Rhetorical Structure and QUDs

We consider two hypotheses about how rhetorical structure and QUD structure might come together to provide a more general pragmatic theory. Taking SDRT ([2]) and some basic principles from [18]’s QUD framework as starting points, we first consider the possibility that rhetorical relations can be modelled as QUDs, and vice versa. We ultimately reject this hypothesis in favor of the possibility that QUDs correspond to topics that bind together the members of complex discourse units. Theories of rhetorical structure [2, 13] and theories of discourse structure centered around a Question Under Discussion or QUD [8, 18] share many of the same principles.3 Both approaches hold that the interpretation of a given sentence or elementary discourse unit (EDU) depends in part on that EDU’s relation to other moves that have been made in the same discourse.4 Both also hold that a discourse context must therefore keep track of not only (some subset of the) prior discourse moves but also certain structural relations between these moves. These structural relations are believed to play an integral role in discourse coherence and the relevance of individual discourse moves, and as such, to influence various semantic and pragmatic phenomena, including ellipsis of various sorts, anaphora (rhetorical theories), and prosody (QUD theories). To clarify their potential contribution to analyses of semantic and pragmatic phenomena, we need a better understanding of how rhetorical and QUD frameworks are related: are they fundamentally distinct, but complementary theories, or do they aim to model the same phenomena? If the latter, do they end up describing two sides of the same coin or are they in conflict? The goal of this paper is to propose and evaluate two hypotheses about how the two frameworks might correspond, so that we can eventually come to a better understanding of the respective roles that these frameworks play with regard to phenomena from the semantics-pragmatics interface. The first hypothesis, which we develop and ultimately reject in §2.1, is that there is a direct correspondence between instances of discourse relations and QUDs in a discourse. The second hypothesis, which we present in §2.2, is that QUDs correspond to complex discourse units in a discourse graph. We judge this hypothesis to be more promising. Our discussion focusses on two particular theories. On the side of rhetorical structure, we adopt Segmented Discourse Representation Theory [2], which we briefly introduce in §1.1. Of all of the rhetorical theories, SDRT is the most developed from a ? We gratefully acknowledge support from the ERC grant 269427 and Marie Curie FP7 Grant, PCIG13-GA-2013-618550. 3 See also [21] for a framework that combines QUDs and rhetorical relations. 4 How a discourse is broken down into basic units can vary from theory to theory, but all rhetorical theories and QUD theories must take a stand on what constitutes a basic discourse move. semantic point of view; it assigns semantics to each of its discourse relations and posits semantic constraints on how its hierarchical discourse representations, which capture the contents of full discourses, can be constructed. These choices are fuelled by concerns about anaphora resolution, presupposition, temporal interpretation and other phenomena relevant to the semantics-pragmatics interface, putting SDRT in an ideal position to be compared to alternative semantic/pragmatic theories of discourse structure. On the side of QUDs, we take inspiration from [18] whose proposal, we think, is closest in spirit to SDRT. Like SDRT it aims to capture relations between discourse moves and provide hierarchical discourse structures that model global features of the discourse context. However, because the account outlined in [18] less thoroughly developed than SDRT, the possible formulations of QUD frameworks that we consider in §1.2 sometimes move beyond basic principles laid out in [18] in directions that we think deserve consideration, but which may not be endorsed by Roberts. 1 Background on SDRT and QUD 1.1 A very brief introduction to SDRT A fundamental principle of rhetorical theories, including SDRT, is that the relations that utterances stand in to one another affect on the one hand the interpretation of the discourse in which the utterances figure and on the other, the interpretation of the utterance contents themselves. Consider (1): (1) I missed my meeting this morning. My car broke down. Seeking a connection between the two parts of (1), an addressee will naturally understand the content of the second sentence as providing an explanation for the content of the first. This interpretation of the discourse comes with its own truth conditions: it is true just in case the speaker was late for her meeting, her car broke down, and the latter event was the cause of the former. At the same time, the inferred causal connection affects the interpretation of the two sentences in (1). Both sentences are in the past tense, which indicates that the events that they describe occurred before the speech time. The inferred causal relation between the two events entails that in addition, the second event must have occurred before the first—a cause must occur before its effect. This sequence of events does not follow from the tense of the verbs and the order in which the events are described in the discourse, but only from understanding how the two sentences are coherently or rhetorically related. Of course, in the absence of an explicit marker for causality (e.g. because), this causal connection is at best an implicature; nevertheless, speakers regularly accept and act on such implicatures. SDRT aims to model such connections by developing rules for constructing logical forms for discourses and by providing models to interpret these logical forms. To capture the rhetorical connections in a discourse, rhetorical theories must accomplish three tasks. First, the discourse must be segmented into EDUs according to rules set by the theory. In SDRT, the aim is to have each EDU denote a single eventuality. The next task is to attach each EDU to some other part of the discourse. SDRT maintains that a new move can only be a coherent extension of a given discourse if it is relevant to some other move that has been made previously; thus, each EDU will be attached to at least one other EDU that was discourse prior to it, with the exception of the initial EDU, which will only have attachments to later moves. In this way, a rhetorical relation is a kind of anaphoric relation, but one that holds between discourse units rather than referring expressions and discourse referents. The final task to be accomplished, which is in practice accomplished in tandem with labelling and even segmentation at times, is the labelling of each discourse attachment with a rhetorical relation (Explanation, Elaboration, Contrast, Narration, etc.). Attachment and labelling involve default reasoning about the contents of the EDUs involved and world knowledge. Importantly, however, they also take into account global features from the discourse. For example, because discourse relations are often implicatures, the inference of an attachment/label that might otherwise be justified can be blocked by information carried by other EDUs in the discourse. For this reason, SDRT provides defeasible rules for inferring discourse relations and stresses the need for global constraints on the development of a discourse structure. One such global constraint, which affects EDU attachment, is the Right Frontier Constraint (RFC). SDRT, along with other discourse theories [16, 15], posits that even if the content of a new EDU satisfies the semantic conditions for being attached to some other EDU in a discourse, this connection is only coherent if the prior EDU is accessible5, that is, if it is on the Right Frontier (RF). Compare (2-a)-(2-c): (2) a. [John speaks German.]π1 [He can translate for you while you’re in Berlin.]π2 b. [John speaks German]π1′ [and his sister speaks French.]π2′ ?? [He can translate for you while you’re in Berlin.]π3′ c. [John speaks German.]π1′′ [(because) He lived in Stuttgart for 10 years.]π2′′ [He can translate for you while you’re in Berlin.]π3′′ While π1 and π2 in (2-a) are related by Result—a connection supported by the contents of the EDUs and world knowledge about what being able to translate in Berlin would normally require—the same connection is blocked for π1′ and π3′ in (2-b). (2-c), however, shows that it is not blocked by just any intervening EDUs. SDRT defines the RF so as to reflect facts about accessibility like those illustrated in (2-a)-(2-c). We start by introducing SDRT’s discourse structures, which are connected graphs rather than trees. Graphs are needed to model certain facts about discourse, two of which are relevant for this paper: (i) some units can have incoming links from two different EDUs, and (ii) multiple DUs can work together to form a complex discourse unit (CDU), which serves as a single argument to a discourse relation. (3) a. Sam is being punished.π1 She took her parents’ car without permission,π2 so they’ve grounded her for 2 weeks.π3 b. Explanation (π1, π2), Result (π2, π3), Elaboration (π1, π3) (4) a. π1′ + π2′ , but their parents don’t speak any foreign languages.π3′ b. Continuation(π1′ , π2′), Contrast ([π1′ , π2′ ], π3′ ) In (3), π3 is the second argument for an instance of Result and an instance of Elaboration, whose first arguments are distinct. In (4), which builds on π1′ and π2′ from (2-b), 5 SDRT does allow for violations of the RFC in cases that [1] calls discourse subordination, but such violations need to be explicitly signalled, e.g. Let’s go back to your first point. not only are both π1′ and π2′ needed to provide the necessary antecedent for their in π3′ , but these units are attached to each other and both satisfy the conditions needed to support the Contrast with π3′ . SDRT would therefore group them into a single, though internally complex, argument for Contrast. SDRT’s graphs thus contain two types of edges: (i) edges that are labelled with discourse relations, and (ii) edges that relate each CDU to each DU that it contains. Edges of type (i) can be further subdivided into subordinating and coordinating edges, governed by the semantics of the relations that label these edges. When a DU is attached to another DU via a subordinating relation, both arguments remain accessible for further attachments. In (2-c), for example, the role of π2′′ is to provide background information that explains how John came to speak German (π1′′ ) or to simply back up that claim. π1′′ therefore remains central to the discussion and salient enough to be accessible to π3′′ despite the intervening π2′′ . Subordinating relations include Explanation, Elaboration, Background, and Question-Answer Pair, among others. When a DU is attached to another DU via a coordinating relation, the RF is pushed dynamically forward so that the second DU is on the RF, but the first is knocked off. In (2-b), for instance, π2′ goes on to tell us that John’s sister speaks French, so the discourse is no longer centered on John’s German speaking abilities when it comes time to add π3′ . An attachment from π3′ to π1′ is thus difficult to achieve. Coordinating relations include Contrast, Continuation, Narration (Sequence), Result, and Conditional, among others. We conclude with a description of SDRT’s RF. A node πx is on the RF of a graph G, i.e. RFG(πx), just in case (a) πx is Last, i.e. πx is the EDU introduced most recently into G following the textual order of the EDUs in G, or (b) ∃πy(RFG(πy)) such that (i) es(πx, πy) for a subordinating edge es or (ii) πy ∈ πx (i.e., πx is a CDU). The Right Frontier Constraint (RFC) simply states that an incoming EDU that needs to be attached to a graph G should be attached to a node along the RF.