MAPTRAITS: MACHINE ANALYSIS OF PERSONALITY TRAITS IN HUMAN VIRTUAL AGENT INTERACTIONS

研究结果表明，性格特征，如外向、随和和开放的经验，与人类在日常生活中遇到的能力和行为密切相关：情感表达、语言表达、人际任务的成功、领导能力、一般工作表现、教师效率、学术能力以及与技术的互动。事实上，人类用户倾向于将计算机和虚拟代理拟人化，将他们视为社会存在，并将他们的行为解释为日常的人与人之间的交互。评估人的个性的问题对于计算机中介的员工评估和培训、人机和人机交互等多个研究和商业领域非常重要。尽管人们对人格特征及其对人类生活的影响越来越感兴趣和重视，并且最近在人类行为信号(例如，声音表达和生理反应)的机器分析方面取得了进展，但专注于人格特征的机器分析的开创性努力直到最近才开始出现：(I)存在少量基于单峰线索的努力，例如书面文本/音频/语音/静态面部特征，(Ii)尽管在多模式人格特征分析、动态(持续时间、速度等)方面进行了试探性的努力。(Iii)尽管人格分析研究表明在所有人中或多或少地存在一种特质(即，一个人可以在从内向到外向的连续体上的任何地方)，但所提出的努力都没有试图在时间和空间上连续地评估人格特质(即，如何在给定的交互时间和背景下沿着多个特质维度对一个人进行评级)，以及(Iv)如何将机器(自动)特质分析用于个性化，社会的、自适应的人与虚拟智能体的交互还没有被研究过。总体而言，无论是常见的日常技术(如个人电脑、智能手机)还是人们现在使用的更复杂的系统(如计算机游戏、辅助技术、具体化虚拟智能体等)。为了解决这些问题和限制，MAPTRAITS项目将带来一套视听工具，可以在连续的时间和特征空间中根据多种非语言线索和渠道(即上半身、头部、面部、声音及其动态)动态分析和预测人类的人格特征。在12个月的时间里，没有希望建立一个完美的人格特征自动分析系统，可以应用于所有可能的应用领域。因此，作为概念验证，MAPTRAITS技术将被开发用于虚拟代理和用户个性的自动匹配，以自动建模什么类型的用户想要与什么类型的虚拟代理接触以达到增强用户参与度的目的。选择这个应用领域的动机在于它的重要性：(I)研究表明，人们对机器和对话代理的态度是基于感知到的代理的个性和他们自己的个性，以及(Ii)人类是社会存在，目前他们的日常生活围绕着与计算机、虚拟代理和机器人的交互，这些计算机、虚拟代理和机器人作为同伴、教练、智能家居的用户界面或家用机器人正变得越来越流行。

Research findings suggest that personality traits such as extraversion, agreeableness, and openness to experience, are tightly coupled with human abilities and behaviour encountered in daily lives: emotional expression, linguistic production, success in interpersonal tasks, leadership ability, general job performance, teacher effectiveness, academic ability, as well as interaction with technology. In fact, human users tend to anthropomorphise computers and virtual agents, treating them as social beings, and interpreting their behaviour similarly to daily human-human interactions.The problem of assessing people's personality is very important for multiple research and business domains such as computer-mediated staff assessment and training, human-computer and human-robot interaction. Despite a growing interest and emphasis on personality traits and their effects on human life in general, and recent advances in machine analysis of human behavioural signals (e.g., vocal expressions, and physiological reactions), pioneering efforts focusing on machine analysis of personality traits have started to emerge only recently: (i) there exist a small number of efforts based on unimodal cues such as written texts/ audio/ speech/ static facial features, (ii) despite tentative efforts on multimodal personality trait analysis, the dynamics (duration, speed, etc.) of multiple cues, which have been shown to be important in human judgments of personalities, have mostly been neglected, (iii) although personality analysis research suggests that a trait exists in all people to a greater or lesser degree (i.e. a person can be anywhere on a continuum ranging from introversion to extraversion), none of the proposed efforts have attempted to assess personality traits continuously in time and space (i.e., how a person can be rated along the multiple trait dimensions at a given interaction time and context), and (iv) how machine (automatic) traits analysis can be utilised for personalised, social and adaptive human - virtual agent interaction has not been investigated.Overall, both the common everyday technology (e.g., personal PCs, smart phones) and the more sophisticated systems people use nowadays (e.g., computer games, assistive technologies, embodied virtual agents, etc.) lack the capability of understanding their human users' personality and behaviour, and of providing socially intelligent, adaptive and engaging human - computer interaction.To address these issues and limitations, MAPTRAITS project will bring around a set of audio-visual tools that can analyse and predict human personality traits dynamically from multiple nonverbal cues and channels (i.e., upper body, head, face, voice and their dynamics) in continuous time and trait space. There is no prospect of building a perfect system for automatic analysis of personality traits that can be used in all possible application domains in 12 months' time. Therefore, as a proof-of-concept, the MAPTRAITS technology will be developed for automatic matching of virtual agent and user personalities, to automatically model what type of users would like to engage with what type of virtual agents to the aim of user engagement enhancement. The motivation for choosing this application area lies in its significance: (i) Research has shown that people's attitudes toward machines and conversational agents is based on the perceived personality of the agent, and their own personality, and (ii) humans are social beings, and currently their everyday life revolves around interacting with computers, virtual agents and robots that are getting increasingly popular as companions, coaches, user interfaces to smart homes, or household robots.