How to Win Arguments - Empowering Virtual Agents to Improve their Persuasiveness

演讲者是否被认为是可信的，不仅取决于他们的演讲内容，而且在很大程度上取决于信息的传达方式。人类的交流不仅以言语为基础，还包括其他渠道，如手势、姿势、面部表情和凝视，这些都影响着听众的感知。在EVA项目中，我们将研究公共辩论中的典型会话模式，包括论点的内容和结构，以及它们如何与听众交流。此外，我们将包括战术对话行为（所谓的闪避动作），例如在真实辩论中经常使用的话题变化或借口。为此，我们将通过具身会话代理模拟人类之间的论证对话。作为一个应用领域，我们关注的是政治话语。由于政治辩论对塑造人们的思想和态度具有巨大的影响，因此为人们提供工具，帮助他们通过以不同形式提出论点来探索论证空间，从而支持他们的决策过程，这一点尤为重要。具身会话代理的使用是一种视听形式的表示，在论证挖掘领域探索较少。它允许我们以一种直观的方式提出论点，并揭示辩论中理性和非理性因素的影响。智能体的语言和非语言行为将结合使用基于规则的方法（由论证理论提供信息）和数据驱动的方法（由人类之间多模态辩论的语料库提供信息）来确定。虚拟代理的参数将自动从使用参数挖掘技术创建的本体中提取出来。我们将依靠强化学习（RL）来优化智能体在与模拟对手互动时的论证策略。EVA提案是在优先方案健全论证机（RATIO）框架内编制的。它解决了RATIO的以下核心问题：如何将论点直观地呈现给用户，以支持决策过程？EVA项目具有跨学科的特点。它将结合多模态行为综合和分析、论证挖掘和对话管理的研究。

Whether speakers are perceived as credible depends not only on the contents of their speech, but also to a large extent on how a message is conveyed. Human communication is not only based on speech, but also comprises other channels, such as gestures, postures, facial expressions and gaze, all of which influence the perception of the audience. In the EVA project, we will investigate typical conversational patterns in public debates including the content and structure of arguments, but also how they are communicated to an audience. Moreover, we will include tactical dialogue acts (so called dodge moves), such as a change of topics or subterfuges, that are often used in real debates. To this end, we will simulate argumentation dialogues between humans through embodied conversational agents. As an application domain, we focus on political discourse. Since political debates have an enormous influence on shaping people's mind and attitudes, it is of particular importance to provide people with tools that help them explore the argumentation space by presenting arguments in different forms and thus support their decision-making processes. The use of embodied conversational agents is an audio-visual form of presentation that has been less explored in the area of argumentation mining. It allows us to present arguments in a way that is intuitive and reveals the effect of rational and non-rational elements in a debate. The verbal and nonverbal behaviours of the agents will be determined using a combination of a rule-based approach that is informed by theories on argumentation and a data-driven approach that is informed by corpora of multimodal debates between humans. The arguments for the virtual agents will be automatically extracted from an ontology that will be created using argument mining techniques. We will rely on Reinforcement Learning (RL) to optimize the agents' argumentation strategies in an interaction with a simulated opponent. The EVA proposal has been prepared within the Priority Programme Robust Argumentation Machines (RATIO). It addresses the following core question of RATIO: How can arguments be presented intuitively to users in order to support decision-making processes? The EVA project has an interdisciplinary character. It will combine research on multimodal behaviour synthesis and analysis, argumentation mining and dialogue management.