of rules likc the one presented in 
section 3 and compiles them into two executable 
Prolog programs one for each translation direc- 
tion. The compiled program includes the selection 
of rules, the control of rule applications and calls 
to external processes if necessary. 
Because both the transfer input and the match- 
ing part of the rules consist of sets we can ex- 
ploit ordered set operations during compilation as 
320 
well as at runtime to speed up the matching pro- 
cess and for computing common prefixes which axe 
shared between different rules. 
The compiled trazlsfcr program is embcdded in 
the incremental and parallel axchitecture of the 
Verbmobil Prototype. Interaction with external 
modules, e.g. the domain model and dialog mod- 
ule or other inference components, is done via a set 
of predefined abstract interface functions whidl 
may be called in the condition part of transfer 
rules. The result is a hilly transpaxent and modu- 
lax interface for filtering the applicability of trans- 
fer rules. 
6 Summary 
This pai)er presents a new declarative transfi'.r 
rule forlnalisin, which provides an iInplementation 
platform for a selnantic-based transfer approach. 
This approa(:h contl)ines ideas fronl a nunlber of 
re('cnt MT proposals and tries to avoid many of 
the well known problems of other transfer and in- 
terlingua approaches. 
The deelaxativc trtmsfer correspondences m'e 
compiled into an executable Prolog program. The 
conlpiler exploits indexing for more efficient search 
of matching rules. There is a nonnlonotonic but 
rule-independent control strategy based on rifle 
specificity. 
Currently, the transfer conlponent contains 
about 1700 transfer rules. Thanks to the set ori- 
entation and indexing techniques we did not en- 
counter any scaling problenls aald the average run- 
time pcrfornlanec for a 15 word sentence is about 
30 milliseconds. 
Fhture work will include tim automatic acqui- 
sition of transfer rules fronl tagged bilingual cor- 
pora to extend tim coverage and an integration of 
domain specific dictionaries. 
References 
B. Abb and B. Buschbeck-Wolf, 1995, A Minimal Transfer 
Conception for Verbmobil. Vcrl)~nobil Report 84, Institute 
for Logic and Linguistics, IBM hlfornlationssysteme Gnfl)ll 
neidelberg, Germany. 
\[t. Alshawi, D. Carter, M. nayner, and B. Gamb~i.ck. 1991. 
Translation by Quasi Logical Form Transfer. In Proceedings 
of the 29th Annual Meeting of the Association for" Compu- 
tational Lingnistics (ACL '91), pages 161-168, Berkeley. 
J. L. Beaven. 1992. Shake-and-Bake Machine Translation. In 
Proceedings of the 14th International Conference on Corn- 
putational Linguistics (COLING '9~), Nantes, France. 
J. Bos, B. Gambaek, C. Lieske, Y. Mort, M. Finkal, and 
K. Worm. 1996. Compositional Semantics in Verl)mobil. 
In Proceedings of the 16th International Conference on 
Computational Linguistics (COLING '96). 
B. Busehbeck-Wolf and R. Niibel. 1995. Die Behandhmg iiber- 
setzungsambiger Pri4positionen im Transfer des Verbmobil- 
Demonstrators. Verbmobil Report 87, IBM, tleidelberg and 
IAI, Saarbriicken, Germany. 
n. Caspari and L. Schndd. 1994. Parsing und Generieruug in 
qYUG. Verbmobil Report 40, Siemens AG, Munich, Ger- 
IOal,y. 
A. Copestake, 11. Flickinger, R. Malouf, S. l/.iehemamm, and 
1. Sag. 1995. Transfer and Minimal Recursion Semantics. 
In Proceedings of the 6th International Conference on 
Theoretical and Methodological Issues in Machine 7~rans - 
Is|ion (TM! '95), Leuven, Belgium. 
A. Copes|eke. 1995. Semantic Transfer in Verbmobil. 
Verbmobil Report 93, IMS, Universitiit Stuttgart, Germany 
and CSLI, Stanford University. 
M. Dorna and M. Emele. 1996. Emcient imlflementatioa 
of a semantic-baaed transfer approach. In Proceedings o\[ 
the 12th European Conference on Artificial Intelligence 
(EC'AI '96). 
M. Dorna. 1996. Ttle ADT-Package for the Verbmobil In- 
terface Term. Verbmob// Report 104, IMS, Universitiit 
Stuttgart, Germany. (to appear). 
B. J. Dorr. 1993. lnterlingual Machine Translation: a Parame- 
terized Approach. volume 6311-21 of Artificial Intelligence, 
pages 429-493. 
B. J. Dorr. 1994. Machine "lYanslation Divergences: a For- 
mal Description and Proposed Solution. volmne 20(4) of 
Computational Linguistics, pages 597-633. 
D. Dowty. 1989. On the Sen,antic Content of the Notion of 
'Thematic Role'. In G. Chierehia, B. Partee, and R. Turner, 
editors, Property Theory, 7Ope Theory and Natural Lan- 
guage Semantics. Dordrecht, Reidel. 
M. Egg and K. Lebeth. 1995. Semantic mtderspecification 
and moditier attachment ambiguities. In Beitrligc zur 5. 
l,'achtaguug der Sektion (\]omputerlingnistik der 1)~fS. 
A. Frank and U. Reyle. 1995. Principle Based Semantics for 
tIPSG. In Proceedings of the 7th Conference of the Euro- 
pean Ghapter of the Association for Computational Lin- 
guistics (EAUL '95), Dublin, Ireland. 
W. J. llutchins and H. L. Somers. 1992. An Introduction to 
Machine Translation. Academic Press, London, UK. 
Hans Kamtl and Uwe Reyle. 1993. From Discourse to Logic: 
An Introduction to Modeltheoretic Semantics of Natural 
Lan9nage. Kluwer, Dordrecht, ttolland. 
tt.. M. Kaplan and J. Wedekind. 1993. Restriction and 
eorrespondance-ba.qed translation. In Proceedings of the 
6th Conference o\] the European Chapter of the Associ- 
ation for C~omputational Linguistics (EAUL '931, pages 
193 202, Utrecht, The Netherlands. 
t¢. Kaplan, K. Netter, J. Wedekind, and A. Zaenen. 1989. 
qYans\]ation by Structural Correspondences. In Proceedings 
of the 4th Uonfereuce of the European Chapter of tht: 
Association for Computational Linguistics (EACL '891, 
pages 272-281, Manchester, UK. 
M. Kay, M. Gawron, and 'I J. Norwig. 1994. Verbmobil: a 
Translation System for Face-to-Face Dialogs. CSLI Lec- 
ture Notes; No. 33. CSLI Center for the Study of Language 
and Information, Stanford University. 
A. Kilger and W. Finkler. 1995. Incremental Generation for 
Real-Time Applications. Research Report RR-95-11, I)FK1 
GmbH, Saarbriieken, Germany. 
M. Nagao, U. Tsujii, and U. Nakamura. 1985. The Japanese 
Government Project for Machine Translation. volmne l 1 (2- 
3) of (:ornputational Linguistics, pages 91-110. 
T. Parsons. 1991. Events in the Semantics of English. M1T 
Press, Cambridge, Mass. 
U. Reyle. 1993. Dealing with Ambiguities by Underspeci- 
lication: Construction, Representation and Deduction. In 
Jonnal of Semantics, volume 10(2), pages 123 179. 
L. Sadler and II. S. Thompson. 1991. Structural Non- 
correspondence in Translation. In Proceedings of thee 5th 
Conference of the European Chapter of the Association 
for Computational Lin.quistics (EAC~L '91), pages 293 
298, Berlin, Germany. 
3. Slocum, W. S. Bennet, J. Bear, M. Morgan, and R. Root. 
1987. METAL: The LIt.C Machine Translation System. In 
S. Michaelson and Y. Wilks, editors, Machine Translation 
today: the state of the art, pages 319-350. Edinburgh Uni- 
versity Press. 
B. Vauquois and C. Boitet. 1985. Automated Translation 
at Grenoble University. volume 1111) of Computational 
Linguistics, pages 28-36. 
W. Wahlster. 1993. Verbmobil -- "lYanslation of Face-to-Face 
Dialogs. In Proceedings of the 4th Machine Translation 
Summit, Kobe, Japan. 
P. Whitelock. 1992. Shake-and-Bake Translation. In Proceed- 
ings of the Igth International Conference on (Jornpnta- 
tional Linguistics (C*OLING '9~), pages 784-791, Nantes, 
France. 
321 
