Proceedings of the Analyzing Conversations in Text and Speech (ACTS) Workshop at HLT-NAACL 2006, pages 58–61,
New York City, New York, June 2006. c©2006 Association for Computational Linguistics
Pragmatic Discourse Representation Theory
Yafa Al-Raheb
National Centre for Language Technology
Dublin City University
Ireland
yafa.alraheb@gmail.com
Abstract
This paper presents a pragmatic approach to Dis-
course Representation Theory (DRT) in an attempt
to address the pragmatic limitations of DRT (Werth
1999; Simons 2003). To achieve a more prag-
matic DRT model, this paper extends standard DRT
framework to incorporate more pragmatic elements
such as representing agents’ cognitive states and the
complex process through which agents recognize
utterances employing the linguistic content in form-
ing mental representations of other agent’s cogni-
tive states. The paper gives focus to the usually
ignored link in DRT literature between speaker be-
liefs and the linguistic content, and between the lin-
guistic content and hearer’s beliefs.
1 Introduction
Developments in dynamic semantics, resulting in
DRT, have led to a framework suitable for the rep-
resentation of linguistic phenomena (van Eijck and
Kamp 1997). This is specifically due to the fact that,
recognizing the importance of context, DRT concen-
trates on updating the context with the processing of
each utterance. In addition, DRT can also be viewed
as an agent’s mental model of the world and not just
a representation of the discourse. It is for these rea-
sons that DRT holds great potential for incorporating
more pragmatic phenomena.
However, despite the suitability of DRT for repre-
senting linguistic phenomena, some pragmatic lim-
itations have been noted in the literature. Simons
(2003) remarks that DRT is a theory of seman-
tics and not pragmatics. Werth remarks that ‘there
is no place in [DRT] for participant roles, setting,
background knowledge, purposes, even inferences’
(Werth 1999: 65). In general terms, we can say that
the pragmatic dimension supplements semantic con-
tent by using context and cognitive states of agents
in dialogue. The discipline of pragmatics is, there-
fore, concerned with the process by which agents
infer information about elements of another agents’
cognitive state such as their beliefs and intentions.
Thus, this paper focuses on extending standard DRT
pragmatically to model agents’ cognitive states in
the pragmatic context of dialogue.
2 A More Pragmatic DRT
This section presents a more pragmatic DRT focus-
ing on the relationship between speaker generation
and the linguistic content, and between the linguistic
content and hearer recognition. Figure 1 represents
the link between our representation of the speaker’s
cognitive state, the speaker’s linguistic content and
the hearer’s cognitive state or DRS (Discourse Rep-
resentation Structure). This relationship has not to
our knowledge been explored in the literature and
deserves investigation.
Generally speaking, to generate an utterance,
there would be some discrepancy between the
speaker’s beliefs and the speaker’s beliefs about the
hearer’s beliefs. The discrepancy leads to an utter-
ance, i.e. linguistic content. The linguistic content
is the window the hearer has onto the speaker’s state
of mind. It is what influences hearer recognition.
By analysis of the linguistic content provided by the
speaker, the hearer can propose a hypothesis regard-
ing the speaker’s state of mind.
58
Speaker Generation:
i you
attitude(i, ‘BEL’, drs1)
drs1:
attitude(i, ‘INT’, drs2)
drs2:
→ Linguistic contentSpeaker’s utterance → Hearer Recognition:
i you
attitude(i, ‘BEL’, drs3)
drs3:
attitude(you, ‘INT’, drs4)
drs4:
Figure 1: Speaker DRS, Linguistic Content and Hearer DRS
2.1 New DR-Structures
The DRT representation introduced here extends
standard DRT language and structure resulting in a
suitable pragmatic-based framework for represent-
ing this pragmatic link. Separate DRSs are created
to represent each agent. DRSs get updated with each
new utterance. Each DRS representing an agent’s
cognitive state includes the two personal reference
markers ‘i’ and ‘you’. When ‘i’ is used in a DRS, it
refers to the agent’s self within that DRS; i.e. if the
agent is the speaker, then ‘i’ refers to the speaker in
the entire DRS. To refer to the other agent, ‘you’ is
used. To follow from the speaker’s example, ‘you’ in
this case refers to the hearer. To account for agents’
cognitive states and their meta-beliefs, a sub-DRS
representing the agent’s cognitive state called the be-
lief DRS is created to include the speaker’s beliefs
about the hearer’s beliefs. Additionally, a new DRS
for representing weaker beliefs called acceptance is
introduced. The same level of embedding offered
to belief DRSs is introduced in acceptance DRSs.
Acceptance DRS includes the speaker’s acceptance
DRS as well as what the speaker takes the hearer to
accept. Provided the speaker has sufficient informa-
tion, the speaker can also have the embedded DRS
within the acceptance DRS that represents what the
hearer takes the speaker to accept.
In addition to expanding the belief DRS, each
agent’s cognitive state contains an intention DRS. In-
tention in the sense used here refers to the agent’s
goals in making an utterance, which are represented
by the corresponding dialogue act marked in the
intention DRS. The hearer’s intention DRS repre-
sents the recognized utterance and contains elements
of utterance-making generally associated with prag-
matics such as the function of an utterance, its dia-
logue act. This pragmatic enriching strengthens the
link between an agent’s intentions and the linguistic
form uttered. What is proposed is that the intention
DRS be designed to include the linguistic content
provided within utterances.
To further enhance the link between agents’ cog-
nitive states and the linguistic content of their ut-
terances, the intention DRS contains the rich prag-
matic information offered by explicitly marking the
presupposition (given information) and the assertion
(new information) of the current utterance. The in-
tention DRS is a separate DRS from the belief DRS.
The beliefs of an agent give the motivation for mak-
ing an utterance, and the intention DRS represents
the speaker’s intended message. The recognition
of an utterance gives the hearer an insight into the
agent’s beliefs. Depending upon the particular dia-
logue represented, the intention DRS could have the
speaker’s intention, the hearer’s intentions or both.
The intention DRS functions as the immediate con-
text, the one containing the utterance being gener-
ated or recognized. The belief and acceptance DRSs
function as background context containing informa-
tion pertaining to the dialogue and not just the cur-
rent utterance. This division of labour context-wise
is useful in that the information represented in the
intention DRS directly feeds into the speaker’s ut-
terance, and is then inferred by the hearer through
the linguistic content. The hearer’s intention DRS
includes the inferred speaker intentions in uttering
the current utterance. This gives the flexibility of
being able to model information that the hearer has
inferred but has not yet decided to accept or believe
and is, therefore, not yet included in either the belief
or acceptance DRS. For instance, while the hearer
in example (1) has recognized S1’s utterance, he has
not yet accepted S1’s utterance. This motivates sep-
arating the representation of beliefs from intentions.
59
(1) S1: Bob’s trophy wife is cheating on him.
H1: When did Bob get married?
2.2 Extending DRT Language
In addition to the three DRSs introduced above, in
order to make the link between speaker generation,
linguistic content, and hearer recognition more ex-
plicit, labels, ‘labeln’, n an integer, are introduced.
The labels mark the distinction between presupposi-
tion and assertion, and the distinction between weak
and strong beliefs. Furthermore, the labels can be
used to refer to a particular predicate by another
complex predicate. The labels increase the expres-
sive power from an essentially first-order formal-
ism to a higher-order formalism. Presuppositions
are marked by a presupposition label ‘pn’. Simi-
larly, DRSs inside the main speaker or hearer DRS
are labeled ‘drsn’. Assertions are marked by ‘an’
to strengthen the connections between the linguistic
form (in the separation between presupposition and
assertion) and the representation of beliefs. Believed
information labeled ‘bn’ inside a belief DRS or ac-
cepted information labeled ‘cn’ inside an acceptance
DRS can be either presupposed or asserted inside the
intention DRS. Thus, the labels in the intention DRS
can only be ‘p’ or ‘a’.
Conditions referring to attitudes (acceptance, be-
liefs, and intentions) have been added to the ex-
tended semantics of DRT. Figure 2 shows three em-
bedded DRSs, acceptance DRS, drs2, belief DRS,
drs4, and intention DRS, drs6 representing:
(2) A: Tom is buying Mary a puppy.
B: That’s sweet.
DRSs are referred to by the attitude describing them.
For example, attitude(i,‘BEL’, drs4) refers to the
DRS containing the speaker’s beliefs, using the la-
bel for the belief DRS, drs4. Other conditions
are allowed to employ ‘i’ as an argument. Atti-
tude(i,‘accept’, drs2) refers to the DRS containing
the speaker’s acceptance DRS, using the label for
the acceptance DRS, drs2. Attitude(i,‘INT’, drs6)
refers to the DRS containing the speaker’s intention
in uttering example (2), using the label for the inten-
tion DRS, drs6. The speaker’s acceptance DRS con-
tains an embedded DRS for the hearer’s acceptance
DRS, drs2. In this case, it is empty, as no weakly be-
lieved propositions have been introduced yet. Simi-
larly, the belief DRS contains space for the speaker’s
beliefs about the hearer’s beliefs, drs5. The intention
DRS contains the linguistic content of the utterance
that the speaker is about to make, drs6, as well as the
relevant dialogue acts.
drs1:
i you t m
drs2:
attitude(you, ‘ACCEPT’, drs3)
drs3:
attitude(i, ‘ACCEPT’, drs2)
attitude(i, ‘BEL’, drs4)
drs4:
p
b1: tom(t)
b2: mary(m)
b3: puppy(p)
b4: buy(t,m,p)
attitude(you, ‘BEL’, drs5)
drs5: b5: tom(t)b6: mary(m)
attitude(i, ‘INT’, drs6)
drs6:
p
p1: tom(t)
p2: mary(m)
a1: puppy(p)
a2: buy(t,m,a1) inform(i,you,a2)
Figure 2: A’s initial Cognitive State
In Figure 2, there are essentially three levels of
embedding in a main DRS. If we look at the belief
DRS, the first embedded DRS is the agent’s own be-
lief DRS. Level two is the agent’s beliefs about the
other agent’s beliefs DRS. Level three is inserted
when necessary and represents the agent’s beliefs
about the other agent’s beliefs about the agent’s be-
liefs DRS. DRSs of the same level of embedding
have similar status. For example, the agent’s accep-
tance and belief DRSs have equal status. However,
the only discourse referents in common are the ones
in the main DRS’s universe. Each equal-level em-
bedding has its own set of discourse referents, as
well as its own conditions.
Discourse referents of same and higher levels of
embedding are accessible to lower levels of embed-
ding and are therefore not represented in the lower
level embedding universe. This does not entail that
when a lower level embedding makes use of a dis-
course referent introduced in a higher level embed-
ding the agent and other agent share the same inter-
nal or external anchors. For example, when talking
about a rabbit, the speaker’s representation of rabbit
60
will be: b1:rabbit(x), whereas the speaker’s repre-
sentation of the hearer’s beliefs will be b2:rabbit(x).
This is to replace Kamp and Reyle’s (1993) use of
different discourse referents, where a new discourse
referent is used every time the same object or in-
dividual is referred to in a new sentence (e.g. rab-
bit(x), then rabbit(y)). The aim is to avoid having
to overtly use the x=y rule every time the same rab-
bit is referred to. The principles behind the equation
predicate are still in place; i.e. every time rabbit is
referred to, it is bound to the rabbit already in the
context. However, we bind it to the previous proper-
ties of rabbit already in context through attaching it
to the same discourse referent, rabbit(x).
Both Kamp and Reyle’s and our representation
face revision when it transpires that the agents in
dialogue have different referents in mind. For ex-
ample, both the speaker and hearer might be talking
about ‘rabbit’. However, they might have a differ-
ent ‘rabbit’ in mind, and assume the other partici-
pant is thinking of the rabbit they have in mind. The
speaker might have a grey rabbit in mind, whereas
the hearer has a white rabbit in mind. In this case,
Kamp and Reyle’s revision would consist of deleting
x=y predicate, and any previous equation predicate
that may have been introduced each time rabbit was
referred to. In our representation, the revision takes
place by changing the other agent’s discourse refer-
ent, b2:rabbit(x) becomes label2:rabbit(y).
Furthermore, the previous pragmatic extensions
to standard DRT have been implemented computa-
tionally to approximate a computational model of
communication and to enable us to see whether the
extended DRT works logically. The implementation
relates the linguistic content of utterances to the be-
liefs and intentions of the agents. The implementa-
tion operates with a specific dialogue, which can be
modified, within a restricted domain. It seems rea-
sonable to conclude on the basis of the implementa-
tion that the conceptual and formal proposals made
provide a basis for further development.
3 Conclusion and Further Extensions
This paper pushes the treatment of linguistic
phenomena in DRT more towards pragmatics,
by bringing more pragmatic elements to the
semantic/pragmatic interface which is DRT. It has
been the aim of this paper to achieve this by (a) ex-
panding DRT structure to incorporate the pragmatic
extensions introduced in this paper, (b) representing
the complex process of speakers recognizing utter-
ances and using the linguistic information in form-
ing mental representations of hearers’ mental repre-
sentations, (c) enhancing the link between speaker
beliefs, and between the linguistic content and the
linguistic content and hearer’s beliefs and (d) putting
all these extensions and enhancements to the prag-
matic side of DRT in a computational model.
While the work presented in this paper offers a
more pragmatic approach to DRT, there is still more
work to be done on making DRT more pragmatic.
The possibility of extending the present treatment
to include more agents remains for future work.
In addition, future work can employ the intention
DRS introduced in this paper, in order to enhance
the complexity of the pragmatic representation of
speaker/hearer intentions. For instance, embedding
turn-taking acts within the intention DRS and relat-
ing them to agents’ beliefs and intentions should be
straightforward. It is also hoped that future work
will address more aspects of context than the two
detailed and implemented in this paper, namely, the
immediate and background context. Furthermore,
the sample implementation of the extensions sug-
gested in this paper serves as an example of how the
extensions to DRT can be implemented. One way
of developing this implementation is to incorporate
it into a dialogue system which aims to achieve a
more balanced approach to the semantic/pragmatic
interface in representing linguistic phenomena.

References
Kamp, H. and Reyle, U. 1993. From Discourse to Logic: In-
troduction to Model Theoretic Semantics of Natural Lan-
guage, Formal Logic and Discourse Representation Theory.
Boston, Dordrecht: Kluwer.
van Eijck, J. and Kamp, H. 1997. ‘Representing Discourse in
Context’. In: J. van Benthem and A. Ter Meulen (Eds.).
Handbook of Logic and Language. pp. 179–237. Amster-
dam: Elsevier.
Simons, M. 2003. ‘Presupposition and Accommodation: Un-
derstanding the Stalnakerian Picture’. Philosophical Studies
112, pp. 251–278.
Werth, P. 1999. Text Worlds: Representing Conceptual Space
in Discourse. New York: Longman.
