Abstract 
This paper describes ongoing work on the choice 
of modal expressions in the generation of recom-
mendations about courses of study within the B.A.-
program of the Ruhr-Universität Bochum. We fo-
cus our work on the German modal verbs müssen 
(must), können (can), the subjunctive form sollte 
(should), and dürfen (may). The idea is to combine 
insights from formal semantics into the meaning of 
modal verbs with requirements for their choice in 
NLG-systems in order to achieve a linguistically 
satisfying model of their choice. The overall model 
is implemented in the CAN-system that plans 
courses of study and generates recommendations 
about them.   
1 Introduction 
From a typological perspective, modality concerns the 
speaker’s estimation of the possible and necessary occur-
rence, respectively, of some state of affairs [van der Auwera 
and Plungian, 1998]. In many languages, modality is lin-
guistically realized by a whole bundle of different means, 
with modal verbs being the most prominent ones.  
In this paper, we will describe the realization of uses of the 
four German modal verbs müssen (must, have to, need to), 
können (can, to be able to), dürfen (may, could, to be al-
lowed to do) and the subjunctive form of sollen, i.e. sollte 
(should). As the English translations indicate, there is no 
simple correspondence between the meaning of these Ger-
man modal verbs and their counterparts in other languages. 
In particular, the pragmatic conditions for their use differ 
widely so that a straightforward adoption of the conditions 
we are giving in this paper to other languages will probably 
not be possible. 
In general, the problem of the choice of modal verbs is the 
fact that the use of modals is, among other things, deter-
mined by information that is usually only implicitly given in 
a discourse. This information must be made available during 
macroplanning for the entire text in a way that enables the 
choice of the contextually appropriate modal verbs.  
In what follows, we will first outline the semantics and 
pragmatics of modal verbs in general, and the four German 
modals plus sollen in particular. Based on this, we will de-
lineate our choice model for modals and its application to 
the aforementioned four modal verbs. Finally, we will show 
how this model is integrated into the CAN-System that we 
are currently developing.  
2 The meanings of modal verbs 
The meaning of modals is deeply rooted in information that 
is only implicitly given in a discourse, but plays a crucial 
role in the full understanding of the underlying speaker’s 
intention. This context-dependence results in different read-
ings. A sentence like Peter must attend the lecture on se-
mantics can be interpreted in numerous ways: It could ex-
press that according to the knowledge of the speaker it is 
necessary that Peter attends this lecture, or that Peter´s 
course of study requires his attendance in the lecture on se-
mantics.  
Since Kratzer´s (1977) seminal work on a uniform and for-
mal treatment of the semantics of modal verbs, the context-
dependence of the meaning of modals is traced back to so-
called conversational backgrounds (also called modal do-
mains or modal bases) that provide that information against 
which the modalized sentence has to be interpreted. Accord-
ing to her widely accepted analysis in formal semantics, the 
meaning of modal verbs depends on the class of permitted 
conversational backgrounds and the so-called modal force, 
i.e. whether the modal verbs express possibility or necessity. 
Possibility and necessity are defined as consistency and en-
tailment with respect to the respective conversational back-
ground: a proposition is possible if it is consistent with the 
relevant background, and it is necessary if it is entailed by 
the conversational background.  
Accordingly, modal verbs differ in their semantics with re-
spect to these two parameters: their inherent modal force, 
and the classes of modal domains associated with them.
1
 
                                                 
1
 In more recent works, Kratzer`s (1977) original work has 
been extended to a doubly relative analysis of modal verbs: Their 
meaning does not only depend on one modal domain, but addition-
ally on a second one, the “ordering source”. An ordering source is 
a set of propositions that describes the plausibility of the states 
given in the primary modal domain. Since this refinement of 
Kratzer’s work does not concern our treatment of the choice of 
When must should be chosen 
Ralf Klabunde 
Ruhr-Universität Bochum 
Department of Linguistics 
D- 44780 Bochum 
klabunde@linguistics.rub.de 
The classifications of relevant conversational backgrounds 
that have been postulated for an adequate analysis of the 
meaning of modal expressions differ widely in the literature. 
They range from two types – a basic distinction between 
epistemic and “root” modals - to more than twenty in von 
Stechow’s (2004) classification.  
Kratzer uses a modal logic to describe the meaning of modal 
verbs. However, she leaves open what conversational back-
grounds really are (besides their characterization as sets of 
propositions), although this concept is the linchpin of a sat-
isfying analysis of the meaning of modals.  
Also discourse-oriented approaches to the meaning of mo-
dals as, e.g., Roberts (1989), Frank (1996), or Stone (1999) 
do not say much about the structure and content of conver-
sational backgrounds. These approaches analyze specific 
uses of modals as anaphoric expressions: in some discourse 
segments, modals pick up modal domains introduced by 
previous modal expressions, and must be interpreted against 
this already introduced hypothetical scenario. Stone (1999), 
for example, gives the following sentence pair:  
 
(1) A wolf might walk into the house. It would eat you. 
 
In order to interpret the second modalized sentence cor-
rectly, the anaphoric connection between the content of that 
sentence and the possible scenario of the first one must be 
made explicit.  
 
2.1 The meaning of müssen, dürfen, sollen, sollte, 
and können 
We will outline what semanticians have to say about the 
meaning of these modal verbs, and characterize those prag-
matic conditions for their use that are of importance for this 
paper.  
Müssen and sollen both express necessity as modal force. 
The semantic difference between these two verbs concerns 
the admissible conversational backgrounds: While müssen 
can be used with any conversational background, the use of 
sollen is restricted to specific ones, viz. those that express a 
specified ideal like laws or social norms. For example, while 
the sentence  
 
(2) Peter muss den Kurs besuchen  
(Peter must attend the lecture)  
 
can, in principle, be interpreted with respect to any kind of 
conversational background, some of these backgrounds are 
not possible for the sentence Peter soll den Kurs besuchen 
(Peter shall attend the lecture). The former sentence can be 
interpreted as: 
1. In view of what is known by the speaker/the lis-
tener/…, it is necessary that Peter attends the lec-
ture (epistemic interpretation) 
                                                                                  
modals in this paper, we will not go into the details. However, we 
will pick up that point in the depiction of our future work. 
2. In view of what the conditions of study determines, 
it is necessary that Peter attends the lecture (deon-
tic interpretation) 
3. in view of someone´s will, it is necessary that Peter 
attends the lecture (volitional interpretation) 
4. in view of achieving certain goals in is necessary 
that Peter attends the lecture (teleological interpre-
tation)  
   and so on.  
 
Sollen, however, cannot presuppose conversational back-
grounds that are evoked by the subject (cf, von Stechow, 
2004; Zifonun et al., 1997). For example, a deontic interpre-
tation of Peter soll den Kurs besuchen is not possible with 
the respective obligations as conversational backgrounds 
determined by Peter himself.  
In addition to the modal verbs, mood is also a linguistic 
means to express modality. Often, mood weakens the modal 
force so that sollte does not express the necessity of the em-
bedded proposition anymore, but a weaker modal force that 
comes close to possibility.  
Können and dürfen seem to be the counterparts of müssen 
and sollen. They express possibility as modal force. Können 
is not confined to specific conversational backgrounds, 
while dürfen can – similar to sollen - only be used with ex-
tra-subjective ideal backgrounds, preferably deontic ones. 
The sentence 
 
(3) Peter darf den Kurs besuchen  
(Peter may attend the lecture) 
 
means: his attendance of the lecture is compatible with his 
duties (imposed by someone else), the conditions of study, 
the will of a third person, some relevant facts, and so on.  
Although this semantic analysis shows convincingly the 
relevance of the modal force and the conversational back-
grounds for a systematic treatment of the meaning of mo-
dals, it cannot be the whole story of the meaning of modals. 
If conversational backgrounds and modal forces would suf-
ficiently explain the meaning of modals, every modal verb 
with a restricted set of admissible conversational back-
grounds could be replaced by the modals können and 
müssen, respectively, without changing the meaning of the 
respective sentence. Although this might be true with re-
spect to the truth conditions of the single sentences, replac-
ing the specific modals by more general ones nevertheless 
results in sentences with different readings. For example, 
replacing the sentence  
 
(4) Du darfst den Semantik-Kurs besuchen und du solltest 
 den Syntax-Kurs besuchen (You may attend the  
semantics lecture and you should visit the lecture on  
syntax) 
 
by the sentence 
 
(4´)  Du kannst den Semantik-Kurs besuchen und du  
müsstest den Syntax-Kurs besuchen (You can attend 
the semantics lecture and you ought to attend the  
lecture on syntax) 
 
leads to different prominent meanings: while the first sen-
tence clearly expresses a permission and a recommendation, 
the second one is more neutral with respect to the relevant 
pragmatic meaning. Additionally, the second clause ex-
presses a weak deontically motivated entailment.  
Hence, the modal verbs are associated with specific prag-
matic conditions for their use. We will refer to those prag-
matics aspects only that are crucial for our treatment of the 
modals. We start from the assumption that the uses of 
müssen and können do not require specific pragmatic con-
straints. Sollte does not express necessity (as the indicative 
sollen does) but possibility, and shall be used if the goal of 
the system is to express the proposition as a strong advice. 
Dürfen shall be used if the system expresses a permission 
that is based on the courses the user has completed success-
fully so far, or if the user has the permission to study at all.  
2 Modality in natural language generation 
systems  
In general, there is not much work on the computational 
treatment of modals, be it in NL Understanding or NL Gen-
eration.  
There seems to be a clear gap between the relevance of mo-
dality in texts and dialogues produced by human speakers 
and its consideration in NLG-systems. While modal expres-
sions appear in all kinds of text genres produced by humans, 
be it newspaper articles or dialogue contributions, their re-
alization in NLG-systems did not receive much attention so 
far.
2
 This disregarding of modality might be due to the na-
ture of the text types that typically describe entities as, e.g., 
animals (Peba-II; Milosavljevic 1997) or air–qualities 
(TEMSIS, Busemann and Horacek 1997). All of these gen-
eration systems present data or knowledge units that are 
“true” in a certain sense. None of these systems have to 
cope with an estimation whether some events will happen 
with a specific probability, given the information state of the 
system and the user. However, estimations of such a kind 
are the basis for the use of modals.  
Our work is loosely inspired by the treatment of modals in 
Crouch and Pulman (1993), but differs from their work in 
several crucial respects. Crouch and Pulman investigate the 
use of temporal and modal expressions in an NL-interface to 
a planner that produces partial plans in an engineering do-
main. Instead of using a modal logic with its underlying 
possible world semantics, the authors determine possible 
and necessary propositions against the range of plans avail-
able to solve a task. The necessity relation holds if the 
proposition (more specifically, the temporal orders and rela-
                                                 
2
 This statement might be too strong in case of the use of sys-
temic grammars. Systemic grammars are able to convert complex 
modality-related speech-functions into corresponding modal ex-
pressions. However, systemic grammars do not consider content 
planning tasks that we are describing in this paper. 
tions in the corresponding formula) can be inferred from the 
plan description. In contrast to this, a formula is possible if 
its negation can be inferred from the plan description.  
3 The choice of modal verbs in CAN 
We are currently developing the CAN-system (Conceptuali-
zation for Modal Expressions) that generates recommenda-
tions about courses of study within the B.A. program of the 
Ruhr-Universität Bochum. A user provides the system with 
his/her current term number and the lectures he/she has fin-
ished so far, and the system generates a recommendation 
which lectures he should/must/shall/may/can take.  
A planner provides the partially ordered plan which de-
scribes the progression of the overall study from the current 
semester to the final degree. The propositions are checked in 
a piecemeal fashion in order to determine whether they nec-
essarily or possibly hold with respect to the conversational 
background. This procedure determines the modal force, 
and a subclass of the relevant modal verbs with it. After-
wards the pragmatic conditions will be checked to select the 
appropriate modal verb, and that verb is inserted into a syn-
tactic template for the corresponding sentence.   
We confine our analysis to epistemic and deontic meanings 
of the aforementioned modal verbs, because these readings 
are the most prominent ones in our domain. Deontic mean-
ings are determined by the structure and content of the plan 
to be verbalized. Epistemic meanings are determined by the 
discourse state. Both knowledge sources will be handled by 
the same procedure which makes an extension to other types 
of modal meanings manageable. In other words, we adopt 
the bipartite distinction between necessary and possible 
formulae, although this might not completely correspond to 
the fine-grained modal forces expressed in natural lan-
guages. Furthermore, we adopt the modeling of possibility 
as consistency with the conversational background, and ne-
cessity as entailment from that background.  
If one of the modal relations is true, we know what the mo-
dal force is, and are able to determine that subclass of modal 
verbs that expresses the proven modal force. Otherwise the 
underlying proposition will linguistically not be modalized. 
The deontic background is determined by the plan. Accord-
ing to the partially ordered plan, some lectures can only be 
attended if other lectures had already been completed suc-
cessfully, but this order does not hold between all lectures. 
The obligation to attend certain lectures during some semes-
ter can directly be read from the plan. The modal force will 
be derived from the plans. 
An epistemic background is given by what the system 
knows of the user´s current course of his study and the in-
formation provided by the system so far. For this, we use a 
discourse record that keeps track of the user´s input and 
what he has been told.  
3.1 Planning for the choice of modals  
The task of the planner is to place the obligatory lectures at 
the right nodes in the plan, and to determine possible nodes 
of the optional lectures. For example, while all students of 
Linguistics have to attend a lecture on pragmatics in the 
third semester, they can attend an additional lecture with a 
topic of their own choice, but the student must have satisfied 
its prerequisites already.  
An “action” of the plan is realized as a pair that represents 
the status of the lecture with respect to the semester (obliga-
tory, optional), and the title of the lecture.  
Suppose a fully specified plan contains the following order 
of single lectures, starting with the second semester:   
 
2
nd
 semester   3
rd
 semester      4
th
 semester 
 
       obl: pragmatics 
obl: formal_ 
semantics   obl: syntax      opt:parsing 
 
       opt: hpsg 
 
If each “action” of this part of the plan shall linguistically be 
realized, only some of the four modal verbs are applicable.  
The use of the modals depends on the status of the lectures 
as being obligatory or optional in a certain semester, and the 
existence of certain preconditions of the planning operators. 
The distinguishing feature of obligatory lectures is that there 
are no alternatives for them; they are necessary courses in a 
certain semester. Hence, the lectures on semantics and prag-
matics will be mentioned in sentences with müssen as modal 
verb. Optional lectures appear to be more interesting be-
cause in principle they can be attended in any semester, as 
long as the prerequisites for their attendance have been 
achieved by the student. The prerequisites are modelled as 
preconditions of corresponding planning operators. Linguis-
tically, optionality is expressed by modals which express 
possibility as modal force, and the use of these modals is 
constrained by additional pragmatic conditions for their use. 
Whether the optional lectures on HSPG and parsing, respec-
tively, will be mentioned in sentences with können, dürfen, 
or sollte depends on the decision to recommend or permit 
these lectures, which will be derived from conditions we 
will mention in the following section.  
To sum up, the choice of the appropriate modal verbs de-
pends on:  
- the status of the lecture, whether it is optional or 
obligatory in one semester, 
- the user´s possible achievement of prerequisites of 
optional lectures, 
- the placement of optional lectures in the overall 
plan, 
- pragmatic functions like expressing a permission or 
a recommendation. 
The first three factors belong to the conversational back-
ground, while the last one seems to be an additional con-
straint for the use of some of the modal verbs. Dealing with 
these aspects of the conversational background can be char-
acterized in the following way:  
Let P be the plan with the obligatory lectures placed in the 
right positions, and T(P) the set of all totally specified plans 
based on P that include possible placements of the optional 
lectures.  
Every lecture in P marked as obligatory is necessary with 
respect to the study process.  
Additionally, for every proposition p we check whether T(P) 
→ p holds or T(P) ∪ {p} is true. If the former formula is 
true, the modal verb müssen can be realized. If the latter 
formula is true, we have to check whether the pragmatic 
conditions for the use of one of the modals verbs that ex-
press possibility are given in the discourse record. The cor-
responding decision results in the choice of the verb dürfen, 
können, and sollte, respectively.  
3.1 An example: the choice of dürfen and sollte  
The choice of dürfen and sollte may illustrate the choice 
mechanism that takes into account the plan structure and the 
discourse state, respectively.  
Both dürfen and sollte express possibility as modal force so 
that they compete with können as the unmarked modal verb 
for expressing possibility. Since dürfen and sollte can only 
be used if specific pragmatic conditions are satisfied (and 
with restricted sets of conversational backgrounds), können 
will always be chosen if the modal force is possibility and 
the other two modals cannot be used. 
Expressing a permission by using dürfen is based on the 
following conditions:  
(a) if the user is willing to begin his studies and CAN 
does not know whether he is registered at the uni-
versity at all, CAN asks for a corresponding con-
firmation. In case of a positive response the per-
mission to study the courses of the first semester 
can be given, and dürfen will be used.  
(b) If CAN does not know whether some courses as 
preconditions of a course are satisfied it asks the 
user whether he attended these required courses. 
The user has to give this information, and the sys-
tem integrates the courses into the plan. Since the 
preconditions for the respective course are now sat-
isfied, the permission to attend this course will be 
signalled by the use of dürfen.   
In general, our impression is that permissions – the prag-
matic function of dürfen when a deontic conversational 
background becomes relevant – can best be derived in dia-
logue situations. An NLG-system that permits to attend a 
certain course must be set into the position to express the 
permission at all. In our scenario, this is the cased if CAN 
requires additional information and asks for it. In this re-
spect permissions are different from recommendations ex-
pressed by sollte, which can be determined without refer-
ence to a dialogue situation.  
Recommendations declare a certain state of affairs as advan-
tageous with respect to some goal. In our scenario, attending 
a course in a specific semester is advantageous for the com-
pletion of the course of study, if its attendance optimizes the 
number of semesters that are necessary for that completion.  
The system checks whether the course can successfully be 
integrated into the overall plan and computes the corre-
sponding consequences: If attending this course during the 
current semester does not increase the number of semesters 
required to achieve the completion within the normal dura-
tion of study, but a different course does, that is also offered 
for the current semester, then it is advantageous to attend 
this course, and the modal sollte is chosen. In other words, 
the system compares the consequences of attending courses 
with respect to the length of study.  
4 Future work 
The work reported in this paper describes the current state 
of the CAN-system. Our long-term goal is a wide coverage 
of modal expressions including all German modal verbs as 
well as other modals as, e.g., modal nouns, adverbs, and 
particles. The evaluation of the generated modal expressions 
is an additional research topic, as well as the interplay of 
modality with mood and temporality. 
Currently, our discourse record is a simplifying collection of 
propositions, but we will need a more sophisticated version 
of this knowledge store if modal subordination [Roberts, 
1989; Stone, 1999] becomes a topic of our research. Modal 
subordination becomes relevant in sentence pairs of the fol-
lowing kind: 
 
(5)  Du könntest den Semantik-Kurs besuchen. Dann musst 
du im folgenden Semester den Pragmatik-Kurs besu-
chen. (You could attend the semantics lecture. In this 
case you must attend the lecture on pragmatics in the 
next semester.) 
 
In order to choose müssen for the second sentence, the hy-
pothetical event introduced by the first sentence must be 
considered as background so that ordering sources will be-
come an additional factor for a discourse-sensitive choice of 
modals.  
Finally, the decision to modalize a proposition is often 
grounded in specific dialogue settings like the existence of 
social power relations between the participants, or their re-
spective beliefs of the beliefs states of the addressees. We 
assume that a dialogue situation explains the use of modals 
in a more natural way than the current model, as our current 
treatment of permissions already indicates. Therefore, in the 
long run CAN shall be able to react dynamically to user re-
quests. 

References 
[Busemann and Horacek, 1997] S. Busemann and H. 
Horacek. Generating air quality reports from environ-
mental data. Proceedings of the DFKI Workshop on 
NLG, pages 15-21, 1997.  
[Crouch and Pulman, 1993] R.S. Crouch and S.G. Pulman. 
Time and modality in a natural language interface to a 
planning system. Artificial Intelligence, 63: 265-304, 
1993. 
[Frank, 1996] A. Frank. Context dependence in modal con-
structions. Doctoral Dissertation, University of Stuttgart, 
1996. 
[Kratzer, 1977] A. Kratzer. What „must“ and „can“ must 
and can mean. Linguistics and Philosophy, 1:337-355, 
1977. 
[Milosavljevic, 1997] M. Milosavljevic. Content selection 
in comparison generation. Proceedings of the 6th Euro-
pean Workshop on Natural Language Generation, pages 
72-81. Duisburg, Germany, 1997. 
[Roberts, 1989] C. Roberts. Modal subordination and pro-
nominal anaphora in discourse. Linguistics and Philoso-
phy, 12: 683-721, 1989. 
[Stone, 1999] M. Stone. Reference to possible worlds. 
Technical report, Rutgers. Available at: 
http://www.cs.rutgers.edu/~mdstone/ 
[van der Auwera and Plungian, 1998] Modality´s semantic 
map. Linguistic Typology, 2: 79-124, 1998. 
[von Stechow, 2004] A. von Stechow. Schritte zur Satzse-
mantik. Ms., University of Tübingen. Available at: 
http://vivaldi.sfs.nphil.uni-tuebingen.de/ 
~arnim10/Lehre/Japan2004/Schritte1_17.pdf 
[Zifonun et al., 1997] G. Zifonun, L. Hoffmann and B. Stre-
cker (eds.) Grammatik der Deutschen Sprache, Vol. III; 
Chapter F5: Bedeutung der Modalverben. Berlin: Walter 
de Gruyter, 1997. 
