Discourse particles and routine formulas 
in spoken language translation 
Manfred Stede and Birte Schmitz 
Technische Universit~it Berlin 
Sekr. FR 5-12 
Franklinstr. 28/29 
10587 Berlin, Germany 
Email: {stede \[ birte}@cs, tu-berl in. de 
Abstract 
The problem of polysemy has received 
much attention when dealing with content- 
words, but it is just as difficult for dis- 
course particles: In spoken language, they 
often perform various functions for dia- 
logue management, rather than contribut- 
ing to propositional content. Different lan- 
guages have evolved different conventions 
on using such particles, which renders the 
task for spoken language translation quite 
difficult. We focus here on particles in Ger- 
man, suggest a framework for represent- 
ing their roles in utterances, and sketch an 
approach for adequately translating them 
into English. 
1 Overview 
Discourse particles at first sight seem to be inno- 
cent little words, but they can pose significant prob- 
lems for automatically processing spoken language. 
Their abundance (types and tokens alike) varies no- 
tably from language to language. In this paper we 
are concerned with German and English; in the for- 
mer, a particularly large number of such particles 
is in use. In the next section, we demonstrate that 
these particles can be quite problematic in transla- 
tion, drawing on extensive corpus analyses we per- 
formed in the VERBMOBIL project \[Wahlster 1993\]; 
thus our examples are all from the domain of ap- 
pointment scheduling. Thereafter, we discuss the 
consequences for automatic spoken language trans- 
lation (SLT) and suggest a framework of discourse 
functions to adequately represent the role of the par- 
ticles within utterances. We then point out that 
certain routine formulas can be tackled with the de- 
scription of discourse functions as well. Finally, we 
discuss how our inventory of discourse functions can 
be used to improve translation quality in an SLT 
system such as VERBMOBIL. 
2 Discourse particles: a rich source 
of ambiguity 
Discourse particles are words that are not uttered 
because of their propositional content, but because 
of some pragmatic function for the ongoing dis- 
course. Imagine, for instance, utterance (1) in the 
midst of a discussion; the right at the beginning 
serves mainly to signal turn-taking and initiating 
some kind of break in the conversation. 
(1) Right, now let's turn to the next topic. 
The exact function of a discourse particle is often 
difficult to determine, though, and thus the need for 
disambiguation arises: In the German utterance (2), 
which can be roughly translated as So we're finished, 
theja can be a mere filler that smoothens the intona- 
tion contour, or it can mark the overall information 
as given, e.g. in a situation where the participants 
have just collectively closed off the final topic. 
(2) Dann sind wir ja fertig. 
The disambiguation problem is aggravated consid- 
erably by the fact that the vast majority of words 
that can be used as discourse particles also have a 
"standard", lexically-determined, reading) In (1), 
the now can be read as a temporal adverb or as a 
semantically empty particle. Similarly, in (2) the 
dann is most likely just an 'uptake' particle, but in 
general it is a temporal adverb as well. 
We will follow the terminology of Hirschberg and 
Litman \[1993\], who distinguish between the senten- 
tial usage and the discourse usage of such words. Ill 
these terms, the problem is that many particles have 
a sentential usage, but it may very well be irrelevant, 
in certain contexts, where only the discourse usage 
of the particle is relevant -- and often that. usage has 
no systematic relationship to the sentential usage. 
1Exceptions are, for instance, the English oh or the 
German ach, which are always discourse particles. 
3 
Several studies of English discourse particles have 
suggested that utterance-initial position is the cen- 
tral criterion to distinguish sentential from dis- 
course reading (e.g., Reichman \[1985\]), and cer- 
tainly, adopting such a criterion makes it much eas- 
ier to gather and evaluate the data (cf. Byron and 
Heeman \[1997\]). The position-criterion, however, 
holds for a certain class of discourse particles at best. 
Hirschberg and Litman \[1993\] give some counter- 
examples to the hypothesis. Furthermore, it does 
not travel well to the German language, despite 
tile fact that Ripplinger and Alexandersson \[1996\] 
also give the utterance-initial criterion for identify- 
ing German discourse particles, such as denn, also, 
ja, in the VERBMOBIL corpus. In fact, these and 
other particles can occur almost everywhere in the 
utterance. 
For this and other reasons, the class of discourse 
particles is a heterogeneous one and very difficult 
to demarcate in syntactic or other formal ways. 
Schiffrin \[1987, p.31ff\] develops a lengthy defini- 
tion of units of talk and then treats discourse parti- 
cles (markers in her terminology) as bracketing such 
units. For our purposes here, we do not attempt to 
give formal criteria for what constitutes a discourse 
particle; instead we are content with the functional, 
dual-usage description. This results from taking a 
rather broad perspective on the range of functions 
of discourse particles, which will be developed in sec- 
tion 3. 
For the task of translating utterances containing 
discourse particles, the monolingual ambiguity prob- 
lem is extended by that of finding an appropriate 
translation -- and different languages have devel- 
oped quite different conventions for using particles. 
German is known to offer an especially wide range of 
discourse particles, and therefore it is not surprising 
that many of them do not have any straightforward 
English translation at all -- instead, their function 
in discourse needs to be signalled by different means. 
And in many cases a particle is best dropped from 
the translation process altogether, if its existence is 
due merely to certain conventions in spoken German, 
which do not carry over to the target language. 2 The 
problem is, given a particular utterance containing 
a particle, to tell which case is at hand. 
~Furthermore, the problem is amplified by the fact 
that German particles can be combined to form con- 
glomerates, as in Wir sollten wohl doch noch rnal einen 
Terrnin ausrnachen. To what extent these can be an- 
alyzed and translated compositionally is an open ques- 
t.ion. In the present paper, though, we deal only with 
individual part.icles. 
2.1 German examples 
The VERBMOBIL system currently operates with a 
German vocabulary of 2300 words, and among these 
we have identified 49 particles that cause problems 
of the kind just described. (This figure does not 
include modal and other particles that can pose sig- 
nificant problems for translation but do not have a 
discourse usage.) Although we have not performed 
large-scale frequency tests, some initial counts sug- 
gest that a typical dialogue from the VERBMOBIL 
corpus, which consists of about 15 turns, contains 20 
to 30 occurrences of such particles. For this paper, 
we have chosen 7 particles to illustrate the situation 
for disambiguation and translation. 
doch If Doch. t is the sole response to an ut- 
terance, dochl denies what the other participant 
has just said and at the same time reaffirms the 
(opposite) opinion of the speaker. In English, one 
would say something like Yes it is! (or use the more 
specific verb in question). However, when doch is 
used within a sentence, it has at least the following 
three functions. In Das ist doch klar, doch~ signals 
that the proposition is assumed to be shared knowl- 
edge, or self-evident. A suitable English rendering 
is That's clear anyway. 
(3) Lassen Sie uns doch einen Terrain ausmachen. 
But when a sentence like (3) introduces an ex- 
change, doch3 merely smoothens the utterance and 
should not be translated at all, since English does 
not offer a corresponding word. Thus, a translation 
is Let us arrange an appointment. 
(4) Dann nehmen wir DOCH Dienstag. 
Finally, in an utterance like (4), where doch4 is 
prosodically marked, it signals a return to a previ- 
ous state of the negotiation: Tuesday had been sug- 
gested earlier but was rejected, and now the rejection 
is taken back. Again, there is no equivalent English 
word; instead, a speaker can signal the reversal of 
her position by saying, for instance, All right, so we 
DO take Tuesday. 
noch Another example is noch, which can be a 
semantically empty smoothening particle, or a focus 
particle meaning another. 
(5) Wir miissen noch einen Terrain ausmachen. 
Sentence (5) is ambiguous between We have to 
arrange an appointment and We have to arrange 
another appointment. Often, prosody indicates the 
distinction \[Bos et al. 1995\]; otherwise, analysis of 
the preceding context is necessary: Has a different 
appointment been scheduled already, so that another 
one can be dealt with now? 
also The "literal" meaning (according to dictio- 
naries) of also1 is therefore or so. 3 In a related usage, 
also2 can introduce a reformulation or specification 
of information: 
(6) Treffen wir uns am Dienstag, also am dritten 
Juni. 
Sentence (6) can be translated as Let us meet on 
Tuesday, hence on the third of June (the more formal 
therefore would be misplaced here). 
(7) Also, da muss ich real schauen. 
However, in (7), alsoa does not contribute to sen- 
tence meaning but merely signals that the speaker 
is taking her turn. In English, this is typically per- 
formed by now, ... or well, ..., etc. 
bitte Sometimes, the dialogue act of the preced- 
ing utterance can be the decisive piece of informa- 
tion for disambiguation. The single-word utterance 
bitte is ambiguous between you're welcome! and ex- 
cuse me?, asking the other participant to repeat the 
utterance. Now, in case the dialogue act of the pre- 
ceding utterance is THANK (which abstracts over 
the various linguistic means of expressing gratitude), 
bitte is translated as you're welcome. (Prosodic in- 
formation can also help to distinguish the two forms 
of bitte, but is not always reliable.) 
beziehungsweise This is a curious German word 
that can be employed for quite a variety of purposes. 
(8) Mir passt es am Dienstag beziehungsweise am 
Freitag gut. 
For example, in (8) the speaker lists two alterna- 
tives, so that beziehungsweisel is best translated as 
or, thus: For me, Tuesday or Friday is fine. 
(9) lch bin am Dienstag und Mittwoch, also am 
dritten beziehungsweise am vierten, in Hamburg. 
In an enumeration with specification, such as 
(9), the English and ... respectively is an ade- 
quate translation for beziehungsweise2; only in this 
case, a relation between different elements is ver- 
balized, which is part of the "literal" meaning of 
beziehungsweise. It requires a certain amount of 
reasoning, though, to notice this relation between 
'Tuesday' and 'third', and between 'Wednesday' and 
'fourth', so that the translation I am in Hamburg 
on Tuesday and Wednesday, hence on the third and 
fourth, respectively is achieved. 
(10) Am Montag bin ich in Frankfurt, beziehungs- 
weise in Eschborn. 
Often, however, beziehungsweise3 merely starts a 
self-correction, specification, or reformulation, best 
translated as or rather, as in (10): On Monday I am 
in Frankfurt, or rather in Eschborn. (Eschborn is a 
smaller town neighbouring Frankfurt). 
3Notice that German also has no relationship to En- 
glish also -- they are genuine false friends. 
vielleicht Dictionaries give the meaning maybe 
for this adverb, but in spoken language it is often 
produced merely as a filler word, partly indicating 
vagueness, partly buying time for the speaker to 
do more thinking. In (11), the vielleicht2 is to be 
treated as such a filler; the literal translation Then 
let us maybe say Tuesday is unnatural. 
(11) Dann sagen wir vielleicht am Dienstag. 
ja As a complete utterance, the meaning ofja is 
quite significant ('yes'), but as a little word accom- 
panying utterances it can be used for many, often 
less significant, purposes. For instance, it can serve 
as a simple uptake, similar to alsoa; or it can be a 
plain filler that should not be translated at all; or 
it can indicate that the speaker assumes something 
to be presupposed by the hearer as well -- in a ne- 
gotiation, someone would say (12) if he had already 
given the information On Monday I am in Hamburg 
earlier. 
(12) Montag bin ich ja in Hamburg. 
2.2 English examples 
In English, particles are generally less frequent than 
in German. Correspondingly, using particles for dis- 
course purposes is also less common than in Ger- 
man. While we have not done comprehensive cor- 
pus studies on English data yet, the limited role of 
particles can be inferred from the fact that the re- 
search literature often deals merely with those occur- 
ring utterance-initial (e.g., \[Byron, Heeman 1997\]). 
Especially the structuring particles are well-studied; 
for instance, by the way and but can signal the be- 
ginning of a digression, and anyway the return to a 
previous topic. 
Other prominent utterance-initial particles are 
those that signal turn-taking (we call them 'up- 
takes'), as in (13) and (14), taken from the English 
VERBMOmL data. In (14), the well is not just a 
neutral uptake but also indicates reservation. 
(13) Alright, why don't you come to my of/ice. 
(14) Well, the morning of the eighteenth is bad. 
Some studies look at particles in other positions 
as well. For instance, Redeker \[1990\] considers 
utterance-final tags like okay? or right?, which serve 
to elicit acknowledgement from the listener, and 
comment-clauses, which can be used as 'enquoting 
devices' to signal the beginning of quoted speech. 
Other particles, which we also find in the English 
VERBMOBIL data, include repair markers such as the 
oh in (15) and the no in (16). Incidentally, (16) is 
an example for a correction that does not pertain to 
the part of the utterance immediately preceding the 
correction marker. 
(15) I'm going to be in Cleveland on the sixth, 
discourse function 
structuring coherence-marking attitudinal 
push pop uptake check repair, positive negative indifferent given surprise 
smoothening 
filler 
Figure 1: Taxonomy of discourse functions 
how about sometime, oh, actually maybe this week 
isn't going to be so good. 
(16) I'm on vacation from the second to the 
twelfth, I'm going to see the Grand Canyon, no, ac- 
tually to the fifteenth. 
2.3 Consequences for automatic translation 
Summarizing the examples given above, a particle 
can have 
• a "literal" lexical translation, where semantics 
(truth conditions) is relevant (there can still be 
ambiguity between several sentential readings), 
* a "non-literal" lexical translation, where prag- 
matic intent is relevant (again, there can be am- 
biguity between several discourse readings), 
• a non-lexical translation, i.e., it is rendered by 
a syntactic or intonation feature, 
• a zero translation. 
For some particles, only a subset of these cases is 
possible, but others can, depending on context, fall 
into any of these groups, such as the notorious ja. 
Moreover, we have seen that a variety of knowl- 
edge sources is required to find the adequate trans- 
lation: 
• syntactic environment (e.g., for ja), 
. semantic types, 
• previous utterance (e.g., for doch), 
• preceding discourse (e.g., for noch), 
• world knowledge and/or inferences (e.g., for 
bezieh ungsweise) , 
• prosody (e.g., for bitte), 
• genre-specific conventions (see example below 
in section 4). 
Given this situation, it is unrealistic to expect that 
a set of standard transfer rules is sufficient to pro- 
vide adequate translations of particles on a lexical 
basis. Instead, a "deeper" analysis is required, which 
extracts sufficient contextual information and also, 
if necessary, can bring world knowledge into play. 
Similarly, on the side of the target language, the 
best realization corresponding to a source language 
particle is not always straightforward to determine. 
This is another argument against relying solely on 
transfer rules, especially if the realization is an into- 
nation feature or syntactic tag whose position is not 
immediately related to that of the source language 
particle. 
3 Discourse functions 
In VERBMOBIL, the "deep" analysis is undertaken 
in the Context Evaluation (ConEval) module, which 
constructs a conceptual representation (based on a 
domain model coded in a description logic language) 
from the output of the syntactic/semantic analysis 
module. The conceptual representation is used to 
compute the dialogue act \[Schmitz, Quantz 1995\] 
and to perform various disambiguations \[Stede et al. 
1996\], which can involve inferences. For example, we 
reason with date expressions to determine whether 
one date is a specification of another, or a separate 
one (which is sometimes important for disambigua- 
tion). 
In our conceptual representation, the discourse 
particles (in their "pragmatic reading") are repre- 
sented by labels signifying their discourse function. 
Figure 1 shows our current taxonomy of discourse 
functions (an extension of the one given in \[Schmitz, 
Fischer 1995\]), which has resulted from extensive 
corpus analyses of the VERBMOBIL data and from 
examining the relevant research literature. In the 
following, we briefly explain the various functions. 
Structuring These functions have received the 
most attention in the research literature. PUSH 
and POP mark the beginning of a sub-topic or di- 
gression, and the return to the previous topic, re- 
spectively. (Examples: by the way-anyway / davon 
abgesehen-wie auch immer). With an UPTAKE, 
the speaker signals a turn-taking at the beginning of 
a turn and a turn-holding within a turn. It also can 
help the hearer adapting to the acoustic properties of 
the speaker's utterance without losing information. 
(Examples: all right, now/ja, also). CHECK is a 
turn yielding signal, prompting the dialogue partner 
to respond. By using a CHECK, the speaker often 
seeks approving feedback from the hearer. (Exam- 
ple: isn't it? / oder?). REPAIR indicates problems 
in planning and performing the output, signals a new 
start, and thereby is also a turn-holding signal. (Ex- 
amples: I mean, sorry / ach nein, £h). 
Coherence Marking Some particles can be em- 
ployed to facilitate the embedding of the utterance 
within the context, and to check the common basis 
of the participants. (Example: doch, schon). 
Attitudinal We borrowed this label from All- 
wood et al. \[1992\]. While English often uses verbs 
for these purposes, German also offers a range of 
particles for speakers to convey a POSITIVE (ex- 
ample: gem), NEGATIVE (example: leider), or IN- 
DIFFERENT (example: ruhig) attitude towards the 
propositional content in their utterance, or towards 
the last utterance of the dialogue partner. In addi- 
tion, the propositional content of the speaker's own 
utterance can be marked as presupposed, or GIVEN 
(examples: ja, doch). And, some particles indicate 
SURPRISE at an utterance made by the partner 
(example: oh). 
Smoothening Also especially in German, parti- 
cles often help to create an overall appropriate into- 
nation contour, and at the same time can serve to ex- 
press cooperativity and politeness (Examples: denn, 
doch). Specifically, FILLERS allow the speaker to 
plan the output, avoid undue pauses, and help to 
hold the turn. (Examples: ich wiirde sagen, £h). 
4 Routine formulas 
We pointed out that the particles investigated here 
have at least one reading in which the discourse us- 
age is central to their usage, and not semantic con- 
tribution to propositional content. This difference 
points to the notion of "idiomatic" meaning, and 
-- not surprisingly -- the discourse functions intro- 
duced above can often also be realized by idiomatic 
phrases. Without going into detail here, we merely 
give a few examples, again taken from the VERBMO- 
BIL domain. In all these cases (and many others), 
the "literal" compositional meaning is not the point 
of using the phrase, and they typically cannot be 
translated word-by-word. 
As fillers, we often find phrases like lch wiirde 
denken .... or Ich muff sagen, ... In English, the 
translation I must say, ... is not wrong but not con- 
ventionally used in this context. Similarly, the Ger- 
man Wenn ich da real nachsehe, ... should not be 
translated preserving the conditionality, hence If I 
look this up, ... but by the common phrase Let me 
see, ... 
The function check can be realized by phrases 
like Sehe ich das richtig? which also should not be 
translated literally (Do I see that correctly?) but by 
a conventional phrase such as Am I right? 
repair markers can also be phrasal, as in X, oder 
besser gesagt, Y or in X, nein, ich wollte sagen Y. 
Again, literal translations should give way to conven- 
tionalized English formulas, hence X, no, I wanted 
to say Y is less felicitous than X, no, I meant Y. 
5 Towards automatic translation 
Since the problems associated with discourse parti- 
cles are largely absent when processing written lan- 
guage, computational linguistics has for most of its 
history not dealt with these problems. In SLT, how- 
ever, they cannot be avoided, especially when work- 
ing with a language rich in particles, such as Ger- 
man. Given the youth of the field, plus the fact 
that particles at first sight do not exactly seem to be 
the most important challenge for translating spoken 
language, it comes as no surprise that there are no 
satisfactory solutions in implemented systems yet. 
In the VERBMOBIL prototype that was completed 
last year, a number of particles are considered am- 
biguous between scopal/modal/focusing adverb on 
the one hand, and "pragmatic adverb" on the other. 
This class of "pragmatic adverbs" loosely corre- 
sponds to the "discourse usage" we have investigated 
above. The translation framework of VERBMOBIL 
is strongly lexeme-based; thus, for any particle in 
the German source-utterance, the transfer compo- 
nent seeks a corresponding English word on the basis 
of the reading determined. Typically, the ConEval 
module is asked to determine the class of a parti- 
cle, wherupon transfer chooses a target word. As an 
exception, in some contexts a pragmatic adverb is 
suppressed in the translation. 
This procedure is a start, but it cannot deal 
with all the facets of meaning found in discourse 
particles, as outlined above. On the basis of cor- 
pus studies, both \[Schmitz, Fischer 1995\] and \[Rip- 
plinger, Alexandersson 1996\] already demonstrated 
that many German particles have a whole range 
of English correspondents, of which VER.BMOBIL at 
present manages only very few. 
To improve the translations, for the second phase 
7 
of the VERBMOBIL project we propose to build upon 
the framework of discourse functions. The pur- 
pose of computing discourse functions in analysis 
is twofold: it supports disambiguation (not only of 
the discourse particles, but also of the surrounding 
words) and computation of the dialogue act under- 
lying the utterance; and it helps in segmentation, 
i.e., breaking an utterance into portions that serve 
as complete units for further processing. In trans- 
lation, the information on discourse function is im- 
portant for deciding whether to translate a particle 
at all, and how to do that: by inserting a corre- 
sponding target language particle, or by modifying 
the syntactic structure or intonation contour of the 
target utterance. 
Given the wide variety of information required 
for determining discourse functions (listed in sec- 
tion 2.3), the task is best performed in tandem with 
building up the conceptual representation of the ut- 
terance, i.e., in the ConEval module. The decision as 
to what discourse function to associate with a par- 
ticle is seldom a strict one (not even for the human 
analyst). Instead, the different clues from syntax, 
semantics, prosodb;i and world knowledge are typ- 
ically weak and have to be weighted against each 
other in the light of the complete utterance. There- 
fore, we tackle the problem with the same mech- 
anism we use for identifying dialogue acts: a set 
of weighted default rules, implemented in FLEX 
\[Quantz et al. 1996\] as an extension to the standard 
description logic language. The rules are matched 
against the utterance representation, and the accu- 
mulated weights decide on the most likely discourse 
function. We are currently in the process of defining 
this rule set. 
The result will be more fine-grained information 
on discourse particles than is available now in the 
system. The transfer and generation modules can 
use the discourse function to decide whether a lexi- 
cal correspondent should be produced in the target 
language, and if so, which one, and at what position 
of the utterance. Particles that are mere fillers can 
be removed entirely from the translation, and simi- 
larly those particles that are used to smooth the in- 
tonation contour in German. Whether restarts and 
self-repairs get translated or are merged into a sin- 
gle coherent utterance, is an open question. In many 
cases, it would not be difficult to replace the "cor- 
rected" portion of an utterance with the portion that 
"overwrites" it, thereby sparing the hearer from re- 
working the correction herself. 
As for routine formulas, they first of all cause the 
standard problems of idiomatic phrases: they need 
to be recognized as a single unit of meaning, so that 
they can be translated en bloc. This presupposes 
lexical representations that adequately describe the 
possible variants of the expression, e.g., whether ad- 
ditional modifiers may be inserted into a phrase, etc. 
When processing written language, this is difficult 
enough -- with speech and the additional uncer- 
tainties of word recognition, the problems are even 
harder. For the time being, a comprehensive treat- 
ment of routine formulas and other idioms does not 
seem feasible. 
Regarding the overall system architecture, the 
deep-analysis phase, as we have described it, need 
not be necessary for each and every utterance -- if 
the input allows for a standard transfer-based trans- 
lation (e.g., because it doesn't contain ambiguous 
particles), that will typically be sufficient. This es- 
sentially amounts to a mixed-depth analysis in the 
translation process -- an important question that 
we cannot discuss further here. 
6 Summary 
Discourse particles and routine formulas in spoken 
utterances cannot be translated on a simple lexeme- 
to-lexeme basis. We have proposed a taxonomy of 
discourse functions to represent the pragmatic im- 
pact of such particles and formulas. There is, still, 
no 1:1 mapping between particles/formulas and dis- 
course functions in analysis, nor between discourse 
functions and their realization in the target lan- 
guage. Therefore, we use a "deep" utterance rep- 
resentation of dialogue act and propositional con- 
tent, into which discourse functions are integrated. 
In analysis, the deep representation holds all the 
information required for successful processing; the 
transfer and generation components can then decide 
whether discourse functions get realized in the target 
language, and if so, by what means. This decision 
can be made in the context of the target-language 
utterance. 

References 
W. Abraham (ed.). Discourse particles. Amsterdam: 
John Benjamins, 1991. 
J. Allwood, J. Nivre, E. Ahlsen. "On the semantics 
and pragmatics of linguistic feedback." In: Jour- 
nal off Semantics 9:1-26, 1992. 
J. Bos, A. Batliner, R. Kompe. "On the use 
of Prosody for Semantic Disambiguation in 
VERBMOBIL". Verbmobil Memo 82, DFKI 
Saarbrficken, 1995. 
D. Byron, P. Heeman. "Discourse marker use in task- 
oriented spoken dialog." In: Proceedings off the yh 
European Conference on Speech Communication 
and Technology (Eurospeech), Rhodes/Greece, 
1997. 
J. Hirschberg, D. Litman. "Empirical studies on the 
disambiguation of cue phrases." In: Computa- 
tional Linguistics 19 (3), 501-530, 1993. 
J.J. Quantz, G. Dunker, F. Bergmann, I. Kellner. 
The FLEX System. KIT Report 124, FB Infor- 
matik, TU Berlin, 1996. 
R. Reichman. Getting computers to talk like you and 
me. MIT Press/Bradford Books, 1985. 
B. Ripplinger, J. Alexandersson. "Disambigua- 
tion and translation of German particles in 
Verbmobil." Verbmobil Memo 70, IAI/DFKI 
Saarbriicken, 1996. 
D. Schiffrin. Discourse markers. Cambridge Univer- 
sity Press, 1987. 
B. Schmitz, K. Fischer. "Pragmatisches Beschrei- 
bungsinventar fiir Diskurspartikeln und Rou- 
tineformeln anhand der Demonstratorwortliste." 
Verbmobil Memo 75, TU Berlin/Universit~it 
Bielefeld, 1995. 
B. Schmitz and J.J. Quantz. "Dialogue-Act Type 
Recognition in Automatic Dialogue Interpret- 
ing." In: Proceedings of the The Sixth Interna- 
tional Conference on Theoretical and Methodologi- 
cal Issues in Machine Translation (TMI), Leuven, 
1995. 
M. Stede, U. Kfissner, J. Quantz, B. Schmitz. "Dis- 
ambiguation as part of semantic evaluation in 
Verbmobil." In: C. Boitet (ed.): Proceedings of 
the Workshop on Multimodal Interactive Disam- 
biguation (MIDDIM-96), Grenoble, 1996. 
W. Wahlster. "Verbmobil: Translation of face-to- 
face dialogues." In: Proceedings of the Third Eu- 
ropean Conference on Speech Communication and 
Technology, Berlin, 1993. 
