Information structure and the prosodic structure of English : a probabilistic relationship
MetadataShow full item record
This work concerns how information structure is signalled prosodically in English, that is, how prosodic prominence and phrasing are used to indicate the salience and organisation of information in relation to a discourse model. It has been standardly held that information structure is primarily signalled by the distribution of pitch accents within syntax structure, as well as intonation event type. However, we argue that these claims underestimate the importance, and richness, of metrical prosodic structure and its role in signalling information structure. We advance a new theory, that information structure is a strong constraint on the mapping of words onto metrical prosodic structure. We show that focus (kontrast) aligns with nuclear prominence, while other accents are not usually directly 'meaningful'. Information units (theme/rheme) try to align with prosodic phrases. This mapping is probabilistic, so it is also influenced by lexical and syntactic effects, as well as rhythmical constraints and other features including emphasis. Rather than being directly signalled by the prosody, the likelihood of each information structure interpretation is mediated by all these properties. We demonstrate that this theory resolves problematic facts about accent distribution in earlier accounts and makes syntactic focus projection rules unnecessary. Previous theories have claimed that contrastive accents are marked by a categorically distinct accent type to other focal accents (e.g. L+H* v H*). We show this distinction in fact involves two separate semantic properties: contrastiveness and theme/rheme status. Contrastiveness is marked by increased prominence in general. Themes are distinguished from rhemes by relative prominence, i.e. the rheme kontrast aligns with nuclear prominence at the level of phrasing that includes both theme and rheme units. In a series of production and perception experiments, we directly test our theory against previous accounts, showing that the only consistent cue to the distinction between theme and rheme nuclear accents is relative pitch height. This height difference accords with our understanding of the marking of nuclear prominence: theme peaks are only lower than rheme peaks in rheme-theme order, consistent with post-nuclear lowering; in theme-rheme order, the last of equal peaks is perceived as nuclear. The rest of the thesis involves analysis of a portion of the Switchboard corpus which we have annotated with substantial new layers of semantic (kontrast) and prosodic features, which are described. This work is an essentially novel approach to testing discourse semantics theories in speech. Using multiple regression analysis, we demonstrate distributional properties of the corpus consistent with our claims. Plain and nuclear accents are best distinguished by phrasal features, showing the strong constraint of phrase structure on the perception of prominence. Nuclear accents can be reliably predicted by semantic/syntactic features, particularly kontrast, while other accents cannot. Plain accents can only be identified well by acoustic features, showing their appearance is linked to rhythmical and low-level semantic features. We further show that kontrast is not only more likely in nuclear position, but also if a word is more structurally or acoustically prominent than expected given its syntactic/information status properties. Consistent with our claim that nuclear accents are distinctive, we show that pre-, post- and nuclear accents have different acoustic profiles; and that the acoustic correlates of increased prominence vary by accent type, i.e. pre-nuclear or nuclear. Finally, we demonstrate the efficacy of our theory compared to previous accounts using examples from the corpus.