Prosodic marking of contrasts in information structure

Successful dialogue requires cultivation of common ground (Clark, 1996), shared information, which changes as the conversation proceeds. Dialogue partners can maintain common ground by using different modalities like eye gaze, facial expressions, gesture, content information or intonation. Here, we focus on intonation and investigate how contrast in information structure is prosodically marked in spontaneous speech. Combinatory Categorial Grammar (CCG, Steedman 2000) distinguishes theme and rheme as elements of information structure. In some cases they can be distinguished by the pitch accent with which the corresponding words are realised. We experimentally evoke instances of contrasting themes and rhemes to establish the circumstances under which the pitch accents occur in unrestricted spoken dialogue. ‘Contrast’ means ‘alternatives are available’, not ‘contrastive accent’. It is difficult to manipulate context or outcome in quasi-natural engaging situations. Even if contrasting themes and rhemes are available, speakers choose from among a wider set of contrastable elements when framing utterances. Their choice may be difficult to predict: contrasts not apparently critical to the local context may be as important to speakers as ones usually thought to define the situation under discussion. Unscripted dialogue with pressing communicative motivation is difficult to control for genre, topic, and goals. We use a modified map task (Anderson et al. 1991), a restricted-domain route-communication task, which establishes what each participant knows at any time. Without sight of each other’s maps, an Instruction Giver (IG) and Follower (IF) collaborate to reproduce on IF’s map a route printed on IG’s. The route can be adequately described by routecritical landmarks. As Fig. 1 illustrates, map pairs differ in the features of landmarks and in ‘ink damage’ that obscures the colours of some landmarks on IF’s map. Participants know that maps can differ but must learn where and how. The discrepancies between maps do not fully define the alternatives sets speakers may wish to contrast. Instead, speakers define that alternatives set by their intonation. Provided that it is consistent with the context, the hearer will accommodate that set. Take: (1) IF: Do you see the two brown trees and the and the four black trees? IG: You mean THREE black trees right? (1:1–2:T:700.7; 1–1) By deaccenting ‘black’ and ‘trees’ IG presupposes that the alternatives are confined to sets of black trees; specifically to IG’s set of three and IF’s set of four. Both can then adjust common ground incrementally. As there is intense debate about whether the involved pitch accents (L+H* and H*) are actually categorically distinct (Ladd & Schepman 2003, Calhoun 2004), we simply seek to establish that contrasts in the information structure are indeed marked overtly by some form of prominence. We therefore use an undifferentiated notion of perceptual prominence to determine whether contrasts are marked by phonetic means. Our prediction is the following: Only words whose denotation contributes to distinguishing the entity referred to from the other entities in the alternatives set are marked by prominence.