Embodied Conversational Agents (ECAs) are graphical entities with human-like appearance that are able to dialog naturally and spontaneously with users by recognizing and producing verbal utterances and non-verbal signals [1]. One key ability to ensure natural and spontaneous interactions with users is the ability to manage turn-taking [1]. Turn-taking refers to the ability to coordinate speaking turns during a conversation, so that participants speak mostly one at a time in alternating turns [2]. Human participants actively behave to ensure this coordination [3]. These behaviors are, either taking the turn, when the listener becomes speaker after the end of the previous speaker turn, grabbing the turn when the listener tries to become speaker while the speaker's turn is not finished, yielding the turn to the listener or keeping the turn [3]. Participants explicitly signal their behavior by exchanging verbal and nonverbal signals [2]. Concurrently to this alternation of speaking turns, participants often interrupt each other [4]. Interruptions are either cooperative, showing agreement or helping the speaker complete its turn, or competitive, showing disagreement or disinterest towards what the current speaker is saying [4]. Human participants often vary their behavior related to interruptions. They can choose to continue or stop speaking when detecting that their interlocutor interrupts them, or choose to interrupt or not their partner [4]. These behaviors are partly driven by the participants' dominant or submissive attitudes or their current emotional state [4]. The influence of these factors, especially the emotions, should be taken into account when creating an agent able to interrupt or manage the user's interruption. Indeed, in two perceptual experiments, [7] and [8] have observed that the way the agent handled interruptions influenced the user's judgment about the agent's dominant, submissive attitude or its emotions.
However, very few studies have taken into account dominance or emotions in the way the agent controls its behavior related to its emotions. Past works on interruption management in user-agent interactions encompasses models used to detect user interruption attempts [5], or to determine if the agent should resume its interrupted utterance or plan a new utterance after the user interruption [6]. In most of these models, the agent systematically stops its ongoing turn when it detects the user's utterance and waits the end of the user's interruption before starting to speak again. Moreover, very few models allow the agent to interrupt the user on purpose. To our knowledge, only [9] elaborated a model where the agent behavior varied according to the attitude of the agent towards its interlocutors. Yet, no computational model tried to link interruption management to emotions. This requires finding out which emotional state could lead the agent to interrupt the user, and which emotional states could lead the agent to continue or stop speaking when the user tries to interrupt it.
In this paper, we introduce a new model called EmoTurn (Emotional Turn-Taking). EmoTurn is a computational model for real-time user-agent interactions, where one component of an emotion, the action tendency [10] influences the agent's turn-taking behaviors. According to [10], action tendencies refer to readiness or unreadiness to engage in interaction in some particular fashion.
The proposed model relies on a previous computational model elaborated by Jégou et al. [11]. In Jégou et al.'s work, the agent's behavior towards turn-taking was allowed to vary according to a variable called motivation to change role. We based on this model to elaborate a set of rules determining how the motivation to change role varies according to the agent's current emotional state. After presenting our model, we show the ability of an agent controlled by our model to vary its behavior related to interruptions in real-time interactions with a human partner.
In this section, we present the EmoTurn model. We introduce the different variables used in our model, the motivation to change role coming from Jégou et al.'s model [11], and two action tendencies we selected to elaborate our model. We then explain the rules that combine the action tendencies and the agent's motivation to change role.
The model of Jégou et al. [11] is a theoretical model that controls the agent's verbal and nonverbal signals and the moments when the agent speaks based on its partner non-verbal signals. Two variables directly control the agent's behavior: the motivation to change role m and the degree of certainty γ towards the perception of the partner's behavior.
The motivation to change role, m, sets the goals of the agent, namely changing its role (speaker, listener) or keeping its current role. The motivation to change role influences the agent's final behavior, as for a current speaker, a motivation to change role makes it yield the turn, or for a listener, makes it take the turn. Conversely, a motivation to keep role makes the current speaker try to keep its turn and the current listener stay listener. This motivation (m) continuously varies between -1 (the agent strongly wants to keep its current role) and 1 (the agent strongly wants to change role).
However, the agent final behavior also depends on its partner's behavior. The agent continuously monitors the signals displayed by its partner and varies accordingly its own behavior. More precisely, the agent computes a degree of certainty γ about the partner's behavior based on the signals displayed by the latter. This degree of certainty informs the agent about whether its partner is currently trying to change role (become speaker for a listener, or become listener for a speaker) or to keep its role (staying listener or speaker). γ is used concurrently with m to control the agent's behavior. As a result, γ can potentially make the agent give up its current goal (for example, a listener that gives up its interruption attempt due to the turn keeping signals displayed by the current speaker).
The absolute value of the motivation defines the insistence the agent has in accomplishing its goal. For example, if a listener has a motivation value close to 1, it makes it more insistent to take the turn, even if γ indicates that its partner is trying to keep the turn, leading it to try to interrupt the speaker. Conversely, if the listener has a value close to 0, it will not try to take the turn while γ indicates that the current speaker is willing to keep the turn.
Our goal is to combine this mechanism with action tendencies. We selected several action tendencies from [10] to create the EmoTurn model. In this paper, we will focus on a subset of these action tendencies to illustrate the behavior of the model: the "Excited" action tendency, defined by a tendency to be "excited, restless" [10], and the "Inhibited" tendency; defined as a tendency to feel "paralyzed, or frozen" [10].
In our model, the action tendency does not directly influence the agent's behavior but is used to compute the agent's motivation to change role according to the following rules:
if the agent has something to say then if its current role is speaker then m ← motivation to keep turn else m ← motivation to take turn decrease / increase m according to the agent's action tendency else if its current role is speaker then m ← motivation to yield turn else m ← motivation to stay listener
We consider that the agent has four possible behaviors, keep the turn, take the turn, yield the turn or continue to listen. The factor defining the nature of the agent's behavior is whether the agent has something to say or not. Action tendencies do not influence the nature of the agent's behavior, but impact the strength of the motivation, making it more or less insistent in trying to yield or take the turn.
Depending on the action tendency, the value of the motivation to change role varies as specified in Table 1. This mapping is inspired by the semantic descriptions of action tendencies [10]. `"Strongly Unwilling" means that the agent will insist to keep the turn if the user wants to interrupt it whereas "Strongly Willing" means that the agent will insist to take the turn making him try to interrupt the current user. "Weakly Unwilling" means that the agent will yield the turn when detecting an interruption attempt while "Weakly Willing" means that the agent will not try to take the turn while the agent is still speaking. In section 3, we introduce more in details how, in our implementation, we currently compute the motivation values based on these rules.
In this section, we illustrate the behavior of our model in real-time interactions with a human partner. We consider a series of scenarios where the agent computes its motivation to change role according to its action tendency. Moreover, the agent interprets continuously the pitch and the acoustic energy of the user. The agent controls two types of non-verbal signals, its gaze (look towards the user or avert gaze) and its eyebrows (raising or not the eyebrows), which are signals used by humans to coordinate their turns [12], [13]. First, we present how we implemented our model. We then illustrate with four examples, how the agent behaves in real-time interactions with the user.
We have implemented EmoTurn in an existing agent architecture, AgentSlang [14]. In this architecture, we have created components dedicated to the real-time coordination of speaking turns with the user. The components are shown on Fig. 1.
We used the SSI Framework [15] to extract the energy e and pitch p from the user's voice. These values, p and e, are then normalized according to the user's minimum of maximum values of pitch and energy. These values are computed in a previous calibration step. Once the normalized values en and pn determined, they are transmitted to the User Behavior Interpreter component that computes γ. The EmoTurn component computes according to the following formula:
According to the agent's role, the different values of mu and m act are presented in Table 1. Based on m and γ, the Agent Signals Controller is in charge to compute gaze and eyebrows variations. It also determines when to launch or stop the agent's utterance according to an internal variable representing the agent acoustic energy. This component takes as inputs, m and γ. The Agent Signals Controler sends gaze and eyebrows variations commands to the realizer, formulated in BML (Behavior Markup Language) [17]. Based on the theoretical acoustic value, the module decides to launch or stop the agent's utterance. This decision is made such as, when the theoretical energy value is greater than a threshold value (0.2), a launch command is sent to the realizer, and when the theoretical energy value is less than this threshold, a stop command is sent to the realizer. We used MARC [16] as the realizer. MARC modulates the agent's nonverbal signals based on the command it receives, launches and stops audio files corresponding to the agent's utterance, and synchronizes the lips of the agent with the audio. We used audio records of human voices rather than a TTS to generate the agent's utterance.
We present four examples of interactions between our agent and a human partner. of the author of this paper played the human partner. These four examples cover four possible qualitative behaviors related to interruption management the agent can have depending on the agent's current role. As a listener, the agent either interrupts the user or take the turn after the user finished its turn. As a speaker, the agent either continues speaking when reacting to the user interruption or stops speaking and let the user become the speaker. The verbal utterances exchanged by the participants, not presented here, were inspired from [18]. The four scenarios are presented in Fig 2 . For each scenario, we represent the moments when the agent and the user spoke by the waveforms of the audio signals.
In scenario #1, the user speaks at the beginning of the interaction and the agent is the listener. After a few seconds, the dialog manager plan an utterance by setting mu to 0.5. The agent's action tendency is set to "Excited" in this scenario, thus mact=0.5 . As a result, even if γ indicates that the user is continuing its turn, the agent starts speaking while the user is still speaking. The overlap between the agent and the user lasts 1.6 s, after which the user lets the agent continue its turn. In the scenario #2, the agent is the listener. Similar to scenario #1, we simulate in the architecture the fact that the agent planned a new utterance ("has something to say" on Fig 2 .) after few seconds. Here, the agent's action tendency is "Inhibited", thus, mu=0.5 and mact=−0.5. As a result, the agent does not try to interrupt the user and waits for the end of the user's turn before taking the turn, leaving a gap of 1.3 s. In scenario #3, the action tendency is set to "Excited", the agent is the current speaker and the user is the current listener. After a while, the user starts to speak, however, since the agent did not finish its utterance, mu=−0.5 and mact=−0.5. When the user interrupts, γ increases, indicating that the user wants to take the turn. As the acoustic energy remains high during the overlap, the agent continues to speak. In scenario #4, the agent is also the current speaker. However, its action tendency is "Inhibited". Similar to scenario #3, the user tries to interrupt the agent after several seconds. However, since mu=−0.5 and mact=0.5 , the agent stops speaking to let the user speak.
In this paper, we presented EmoTurn, a model designed to make the agent vary the way it manages interruptions with the user according to one component of its emotions, the action tendencies. We then showed how our model integrates in an existing agent architecture, AgentSlang, and how it manages the launch and the stop of the agent's utterance in real-time interactions with a human partner.
We plan to validate the links between action tendencies and motivation to change role established in section 2. To that purpose users will interact in real-time with an agent varying its turn-taking behavior according to our model. At the end of the interactions, questions about the agent's action tendencies will be asked to the users. Finally, we only covered a subset of the type of interruptions that exist in human interactions. Especially, in order to have a complete model, we should distinguish cooperative and competitive interruptions.
Acknowledgments. This work was funded by the Agence Nationale de la Recherche (ANR), project NARECA (ANR-13-CORD-0015).