1
The rhetorical attachment of
questions and answers
Philippe Muller and Laurent Prévot
1.1 Introduction
Placing ourselves within the tradition of dynamic semantic approaches
to discourse in natural language Kamp and Reyle (1993), Groenend-
jik and Stokhof (1991), we investigate here how some of the principles
underlying these approaches can be extended to the study of human
dialogue, taking over some hypotheses of Asher and Lascarides (1998,
2003). In this perspective, a dialogue is made of segments just like any
discourse; the semantics of each one of these segments is represented in
a logical framework and linked to other segments by so-called rhetori-
cal relations Hobbs (1985) that also carry semantic or intentional con-
tent, in the spirit of Segmented Discourse Representation Theory Asher
(1993). Thus can be integrated linguistic phenomena tackled by formal
semantics and more dialogue-specific characteristics such as turn-taking
conventions or common ground establishment.
We will focus here on the issue of question/answer pairs and on the
way they structure some of the established conversational content. More
specifically, we have focused on Yes/No questions (questions for which
the expected answer can be Yes or No, possibly with some additional
material). In order to do so, we have collected a corpus of dialogues from
phone conversations in a specific, constructed setting. This corpus is
made of 21 transcribed 21 conversations. Phone conversations eliminate
deictics and gestures as well as facial expressions in order to focus on
verbal communication. Each dialogue involves a “giver” and a “receiver”:
SPR.
***.
Copyright c 2006, CSLI Publications.
1
,2 / Philippe Muller and Laurent Prévot
the giver explains to the receiver how to go from one place to another,
in the same city. The global subject or topic of this kind of discussion
is the explained route.
After a brief introduction to SDRT (section 1.2.1), we will look in sec-
tion 1.2.2 at the classical proposals of this theory for attaching yes/no
questions in the rhetorical structure of a dialogue, and pursue with a
critical analysis of these solutions. Section 1.3 will focus on the link be-
tween question/answer pairs and the other rhetorical relations. In the
process we will introduce and stress the need for a set of additional re-
lations, especially for treating interrogative narration or correction and
request for confirmation. We will then propose a new solution (section
1.4) for attaching questions and answers.
1.2 SDRT and Dialogue
1.2.1 SDRT : the basics
We will give here a very rough outline of Segmented Discourse Repre-
sentation Theory Asher (1993) and its extension to dialogue. Asher’s
theory assumes that a dialogue is a kind of discourse involving two
participants. SDRT also assumes, as the RST of Mann and Thompson
(1987), that a discourse can be seen as a set of segments linked with
rhetorical relations. These relations can be hierarchical or not. Hierar-
chical relations between segments induce a tree structure which imposes
constraints on the interpretation of current utterances (e.g. anaphora
resolution can be limited to certain segments of the current interpre-
tation).1 More generally, constraints based on the structure induced
by the segmentation determine what is a coherent dialogue. New utter-
ances will be attached to some segment incrementally within an already
existing dialogue structure as they come, by taking into account lexi-
cal semantics, world knowledge and semantic-pragmatic rules selecting
an appropriate relation in context, see Asher and Lascarides (1998).
These aspects –which are often taken for granted in theories taking as
primitives propositions and speech acts related to these propositions–
makes SDRT perspective more fine-grained than that of game-based
or intentional analyses. In spite of the latest proposal made in Asher
and Lascarides (2003) we believe that there is room to improve some
aspects related to dialogue and we will question some of its choices
on the attachment of questions. We will therefore keep the following
principles, while leaving aside the presently less stable aspects of the
theory:
1 Rhetorical relations are thus either subordinating in the tree structure, or coor-
dinating, i.e. linking two daughters of a same mother node.
, The rhetorical attachment of questions and answers / 3
. the global representation of a dialogue is composed of a set of labelled
speech acts (SAs) and rhetorical relations between these occurrences
of SAs. A speech act will be of the form h Speaker,Mood,Content i,
where the mood can be interrogative(?), declarative(.) or impera-
tive(!). A basic semantic content will be a DRS, cf. Kamp and Reyle
(1993), that is a set of linguistic referents and of conditions (predi-
cates) on those referents. Speech acts, seen in other frameworks as
having an intentional content, take on an intentional dimension only
when linked together by relations bearing an intentional content.
. We have divided them into the following categories :
—monologic relations : The set of these relations is given in the
work of Asher (1993) updated in Busquets et al. (2001); it is com-
posed of subordinating relations (elaboration, explanation,. . . )
and coordinating ones (narration,. . . ).
—dialogic relations : The set of these relations coming from
Asher and Lascarides (1998) has been updated in Asher et al.
(2001), Asher and Gillies (2003); it is composed of Plan-
elaboration, Question-elaboration , Question-Answer-Pair 2 , Not-
Enough-Information, Acknowledgement, Correction.. In Asher
and Lascarides (2003) a systematic account of mood is added to
the relations. Each indicative relation of monologue can now be
uttered with interrogative (e.g narrationq , elaborationq 3 ,. . . ) or
imperative mood. There are also some additional relations (e.g
Plan-correction, . . . ).
1.2.2 Attachment of questions
Examples (1.1-1.3) and their discourse structure (Fig. 1) shows how
questions and answers can be related by different relations, A3 A′3 , A′′3
and A′′′
3 being possible continuations of A1 -B2 . Here πi is the label of
the speech act made in turn i (each turn being here only one segment).
Graphically, we will represent a subordinating relation with a vertical
segment between two labels, and a coordinating relation with an hori-
zontal one.
2 Noted QAP, with a distinction as proposed in Karttunen (1977) between QAP,
Indirect QAP (IQAP) and Partial QAP (PQAP).
3 The interrogative nature of these relations makes them subordinating.
