How a Conversational Agent Should Respond to Verbal Abuse

We study how a conversational agent should respond to verbal abuse by a user. Our goal is to understand which agent response style (Avoidance, Empathy or Counterattacking) positively influences those emotions found to intervene in users’ aggressive behaviors (i.e., guilt, shame). We also examine whether the style of agent responses affects the users’ evaluation of the agents

We examine whether a conversational agents’ response style under varying abuse types inﬂuences those emotions found to mitigate peoples’ aggressive behaviors. The participants assigned to one of the abuse type conditions, interacted with the three conversational agents in turn and reported their feelings about guiltiness, anger, and shame after each session. Two experiments were conducted in each text and voice-based environment. We investigated the current status of commercial CAs’ responses to verbal abuse.

Our study results show that, regardless of the interaction type, the agents’ response style has a signiﬁcant eﬀect on user emotions. The verbal abuse type that user exercised has a signiﬁcant inﬂuence on users’ degree of shame only in voice interaction.
The primary findings are that the current IPAs mostly rely on the avoidance strategy in coping with verbal abuse and the agent’s response style has a significant effect on user emotions associated with mitigating verbal aggression and on the evaluations of the agent’s capability.
Participants felt lower anger and higher guilt when dealing with the empathetic agent and evaluated the empathetic agent as the most capable agent.
The users rated the avoidance CA as most inappropriate and incompetent than the other two agents.
User assessments about the counterattacking CA have shown conflicting results.
Considering that major IPAs are generally taking the approach of avoidance,the current strategies of major IPAs for dealing with verbal abuse seems inadequate.