Modeling Politeness in Human-Robot Interaction

Politeness is a linguistic phenomenon central to human communication. The many influences on the choice and interpretation of politeness in language make the phenomenon very complex to model and to account for in Human-Robot Interaction. I therefore question whether and how the implementation of such a social linguistic phenomenon in a robot is desirable. To find answers I conducted studies on politeness in Human-Robot Interaction from different perspectives: users' expectations, users' politeness towards a robot and users' perception of politeness use by a robot. In future research, based on and informed by my results, I aim to computationally model two politeness strategies and implement them in a Furhat robot.

There is extensive research on linguistic politeness (henceforth politeness) trying to account for its complexity with multiple theories and models [discussed, e.g., in 13,26].Brown and Levinson's [3] theory is widely adopted.They argue that the degree of face threat occurring in conversation determines the use of politeness strategies, where face is the public self-image that a person wants to preserve in conversation [3].They also consider three main infuences on the choice of politeness strategy: the power of the hearer over the speaker, the social distance between the speaker and the hearer as well as the rank of imposition of the chosen utterance [3].
Despite critique, especially regarding claims about its universality [2], Brown and Levinson's [3] politeness theory is still often used as a basis for politeness modeling and implementations, also in human-robot interaction research [23].
Previous research on politeness in HRI has produced seemingly contradicting fndings regarding this topic.On the one hand, studies focusing on users' perception of polite machines found mostly positive efects of politeness use by agents, for example, increasing the likability of a robot [24], the trust towards non-humanoid robots [12] and the persuasiveness of a virtual agents [11].On the other hand, studies focusing on the expectations regarding an agents' use of politeness found that users might prefer that robots use more direct language [20] and that indirect or vague language can lead to a bad user experience [7].A possible explanation could be that concepts such as face are not trivially transferable to HRI [6].
Other research has investigated how users behave towards robots.Here, studies have found that users seem to be polite towards machines [22] as well as towards embodied conversational agents [9].These fndings led to the proposal of the 'Computers are Social Actors' (CASA) theory, which claims that humans mindlessly apply social rules, such as politeness, in conversations with agents [21].One challenge for this theory is that conversational goals seem to difer between HRI and HHI in the frst place [8].The observation that users show politeness towards robots also seems to contradict users' expectations and preferences regarding robots' politeness.
My third research question therefore addresses these contradictions: Is there a discrepancy between what humans expect and perceive regarding a robot's politeness?(RQ3).
Additionally, in order to connect linguistic politeness research with HRI research, my work tries to answer another question: How can politeness be modeled linguistically by considering empirical and theoretical research to be implemented in a robot?(RQ4).

RESEARCH APPROACH
The aim of this PhD project is to combine diferent linguistic research methods with computational approaches to answer the research questions described above.
To gain further insights on politeness in HRI, I considered different perspectives by applying a range of research and evaluation methods including quantitative and qualitative methods.I frst collected data with a quantitative online study [16].This study was insightful, but limited to pre-defned categories due to its methodology.To gain more insights on participants' concepts for and expectations regarding politeness in HRI (RQ1), I therefore conducted a qualitative interview study (Study 1 below).To analyze whether these conceptual expectations match participants' actual behavior in interaction with a robot (RQ3), I afterwards conducted an interaction study (Study 2 below).Together with prior politeness research, I will use insights gained from these studies to construct a linguistic computational model of politeness strategies (RQ4).In order to evaluate the model it will be implemented in a Furhat robot and tested by users in an interaction study (future work in section 4).Further, based on the evaluation study I will be able to directly compare expectations (from study 1) with the actual perception of politeness strategies when implemented in a robot (RQ3).

PREVIOUS WORK
Study 1: Infuences and expectations regarding politeness.Through semi-structured interviews, this study collected insights on the expectations and ideas that participants (N=17) as laypersons have on politeness.First, participants discussed politeness in general and then, after being confronted with a Furhat robot, politeness in HRI.A more detailed description of the study procedure and results can be found in [17].I analyzed the interview data using a thematic analysis approach [1].Regarding politeness in general, I identifed two overall types of politeness strategies mentioned by our participants: rule-governed politeness and adaptive politeness strategies.Rule-governed politeness strategies are based on societal and cultural norms and conventions.These include phrases that we are taught as children such as saying "please" and "thank you".Adaptive politeness, on the other hand, refers to strategies that are used to adapt to the interlocutor, situation and topic during interaction.Such strategies include active listening and indirectness.From the Furhat robot, participants mostly expected only rulegoverned politeness.Participants stated that they would only expect a robot to use the more complex adaptive politeness strategies when it is used in a private setting at home.A discussion of the results and further insights are presented in [17].These insights contribute to answering RQ1.To see whether these expectations match with user behavior (RQ3) and to analyze politeness production I conducted an empirical interaction study (Study 2).
Study 2: Users' politeness in disconfrming answers towards a robot compared to a human.In the second study, participants (N=40) frst interacted with a Furhat robot in a Wizard-of-Oz set-up and afterwards with a human researcher.The interaction scenario in the study was designed to elicit disconfrming answers (e.g."no") to the same questions frst asked by the robot and then by the human.More information on the study can be found in [18].The audio and video data we collected was frst analyzed with quantitative measures by comparing the use of indirect disconfrmations and the use of linguistic markers for indirect politeness (e.g., hesitations, hedging and added explanations) towards the robot and the researcher.To fnd and analyze further politeness markers I will conduct a qualitative analysis based on conversation analytical methods in future research.
Overall the results show clear diferences in the use of indirect politeness and politeness markers towards the Furhat robot and the human [18].When talking to the human, participants used fewer direct disconfrmations but more indirect disconfrmations than when answering to the Furhat robot.Further, all politeness markers that were analyzed were used signifcantly more with the human than with the robot.This overall showed a clear tendency for participants to be more direct and use less indirect politeness towards the robot than the human [18].These fndings contradict the CASA theory [21] as they indicate that our participants did not mindlessly apply social linguistic strategies when interacting with the Furhat robot.One possible explanation is that participants did not expect the robot to be able to process more complex politeness cues [18].Arguably, these results are more compatible with the depiction model [5].
Together with the results from the previous study (refecting a self-report perspective), insights gained from this study (refecting an observational stance) can inform us more accurately and from more perspectives about the partner models that users build in interaction with robots.

FUTURE WORK
In future work I will model two politeness strategies based on my previous results and implement them in a Furhat robot to evaluate them in an interaction study.
Considering politeness markers (produced by participants in Study 2) and infuences on politeness found in previous research, I plan to model the adaptive and rule-governed politeness strategies (from Study 1) by implementing them in a Furhat robot.Relevant infuences on politeness that I plan to consider in my model include gender [10], mood [27], emotion [25,27], situation [3,14], and power and distance [3].Further, based on earlier results [16], for the implementation I will also control the role that the robot is presented in, as the role determines the user's relationship with the robot.Building on previous small-scale Bayesian modeling of politeness [15], this probabilistic Bayesian model will choose a politeness strategy based on the parameter setting of the diferent infuences.
To evaluate my model, I plan to conduct a between-subject study with two politeness conditions.In one condition participants will interact with the Furhat robot that uses adaptive politeness strategies while in the control condition, the robot will use rule-governed politeness strategies.I plan to evaluate the results with diferent methods to gain insights on participants' perceptions after interaction through self-report measure (e.g., through RoSAS [4]) and during interaction using conversation analysis as an observational method.The implementation of my model and its evaluation will allow me to validate it and its functionality to answer my RQ4.
Because of the combination of diferent methods, theories and implementation, this PhD project is a comprehensive and novel contribution to research on HRI as well as politeness.