Research on Architectures for Integrated Speech/Language 
Systems in Verbmobil 
Gfinther GSrz~ Marcus Kesseler, JSrg Spilker, Hans Weber 
University of Erbmgen-Nfirnberg 
IMMD (Computer Science) VIII --- Artificial Inl0elligence 
Am Weichselgarten 9 
D-91058 ERLAN(\]EN 
Emaih {goerz, kesseler, spilker, weber}Oinformat ik. uni-erl angen, de 
Abstract 
The German joint research project Verb- 
mobil (VM) aims at the deveh)pment 
of a speech to speech translation sys- 
tem. This paper reports on research 
done in our group which belongs to Verb- 
mobil's subprojeet on system architec- 
tures (TP15). Our specific research areas 
are the construction of parsers for spon- 
taneous speech, investigations in the par- 
allelization of parsing and to contribute 
to the development of a flexible com- 
munication architecture with distributed 
control. 
1 Introduction 
The German joint research project Verbmobil 
(VM) 1 aims at the development of a speech to 
speech translation system. This paper reports 
on research done in our group which belongs to 
Vcrbmobil's subproject on system architectures 
(TP15). The task of this subproject is to provide 
basic research results on incremental and inter- 
active system architectures for the VM research 
prototype and to demonstrate their feasibility in 
the prototypical INTARC system. Our specific 
research areas are the construction of parsers for 
spontaneous speech, investigations in the paral- 
lelization of parsing and to contribute to the de- 
velopment of a flexible communication architec- 
ture with distributed control. The paper is or- 
ganized as follows: Section 2 reports on the de- 
sign and implementation of an incremental in- 
teractive speech parser which integrates statistics 
i 
1This work was flmded by the German Federal 
Mirfistry for Researdl and Tedmology (BMFT) in 
the framework of the Verbmobil Project under Grant 
BMFT 01 IV 101 H / 9. The responsibility for the 
contents of this study lies with the authors. 
with a chart-parser einploying a unification gram- 
mar (UG) formalism. Furthermore, results of ex- 
periments on the interaction between the parser 
and a speech recognizer using expectations, are re- 
ported. In section 3 we present experiences with 
a parallel version of the parser. Section 4 deals 
with distributed control in modular Natural Lan- 
guage/Speech (NLSP) systems. 
2 Design and Implementation of 
Incremental Interactive Speech 
Parsers 
In a Left Right Incremental architecture (LRI), 
higher level modules can work in parallel with 
lower level modules. The obvious benefits of such 
an arrangement are twofold: The system does not 
have to wait for a speaker to stop talking and top- 
down constraints from higher level to lower level 
modules can be used easily. 'lb achieve LR! behav- 
ior the singular modules must fulfil\] the following 
requirements: 
Processing proceeds incrementally along the 
time axis ("left to right"). 
Pieces of output have to be transferred to the 
next module as soon as possible. 
So far in INTARC-1.3 we have achieved an LRI 
style coupling of ibur different modules: Word 
recognition module, syntactic parser, semantic 
module and prosodic boundary module. Our word 
recognition module is a modified Viterbi decoder, 
where two changes in the algorithm design were 
made: We use only the forward search pass, and 
whenever a final ttMM state is reached for an ac- 
tive word model, a corresponding word hypothesis 
is sent to the parser. Itence backward search be- 
comes a part of the parsing algorithm. The LRI 
parsing algorithm is a modified active chart parser 
with an agenda driven control mechanism. The 
chart vertices correspond to the .frames of the sig- 
nal representation. Edges correspond to Word or 
phrase hypotheses, being partial in the case of ac- 
484 
tive edges. A parsing cycle corresponds to a new 
time point related to the utterance, in every cycle 
a new vertex is created and new word hypotheses 
ending at that time point are read and inserted 
into the chart. In one cycle, a ba(:kwar(| search 
is performed to the beginning of the utterance or 
to some designated time point in the past con~ 
stitnting a starting point for grammatical analy- 
sis. Search is guided by a weighted linear com- 
bination of acoustic score, bigram score, prosody 
score, grammar derivation seore and grammatical 
parsability. The search prodecure is a beam search 
implemented as an agenda access mechanism. The 
grammar is a probabilistic typed UG with sepa- 
rate rules for pauses and other spontanous speech 
phelnonmua. 
2.1 Basic Objects 
In the h)llowing we use record notation to refer 
to subcoml)oncnts of an object. A chart ver- 
tex Vt corresponds to frame number t. Vertices 
have four lists with pointers to edges ending in 
and starting in thai; vertex: inactive-out, inactive~ 
in, active-out and active-out. A word hypothe- 
sis W is a quadruple (from, to, key, score) with 
J}vm and to being the start and end frames of W. 
W.Key is the name of the lexical entry of W and 
W.score is the acoustic score of W for the frames 
spanned, given by a corresponding HMM acoustic 
word model. An edge i',' consists of flora, the start 
vertex and to, a list of end vertices. Note that af- 
ter a Viterbi forward pass identical word hypothe- 
ses do always come in sequence, differing only in 
ending time. E.actual is the last vertex added 
to t£to in an operation. Those "families" of hy- 
potheses are represented as one edge with a set of 
end vertices. E.words keeps the covered string of 
word hypotheses while SCORE is a record keeping 
score components. Besides that an edge consists 
of a grammar rule E.rule and F,.next, a pointer 
to some element of the right hand side of l£rule 
or NIL. As in standard active chart parsing an 
edge is passive, if E.ncxt = nil, otherwise it is 
active. E.eat points to the left hand side of the 
grammar rule. SCORE is a record with entries for 
inside and outside probabilities given to an edge 
by acoustic, bigram, prosody and grammar model: 
Inside-X Model scores for the spanned portion 
of an edge. 
Outside-X Optimistic estimates for the portion 
fi'om vertex 0 to the beginning of an edge. 
For every vertex we keep a best first store of scored 
edge pairs. We (:all that store Agenda/ in cycle i. 
2.2 Basic Operations 
'I'here are tive basic operations to detine the op- 
erations of the parsing algorithm. The two op- 
erations Combine and Seek Down are similar to 
the well known Earley algorithm operations Com- 
pleter arm Predictor. Furthermore, there are two 
operations to insert new word hypotheses, Insert 
and Inherit. All these operations can create new 
edges, so operations to calculate new scores from 
old ones are attached to them. hi order to im- 
plement our t)eam search method appropriately 
but sinq)ly, we define an operation Agenda-Pu~q~ , 
which selects pairs of active and passive edges to 
be prmn;d or to be processed in the future. The 
Ct (, notation for l)asic operations are given in -' ,' ' 
simplicity. 
2.2.1 Combine 
For a t)air of active and passive edges (A, l), 
if A.next = I.cat and Lfivm ~ A.to, insert edge 
ti with l','.rule - A.rule, E.cat :-- A.eat, E.nexl --: 
shifl(A.uext), l£fl~)m.-= A.fl'o,n, l£to = A.to. 
For X = Bigram, Grammar and Prosody: 
t£Outside-X = A.Outside-X + I.lnside-X + 
7'tans(X, A, 1) 
l£ lnsidc-X = A.lnside-X 4- 1.1nside-X 4- 
7'rans(X,A,1) 
I"or X = Acoustic: 
E. Outside-X = A. Oulside-X\[l.from\] @ L lnsi&'- X 
"Frans(X,A, I) 
l£1nsidc-X = A.lnsidc-X\[1.flvm\] (t) LInside-X 
Trans(X,A,l) 
'l)he operator ct) performs an addition o\[a nun> 
her to every element of a set. Trans(X,A,l) is the 
specilic transition penalty a model will give to two 
edges. In the ctLse of acoustic scores, the penMty is 
always zero and can be neglected. In the. bigram 
c~use it will be the transition from the last word 
covered by A to the tirst word covered by B. 
2.2.2 Seek Down 
Whenever an active edge A is inserted, insert an 
edge E for every rule 1~ such that A.next - E.cat, 
I','. rule = If, F,.flvm - A. actual, t3. to = {A. actual} 
• For X --- Acoustic, Prosody and l)igraln: 
E.lnside-X = 0 
I£ Outside-X = A. Outside-X 
For X == Grammar: 
1£ lnside-X = grammar score of II 
l£Outside-X = A.Outside-X + 7~rans(X,A,l';) + 
I~.inside-X. This reeursive operation of introduc- 
ing new active edges is precompiled in our parser 
and extremely etlicient. 
485 
2.2.3 Insert 
For a new word hypothesis W = (a,i, key, score) 
such that no W' = (a,i-i,key, score') exists, in- 
sert an edge E with E.rule = lex(key), E.cat = 
lex(key), E.from = Va, E.to = {~} and for X = 
Acoustic: 
E.Inside-X = E. Outside-X = { (i, score)}, 
for X = Prosody and Bigram: 
E.Inside-X = E.Outside-X = O, 
for X = Grammar E.Inside-X = E.Outside-X = 
grammar score of lex(keg). 
2.2.4 Inherit 
For a new word hypothesis W = (a,i, key, score) 
such that a W' = (a,i-l,key, score') exists: 
For all E in Vi_l.inactive-in or Vi_l.active- 
in: If last(E.words) = key then add {~} to 
E.to, add (i,E.Inside-Acoustic\[i-l\]- score' + 
score) to E.Inside-Acoustic and add (i,E.Outside- 
Acoustic\[i-I\]- score' + score) to E.Outside- 
Acoustic. 
If E is active, perform a Seek-Down on E in ~. 
2.2.5 Agenda Push 
Whenever an edge E is inserted into the chart, 
if E is active then for all passive A, such that 
A.from 6 E.to and combined-seore(E,A) > Beam- 
Value, insert (E,A, combined-score(E,A )) into the 
actual agenda. If E is passive then for all active A, 
such that E.f~vm 6 A.to and combined-score(A,E) 
> Beam- Value, insert (A,E, combined-score(A, E)) 
into the actual agenda. Combined-Score is a lin- 
ear combination of the outside components of an 
edge C which would be created by A and E in a 
Combine operation. Beam-Value is calculated as 
a fixed offset from the maximum Combined-Score 
on an agenda. Since we process best-first inside 
the beam, the maximum is known when the first 
triple is inserted into an agenda. Agenda-Pop will 
remove the best triple from an actual agenda and 
return it. 
2.3 A simple LRI lattice parser 
The follwing control loop implements a simple LRI 
lattice parser. 
1. T = 0. Create VT 
2. Insert initial active edge E into VT, with 
E.next = S 
3. Increment T. Create VT 
4. For every W with W.end = 7': Insert(W) or 
Inherit(W) 
5. Until Agenda\[T\] is empty: 
(a) Combine(Agenda-Pop) 
(b) When combination with initial edge is 
successful, send result to SEMANTICS 
6. Communicate with PROSODY and go to 3 
2.4 The Grammar Model 
The UG used in our experiments consists of 700 
lexical entries and 60 rules. We used a variant 
of inside-outside training to estimate a model of 
UG derivations. It is a rule bigram model simi- 
lar to PCFG with special extensions for UG type 
operations. The probability of future unifications 
is made dependent from the result type of ear- 
lier unifications. The model is described in more 
detail in (Weber 1994a; Weber 1995); it is very 
similar to (Brew 1995). 
2.5 LRI Coupling with Prosody 
In INTARC we use three classes of boundaries, 
B0 (no boundary), B2 (phrase boundary), B3 
(sentence boundary) and B9 (real break). The 
prosody module, developed at the University of 
Bonn, classifies time intervals according to these 
classes. A prosody hypothesis consists of a be- 
ginning and ending time and model probabilities 
for the boundary types which sum up to one. A 
prosodic transition penalty used in the Combine 
operation was taken to be the score of the best 
combination of bottom-up boundary hypothesis 
Bx and a trigram score (lword, Bx, rword). Here 
lword is the last word of the edge to the left and 
rword is the first word spanned by the edge to the 
right. Prosody hypotheses are consumed by the 
parser in every cycle and represented as attributes 
of vertices which fall inside a prosodic time inter- 
val. In a couple of tests we already achieved a re- 
duction of edges of about 10% without change in 
recognition rate using a very simple trigram with 
only five word categories. 
2.6 Experimental Results 
In a system like INTARC-1.3, the analysis tree 
is of much higher importance than the recovered 
string; for the goal of speech translation an ad- 
equate semantic representation for a string with 
word errors is more important than a good string 
with a wrong reading. The grammar scores have 
only indirect influence on the string; their main 
function is picking the right tree. We cannot mea- 
sure something like a "tree recognition rate" or 
"rule accuracy", because there is no treebank for 
our grammar. The word accuracy results cannot 
be compared to word accuracy as usually applied 
to an acoustic decoder in isolation. We counted 
only those words as recognized which could be 
486 
built into a valid parse from the beginning of the 
utterance. Words to the right which could not 
be integrated into a parse, were counted as dele- 
tions --- although they might have been correct in 
standard word accuracy terms. This evaluation 
method is much harder than standard word accu- 
racy, but it appears to be a good approximation 
to "rule accuracy". Using this strict method we 
achieved a word accuracy of 47%, which is quite 
promising. 
Results using top down prediction of possible 
word hypotheses by the parser work inspired by 
(Kita et. al. 1989) have already been published 
in (Hauenstein and Weber 1994a; ltmlenstein and 
Weber 1994b), (Weber 1994a), and (Weber 1995). 
Recognition rates had been improved there for 
read speech. In spontaneous speech we could not 
achieve the same effects. 
2.7 Current Work 
Our current work, which led to INTARC-2.0, uses 
a new approach for the interaction of syntax and 
semantics and a revision of the interaction of the 
parser with a new decoder. For the last case we 
implemented a precompiler for word-based pre- 
diction which to our current experience is clearly 
superior to the previous word-class based predic- 
tion. For the implementation of the interaction 
of syntax and semantics we proceed as follows: A 
new turn-based UG has been written, for which a 
context-sensitive stochastic traiuing is being per- 
formed. The resulting grammar is then stripped 
down to a pure type skeleton which is actually 
being used for syntactic parsing. Using full struc- 
ture sharing in the syntactic chart, which con- 
tains only packed edges, we achieve a complex- 
ity of O(n3). In contrast to that, for semantic 
analysis a second, unpacked chart is used, whose 
edges are provided by an unpacker module which 
is the interface between the two analysis levels. 
The unpacker, which has exponential complexity, 
selects only the n best scored packed edges, where 
n is a parameter. Only if semantic analysis fails it 
requests further edges from the unpacker. In this 
way, the computational effort on the whole is kept 
as low as possible. 
3 Parallel Parsing 
One of our main research interests has been the 
exploration of performance gains in NLP through 
parallelization. To this end, we developed a par- 
allel version of the INTARC parser. Although 
the results so far are yet not as encouraging 
as we expected, our efforts make for interest- 
ing lessons in software engineering. The parallel 
parser had to obey the tbllowing restrictions: Run- 
ning on our local shared memory lnnltiprocessor 
(SparcServerl000) with 6 processors, paralleliza- 
tion should be controlled by inserting Solaris-2.4 
thread and process control primitives directly into 
the code. The only realistic choice we had was 
to translate our parser with Chestnut Inc.'s Lisp- 
to-C-Translator automatically into C. Since the 
Lisp functions library is available in C source, we 
could insert the necessary Solaris parallelisation 
and synchronization primitives into key positions 
of the involved fnnctions. 
3.1 Parallelization Strategy and 
Preliminary Results 
For effective parallelization it is crucial to keep 
communication between processors to a minimum. 
Early experiments with a fully distributed chart 
showed that the effort required to keep the partial 
charts consistent was much larger that the poten- 
tial gains of increased parallelism. The chart must 
be kept as a single data structure in a shared mem- 
ory processor, where concurrent reads are possi- 
ble and only concurrent writes have to be serial- 
ized with synchronisation primitives. An analysis 
of profiling data shows that even the heavily op- 
timized UG formalism causes between 50% -and 
70% of the compntational load in the serial c~e. 
Therefore we provide an arbitrary number of uni- 
fication workers running in parallel which are fed 
unification tasks from the top of an agenda sorted 
by scores. Due to the high optimization level 
of the sequential parser, load-balancing is faMy 
poor. Namely, the very fast type check used to 
circumvent most unifications, causes large dispar- 
ities in the granularity of agenda tasks. Further- 
more, pathological examples have been found in 
which a single unification takes much longer than 
all other tasks combined. 
I ~::~ :: "~- ~'~1 
4~ 
-IM • , , • , , , , , 
3ttxl ~2 ~tz3 ~t4 ~5 Sta6 $*1z7 Sstz$ ~lz9 r,~tlO 
Figure 1: PercentuM gains and losses over at- 
tained over 10 ditferent sentences (Spilker 1995) 
487 
4 Distributed Control in 
Verbmobil 
The question of control in VM is tightly knit 
with the architecture of the VM system. As yet, 
the concept of architecture in VM has been used 
mostly to describe the overall modularization and 
the interfaces implied by the data flow between 
modules. This socalled dornair~ architecture is in- 
complete in the sense that it does not specify 
any interactio~ strategics. Within our research 
on interactive system architectures we developed a 
modular communication framework, ICE ~, in co- 
operation with the University of Hamburg. Now, 
ICE is the architectural framework of the VM re- 
search prototype. 
4.1 The INTARC Architecture 
The INTARC architecture as first presented by 
(Pyka 1992) is a distributed software system that 
allows for tile intcrconncction of NLSP modules 
under the principles of incrementality and inter- 
activity. Figure 2 shows the modularization of 
INTARC-1.3: There is a main broad channel con- 
necting all modules in bottom-up direction, i.e., 
from signal to interpretation. Furthermore, there 
are smaller channels connecting several modules, 
which are used for the top-down interactive dis- 
ambiguation data flow. Inerementality is required 
for all modules. ICE assumes that each module 
has a local memory that is not directly accessi- 
ble to other modules. Modules communicate ex- 
plicitly with one another via messages sent over 
bidirectional channels. This kind of communica- 
tion architecture is hardly new and eonl¥onts us 
directly with a large number of unresolved issues 
in distributed problem solving, ef. (Durfee et al. 
1989). In the last 20 years there have been numer- 
ous architecture proposals for distributed prob- 
lem solving among computing entities that ex- 
change information explicitly via message passing. 
None of these models include explicit strategies 
or paradigms to tackle the problem of distributed 
control. 
4.2 Structural Constraints of Verblnobil 
Modularity, being a fundamental assumption in 
VM (Wahlster 1992), does still leave us with two 
problems: First, modules have to communicate 
with one another, and second, their local behav- 
iors have to be somehow coordinated into a coher- 
ent global, possibly optimal, behavior. Unfortu- 
nately, the task of system integration has to obey 
some structural constraints which are mostly prag- 
matic in natnre: 
2based on PVM (parallel virtual madfine) 
semRnb~ 
represenl~(m 
t 
I 
. ............. / 
-.. ~-?a,fi~f~.~ ~,, ) 
~ho.. 
~'~z,- Ol~mb~g~ogonDet*lFIow(TopDo~n) 
MoJtl Da~ Flo* (l-lottom Ui)) 
Figure 2: 'l)he interactive, incremental INTARC- 
1.3 architecture 
Some of the modules are very complex soft- 
ware systems in thelnselves. Highly parame- 
terizable and with control subtly spread over 
many interacting submodules, understanding 
and then integrating such systems into a com- 
mon control strategy can be a very daunting 
task. 
Control issues are often very tightly knit with 
the domain the module is aimed at, i.e., it 
is very difficult to understand the control 
strategies used without sound knowledge of 
the underlying domain. The problem even 
gets worse if what is to be fine-tuned is the in- 
teraction between several complex modules. 
These two arguments are similar in nature, but 
diflhr in the architecturM levels that they apply to. 
'File former is implementation related, the latter 
algorithm arid theory related. 
4.3 Layers of Control 
Modules have to colnmunicate with one another 
and their local behaviors have to be coordinated 
into a coherent global, possibly optimal, behavior. 
In highly distributed systems we generally tind the 
following levels of control: 
System Control: The minimal set of operat- 
ing system related actions that each participat- 
ing module must be able to per\[brm which will 
488 
typically include means to start up, reset, moni 
tot, trace and terminate individual modules or the 
system as a whole. 
Is()lated Local Controh The control strate- 
gies used within the module, disregarding any in- 
teractions beyond initial input of data and final 
output of solutions. 'Fhere is only one thread of 
control active at any time. 
lnteraetiv(; Local Controh ll.oughly, this can 
be seen as isolated local control extended with in- 
teraction cal)al/ilities, lncr~"mentalitp is given by 
the l)ossibility of control flowing back to a certain 
internal stake aftex an outl)ut operation, lligher 
mtcraclivily is made possible by entering a state 
more often fl:om w~rious points within the rood- 
tile and by adding a new waiting lool/to cheek for 
any tot)-down requests. The requirement for any. 
time behavior is a special case of that (G6rz and 
Kesseler \]994). 
in ore: experience ~he change to interactive COl> 
trol will tremendously increase the complexity of 
the resulting (:ode. But we are swill making the 
simplifying assumptions that tile algorithm can 
be used increnlentally - but there are algorithms 
m~suitable for incremental processing (e.g. A*). 
h~crementality can lead to the (\[elrlalld for a eoln- 
i)\]ete redesign of a lnodule. Furthexmore we as- 
sume that simply by exchanging data and doing 
simple extensions in the control \[tow everything 
will balan(:e out nicely on the system scale which 
is enorlnously naiv(:. Even for the sequential archi- 
tecture implied by the case of isolated local con- 
trol, we have to solve a whole plethora of uew 
problems that corne along with interaetivity: 
* Mutual deadlock 
- Mutual live-lock 
- Race conditions (missing synchronization) 
,, Over-synchronization 
Dialogue Controh In systems like VM there is 
a module that comes close to possessing the "inte-- 
grated view" ot'a centralized blackboard control: 
the dialogue module. So it seems the right place to 
handle some of the global strategic control issues, 
like: 
- l)omain error trundling 
. Observe timeont constraints 
. ll,esolve, external ambiguitie.s/unl(nowns 
The fact that tile dialogue module exercises a kind 
of global control does not invalidate what has bee, n 
said about the unfeasability of central control, be.. 
cause the control exercised by it is very coar.',e 
grained. To handle liner grained control issues in 
any rood ule would take us back to memory and/or 
eomm,mication system contention. 
References 
Chris Brew. Stochastic t\]I'SG, l)roccx,dings of l.he 
\]'hn'ol)ean ACI, Conference 1995, b;din/)urgh, 1995 
l';dnmnd 11. I)urfec, Victor I{. I,cssev, and I)anM I). 
Corkill. Coop~rative l)istribut{d Probh:m Soh.,ing, 
pages 83 147. Volume 4 ot: Aw'on I{arr et al. (b;d.). 
The ll.ndbook of Arlifi'ciul lntclli.qence l{eading, 
Mass.: Addison-Wesley, 1989. 
(',iitdher G6rz and Marcus I<esschw. Anytime AL 
gorithms for Speech Parsing? Proceedings of 
(\]OIAN(L94, Kyoto, 19{)4 
Andrcas llauensl,eiu and \[laus Weber. An invesligao 
tion of tightly coupled time synchronous speech lan- 
guage inlcEfaccs usin 9 a unification grammar. In: 
l'. McKevitt (Ed.): Proceedings of tile Workshop on 
Integration of Natural Language and Speech Pro- 
(:esslng, AAAI-94, Seatde, \[994, 42 49 
Andrcas lIauenstein and Hans Weber. An lnveatiga- 
lion of Tightly Coupled Time Synchronous Speech 
Language hderfaces. Proceedings of K()NVI'\]NS 94, 
Vietma: Springer, September 19!)4, 141 150 
Marcus Kcsseh!r. l)istribvted Control in Vcrbmo- 
bil+ Uniwwsil.y of Erlangen-N iit'nberg, IMM D V\[ l 1, 
Verbmot)il l{el)Ort 24, 1994. 
Kit.a, K., I<awabata, T. and Saito, \[I. HMM conliu- 
ous speech recognition usin9 predictive L 1~ parsiwI. 
IEEE IC,4,S'SI ~ Proceedings , 1989, 703 706. 
(\]laudius \[)yka. Management of hypotheses in an iutc~ 
.qralcd spccch-la~*guage architecture. Proceedings of 
EUAl-92, 1992, 558 560 
36vg Spilker. Parallclisicrun 9 eines inkrcmc.nlellcn, 
aklivcn (\]hartparscrs. Diploma thesis, Uniw~rsil;y 
of F, rlangen-N6rnberg, l",rlangen, December 1995. 
Wolfgmlg Wahlstcr and Judith l';ngelkamp, editors. 
Wissenschafllichc Ziele und Nctzpldnc filr das 
VERBM 0 BlL-Projekt. I)FK 1, SaarbriM~en, \] 992. 
Ilans Weber. Time Synchronous Uhart Parsing 
of ,5'pccch lntcgratin 9 UniJication Grammars wilh 
Statistics. Speech arid l,anguage l~\]ngineering, Pro- 
ceedings of {;he l';ighth Twente Workshop on Lan- 
guage Technology, (1,. Bores, A. Nijholt, Ed.), 
Twentc, 1\[)94, 107-119 
tlans Weber. I,R-inkrcmcntclles probabili- 
stisches Chartparsing yon Worthypothcsenmengcn 
mit Unifikations.qrammatiken: Eine engc Kopplung 
yon Suche und Analyse. Ph.l). Thesis, University 
of Hamburg, 1995, Verbmobil Report 52. 
489 
