Developing a Personality Model for Speech-based Conversational Agents

Speech-based conversational agents have become increasingly popular, often presented as helpful assistants in everyday tasks. Due to their intended use and conversational nature, this type of user interface seems much more likely to be seen as a “being with a personality,” for example Siri sometimes gives funny answers or the Google Assistant uses filler sounds such as “mmhmmm”. Hence, users unconsciously attribute personality traits to voice assistants, like friendly, helpful, or funny. According to previous research, acceptance and credibility of voice assistants are determined by their abilities to be perceived as having a consistent and coherent personality. However, systematically designing agent personalities remains a challenge. As of today, no dedicated personality model for speech-based conversational agents exists. Thus, most researchers have turned to the Big Five personality taxonomy. The Big Five Taxonomy is the most predominant model for describing human personality in Psychology. As this model was developed for humans, it remains unclear if it is actually suitable to describe agents. For example, the Big Five dimension of openness might be less applicable or important for agents. Moreover, further dimensions beyond those in human models might be necessary to sufficiently describe agents. For example, in previous research, people used adjectives like “logical”, “robotic”, or “technical” to describe a voice assistant’s personality.

Research Questions

RQ1 How can personality of conversational agents be described systematically?

To answer our research question of how we can symstematically describe conversational agent personality, we follow the psycholexical approach, which was also used to derive the established Big Five personality model in psychology [de Raad 2000]. The lexical hypothesis assumes that since people notice and talk about individual differences, these “will eventually become encoded into their language”. The psycholexical approach consists of two steps: (1) Item Pool Generation, in which unique phrases for describing personality are collected, followed by (2) Exploratory Factor Analysis, which explores the structure and relationship between the items. Determining the optimal procedure for compiling a representative set of descriptors from the entire lingual lexicon represents a major challenge of the psycholexical approach. For example, Norman identified 27,000 personality terms in a dictionary. Researchers reduced this number of descriptors by successively excluding unfamiliar, slangy, redundant, etc. terms based on expert and empirical ratings. Inspired by traditional test construction theory, we used a new multi-method approach to collect potential descriptors.

We asked 744 participants on Mechanical Turk to rate a voice assistant they are familiar with on all of the 349 descriptors in an online survey. We investigated the descriptor set’s underlying structure with an exploratory factor analysis based on the correlation matrix of all 349 descriptors. The resulting ten dimensions do not match the Big Five, neither in number nor in content.