Sarcasm is a difficult concept to define accurately and completely and is similarly hard to identify in natural communication. In three works, this dissertation develops a deeper understanding of sarcasm, both as a concept and as a phenomenon. Using a computational approach, I describe novel markers of sarcasm that function as signals for the reader to pause, wait, or slow down, and show that they are often copresent with sarcastic content. Using experimental methods, I show that although sarcasm can be elicited in the laboratory at high levels, agreement on what statements are sarcastic is extremely unreliable. Finally, using a qualitative approach, I ask participants to modify statements to make them sarcastic, and consider the strategies they report using, as well as their success at doing so. Overall, my results suggest that sarcasm is an extremely diverse phenomenon, and that future research should adjust its focus to use a broader view of sarcasm, rather than using a specific definition.

In this dissertation I investigate the process of negotiation in conversation through the use of hedging words. My focus is on the ways we enter into and exit negotiation processes during conversations, and how we can explicitly signal this to our partners. I also examine the degree to which speaker perception is affected by the use of negotiation words. I propose that words that signal negotiation can be measured on two scales (correction and telling), and that there are words that signal more or less negotiation depending on context, conversational medium, and personal relationship.Experiments 1 and 2 are focused on identifying the scales upon which 13 words of negotiation fall. In Experiment 1, I investigate the words in isolation, asking participants to rate the words on a scale of telling-negotiation and correction-negotiation. In Experiment 2, I investigate the words in context, asking participants to read carrier scenarios and answer questions. I found that these words fall in roughly two groups - one that indicates high-telling low negotiation, and one that indicates low-telling high negotiation. Study 1 is a corpus analysis with the goal of investigating the use of I don’t know as well as absolutely, totally, kinda and sorta. I examine these words across different conversational settings (face to face, audiovisual, and text-only), different conversational types (chit-chat versus task), and different relationships (friends versus strangers) in order to build an account of how these words are used in various situations. I don’t know is most often used to indicate a lack of knowledge, and differs in forms across corpora - idk is more often used in the text-based corpus, compared to I dunno and I don’t know in the spoken corpora. Experiment 3 investigates how perceptions of the speaker change when negotiation words are used. Authority is manipulated using a professor - TA - student manipulation, and perceptions of knowledgeability, politeness, friendliness, and professionalism are probed. When looking at knowledge, the words that are low-telling high negotiation lower the perceived knowledge of the speaker, but the words that are high-telling low negotiation do not boost the perceived knowledge of the speaker. For friendliness and professionalism, clearly and obviously emerge as markers of speaker feeling, and for professionalism, peers are given more leeway than those with higher perceived authority. Understanding how we signal negotiation to our conversational partners, as well as what social effects these cues might have, has implications for understanding not just human interaction but human-computer and artificial agent interaction as well. Boosters and hedges, long examined as separate phenomena, can be categorized based on the negotiation functions they serve and the social effects they have.

This dissertation describes a series of studies testing the role of laughter in spontaneous conversation. Though laughter has typically been considered a response to humorous stimuli, one proposed function of laughter in naturally occurring talk is as a communicative signal indicating and attending to a potential source of social discomfort. We argue that laughter in conversation is a paralinguistic response to the desire for maintenance and management of social relationships during conversation. Laughter sends a message that some potentially alarming event need not be taken seriously, and invites the sharing of relief and mirth with others. Three studies tested this proposal. In the first, laughter was found to affect tension levels in conversation, and affect them differently depending on the initial tension level and who is laughing. In the second, the nature of the language just preceding laughter was demonstrated to be surprising and to co-occur with perception of fault on someone's part. And in the third, laughter was found to virtually escape notice in conversation as compared with two other non-linguistic noises, a sine tone and a cough sound, even when preceding talk was surprising or indicative of fault, and regardless of who was laughing. These findings are discussed as supportive of a model of laughter as a pragmatic device in conversation used to attend to and affect the health of social relationships.

Whether another person is perceived as similar or different from oneself can lead to differing impressions of their inherent human-like qualities (Haslam, 2006). As the boundaries between human-computer interaction and computer-mediated communication become increasingly fuzzy, it is important to assess in which ways these humanization measurements may also apply to non-human agents. There is evidence for interventions involving motion synchrony that can strengthen interpersonal bonds and perhaps instill a sense of similarity between individuals who otherwise have little in common (Wiltermuth & Heath, 2009), but it is as yet unclear whether similar effects can be observed for interactions between human and non-human agents. In a series of three experiments, I investigate the effect of synchrony on interactions with machine agents in three health contexts (mental, physical, and social health), each designed to be successively more applicable to real-life interactions between humans and machines (using static robot therapists, a physical robot exercise coach, and then an anonymous on-line debate and a game-like activity with a supposedly human or machine addressee). Humanization ratings did correspond to robot preference, and participation in an optional mimicry task after the initial training did lead to more positive humanization scores for the robot that was designed to help, rather than replace human therapists (Experiment 1). Synchronized stretching exercises resulted in warmer responses during interactions with a physical robot than non-synchronous exercises (Experiment 2). People gave higher humanization ratings to supposedly human agents, and are more persuaded by synchrony with anonymous human agents or non-synchrony with machine agents (Experiment 3). This research provides evidence for the importance of synchrony and backstories in building positive interactions between humans and machine agents, as well as between humans and other humans in anonymized contexts. However, it is important to bear in mind that attitudes around the capabilities of machines as compared to humans can lead to different responses following synchrony exercises. For example, while participants are more persuaded by emotionally distant machine agents, non-threatening outgroup members receive greater humanization benefits from motion-matching activities.

Language is fundamental to the human ability to work collaboratively on shared tasks. Current theories of dialogue, the format in which language use is shared and coordinated, emphasize cognitive processes that bring conversational partners’ talk into alignment. Driven by automatic priming of linguistic representation and the synchronization of embodied actions, the success of an interacting pair is ascribed to the degree to which the two converge. Such convergence, however, cannot explain how language use allows dyads to extend their abilities beyond those at the individual level, nor how a dyad might produce a novel idea. Instead, a new theory of dialogue must be developed, one that takes into account both convergent processes as well as processes that support divergence and complementarity. Borrowing a theoretical framing from dynamical systems and extended cognition, such a model is described as an interpersonal synergy. The role of divergence in supporting collaboration is tested across three domains in which collaboration has been previously tested. In Experiment 1, we test whether linguistic divergence supports joint decision making in a perceptual task, replicating and extending previous work that found that indiscriminate alignment negatively correlates with collective benefit. In Experiment 2, dyads engage in a creative humor production task together and alone. Here we correlate turn-by-turn divergence with collaborative success. In Experiment 3 we explore the extent to which overlapping or complementary contributions assist in the construction of a collaborative memory over repeated conversational rehearsals. Across these three domains different methods for measuring conversational divergence are established. We find that turn-by-turn progressivity predicts the extent to which pairs produce more humorous and creative jokes. However, we fail to correlate measures of conversational divergence with collaborative benefit in the decision-making and memory tasks. These studies and new methods represent a first step in developing a fuller theory of dialogue as the basis for the coordination of distinct information and contributions, such that both convergent and divergent processes may benefit the ability of conversational partners to engage in socially extended cognitive activities.

Every day, we have many conversations with others. What makes you consider these conversations as being good or bad? Three studies explored subjective conversational experience, or how people feel about conversations. Study 1 focuses on the replication of the reciprocity in conversation effect as well as the influence of off-task conversation on conversational perception (Guydish et al., 2021; Guydish & Fox Tree, 2022) by having individuals converse over videoconferencing software. This study showed that having an opportunity to engage in off-task conversation improves both conversation enjoyment but also perceptions of conversational partners. Study 2 focused on the influence of discourse markers on subjective conversational experience by analyzing their use in three corpora. While no relationship between discourse marker use and subjective conversational experience was observed, differing use of discourse markers between on-task and off-task conversation across corpora was observed. Study 3 focused on how memory plays a role in making distinctions between good and bad conversations by having individuals read and answer questions about previously transcribed conversational snippets. Study 3 showed the importance of contribution balance in making distinctions between good and bad conversations. These studies provide a more holistic picture of how we make determinations between good and bad conversations. In a world that is increasingly polarized, learning what and why our behaviors influence our subjective conversational experiences becomes increasingly important.

The Context Model of language processing describes the effect of context on the interpretation of linguistic input. This series of experiments builds on that model, shedding light on how social context, specifically attitude, affects language processing. Listener attitude affected attention to content (Experiment 1), but this effect itself depended on how attitude was induced (Experiment 2), and was found to affect the way the story was recalled, in both a verbal and written format (Experiment 3). In Experiment 1, the more a listener liked a speaker, the more they attended to the content of the speaker's narrative. In Experiment 2, the more similar or more different to the speaker a listener believed themselves to be, the more they attended to what was said. In Experiment 3, the more a listener liked a speaker, the more affective words they used. Overall, it was found that attitude affected language processing in terms of attention, comprehension, and reproduction.