Which word gets the nuclear stress in a turn-at-talk?

Abstract

Information in turns-at-talk is hierarchically structured with some information made more prominent than other information. A key indicator of information structure is nuclear stress. In this paper we aim to identify factors that correlate with nuclear stress placement.Based on exhaustive manual annotation of ten-word turns extracted from the demographically-sampled subcorpus of the British National Corpus (BNC-C), we consider a large number of factors potentially affecting nuclear placement, including duration (measured in Praat, Boersma and Weenink, 2019), word class (based on PoS tags), position in the turn (each word slot representing a position), frequency (derived from BNC), surprisal (predictability of word B given word A; e.g., Jaeger, 2010), and turn part (pre-start, turn-constructional unit [TCU], post-completer; cf. Sacks et al., 1974).To determine which variables correlate with nuclear stress, we fitted a fixed-effects binomial logistic regression model to the data. The regression analysis confirmed that nuclear stress is affected by several factors simultaneously and that the probability of a word receiving stress is highest for long, surprising, infrequent words occurring late in the turn and integrated into the TCU rather than being peripheral unless they have (extremely) high surprisal values. The study offers implications for our understanding of turn structure, arising from constraints set by turn-taking, and differential action types facilitated in different turn parts. Moreover, the results provide strong support for the assumption that speakers have access to probability distributions (e.g., Jaeger, 2010) and skillfully rely on them during turn production for interactional and informational needs arising from situated discourse.