The Phonetics of Stance-taking

Abstract

Stance -- attitudes and opinions about the topic of discussion -- has been investigated textually in conversation- and discourse analysis and in computational models, but little work has focused on its acoustic-phonetic properties. It is a challenging problem, given the complexity of stance and the many other types of meaning that must share the same acoustic channels, all of which are overlaid on the lexical and syntactic material of the message. With the goal of identifying automatically-extractable, acoustically-measurable correlates of stance-taking, this dissertation presents a new audio corpus of stance-dense interaction and three phonetic experiments which find signals of stance in prosodic measures of pitch, intensity, and duration. The ATAROS corpus contains pairs of speakers engaged in collaborative conversational tasks designed to elicit frequent changes in stance at varying levels of involvement. Interactions are transcribed, time-aligned to the audio, and manually annotated for stance strength (none, weak, moderate, strong), polarity (positive, negative, neutral), and stance type (e.g., opinion-offering and soliciting, (dis)agreement, persuasion, rapport-building, etc.). In the first experiment, combinations of pitch and intensity contours are shown to differentiate four discourse functions within a small sample of instances of the word `yeah' that contribute to negative stances. In the second experiment, vowel duration and intensity separate six common stance-act types in over 2200 `yeahs,' changes in pitch and intensity correlate with stance strength, and all three measures are involved in signaling positive stance. The third and largest experiment examines over 32,000 stressed vowels in content words spoken by 40 speakers and finds that pitch and intensity increase with stance strength, longer vowel duration is the primary signal of positive polarity, and a combination of these measures helps distinguish several notable stance-act types, including: agreement in general, weak-positive agreement, rapport-building agreement, reluctance to accept a stance, stance-softening, and backchannels. These results, and the corpus itself, contribute to the study and understanding of the acoustic-phonetic properties of the social and attitudinal messages conveyed in natural speech, information which may be of use to future work in theoretical, experimental, and computational linguistics.