III: Small: Purposeful Conversational Agents based on Hierarchical Knowledge Graphs

最近，通过利用神经响应生成器和大型、多样化的训练语料库，在开放域会话代理（CA）领域取得了巨大进展，也被称为聊天机器人或社交机器人。最近的Alexa奖竞赛产生了可以就任何话题进行长时间对话的社交机器人，为有意义的使用铺平了道路。这些系统仅限于闲聊，谈话从一个话题到另一个话题，没有任何特定的目标。现在有机会（也需要）为有目的的对话开发模型，以社会目标为动力。这促成了一些令人信服的解决方案，比如聊天机器人来帮助解决心理健康问题，为老年人提供陪伴，甚至打击虚假信息。ca可以针对不同的个性、多样化和多语言人群进行配置，因此具有大规模解决社会问题的巨大潜力。实现这一目标需要在多个方面取得进展，包括道德问题和可信度。该项目解决了两个关键的技术挑战。首先，该项目消除了神经反应产生的工件，如不一致、不连贯和重复。这些反应削弱了社交机器人的可信度，从而削弱了它们的影响。其次，该项目解决了有目的对话的计算框架，特别是说服。这个项目将产生一个功能齐全的CA，它将事实知识、逻辑推理和人与人之间对话中的对话策略结合起来，在与社会问题相关的几个主题上进行连贯、有趣和有说服力的对话。有目的的对话主体必须有动机地进行有意义的交流：有说服力的对话可能会在询问、谈判或审议中循环进行，其间穿插着闲聊和轶事。为了实现这些目标，该项目结合了专门构建的多层知识图的使用，以及最先进的神经响应生成器、意图分类器和强大的自然语言理解和对话管理模块。开放知识网络的核心层由来自现有开放知识网络和因果图的整合概念和关系组成，并适当增强以促进基于知识的对话的神经符号方法。外围层构成论证图，通过对已有的说服性对话语料库进行处理，提取基本话语单位及其之间的关系，自动构建论证图。一种创新的神经符号方法是必要的(i)避免典型的“幻觉”表现出单独的神经反应发生器和（ii）智能地导航在一个有说服力的谈话的各个阶段。神经反应发生器被配置为有效地检索、推理和整合来自开放知识网络的信息，以产生事实正确、连贯、引人入胜和有说服力的反应。为了评估生成的对话的质量并衡量其影响，将使用新颖和传统的度量标准。将对代表性数据集进行自动化评估；专家还将对原型进行评估，并就未来部署的可行性提供反馈。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

There has been tremendous progress recently in the area of open domain conversational agents (CA), also referred to as chatbots or socialbots, by leveraging neural response generators and large, diverse training corpora. The recent Alexa Prize competition resulted in socialbots that can engage in prolonged conversations on any topic, paving the way for meaningful use. These systems are limited to chitchat, where the conversation meanders from topic to topic without any specific goal. There is now an opportunity (and need) to develop models for purposeful conversations, motivated by societal goals. This facilitates compelling solutions such as chatbots to assist with mental health issues, providing companionship to senior citizens, and even combating disinformation. CAs can be configured for different personalities, diverse and multilingual populations and thus hold great potential to address societal problems at scale. Realizing this goal requires advances on multiple fronts, including ethical issues and trustworthiness. This project addresses two key technical challenges. First, the project eliminates artifacts of neural response generation such as inconsistency, incoherence, and repetition. These responses diminish the trustworthiness of socialbots and hence their impact. Second, the project addresses the computational frameworks for purposeful conversations, specifically, persuasion. This project will result in a fully functional CA, that incorporates factual knowledge, logical reasoning, and conversational strategies found in human-to-human dialogue, to engage in coherent, interesting and persuasive conversations across several topics related to societal issues.Purposeful conversational agents must engage in meaningful exchanges with a motive: a persuasive conversation may cycle through inquiry, negotiation, or deliberation interspersed with chit-chat and anecdotes. In order to realize these objectives, this project incorporates the use of specially constructed multi-layered knowledge graphs, along with state-of-the-art neural response generators, intent classifiers, and a robust natural language understanding and dialogue management module. The core layers of the open knowledge network consist of consolidated concepts and relationships from existing open knowledge networks and causal graphs, suitably enhanced to facilitate neuro-symbolic approaches for knowledge grounded conversation. The peripheral layers constitute argumentation graphs which are automatically constructed through processing existing corpora of persuasive conversations and extracting elemental discourse units and relationships between them. An innovative neuro-symbolic approach is necessary to (i) avoid the typical “hallucinations” exhibited by neural response generators alone and (ii) intelligently navigate through the various stages of a persuasive conversation. The neural response generators are configured to efficiently retrieve, reason and incorporate information from the open knowledge network, in order to generate factually correct, coherent, engaging and persuasive responses. Both novel and traditional metrics will be used in order to assess the quality of the generated conversations and measure its impact. Automated evaluation will be conducted on representative data sets; the prototypes will also be evaluated by experts who will provide feedback regarding viability for future deployment.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.