EAGER: A hybrid dialogue system architecture for symbolic control of deep learning networks

在过去的15年里，技术的进步将Siri和Alexa这样的对话助手带到了整个社会。虽然这样的助手成功地支持了明确指定的任务，但他们对用户真正想要的东西的理解仍然有限，他们的答案可能是不正确的、重复的或有偏见的，而且通常很难解释他们是如何得出某个答案的。他们的成功依赖于收集大量数据和发现统计模式；然而，几乎没有数据用于以教育或提供建议为目的的对话，例如患者与医疗保健教育者的互动。这一早期的探索性研究拨款(AGER)调查了新颖的对话系统架构，这些架构将专家设计的策略与当今强大的统计模型相结合，但适用于小数据集。如果成功，这项研究将提供一种新的方法来开发具有文化能力的会话助手，用于许多用户，特别是来自代表人数较少的群体的用户寻求建议或知识的应用。这一迫切的项目将探索结合符号和神经方法的新推理机制来开发会话助手。我们将开发一种将明确设计的对话管理器模块与当今有效的神经编解码器方法相结合的体系结构。一个象征性的对话管理器将控制系统的输出，并帮助解释系统的决定。此外，它将内在地改善对大型数据集的需求，因为可以使用显性的专家洞察力，而不是成为潜在变量。同时，我们将探索如何将现代意义表示与数据扩充模型相结合来开发更大的语料库。这项研究的动机是有文化能力的患者教育，这对更好的健康结果至关重要，特别是在少数群体中。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Technological advances in the last fifteen years have brought conversational assistants like Siri and Alexa to society at large. While such assistants are successful at supporting well-specified tasks, their understanding of what the user truly wants is still limited, their answers can be incorrect, repetitive, or biased, and it is often hard to explain how they arrived at a certain answer. Their success depends on harvesting huge amounts of data and uncovering statistical patterns; however, very little data exists for conversations whose purpose is education or advice-giving, such as patient-healthcare educator interactions. This EArly Grant for Exploratory Research (EAGER) investigates novel dialogue system architectures that combine strategies designed by experts with the powerful statistical models of today, but applied to small datasets. If successful, this research will provide a new approach to develop culturally competent conversational assistants for many applications where users, especially from underrepresented groups, seek advice or knowledge.This EAGER project will explore new inference mechanisms that combine symbolic and neural approaches for the development of conversational assistants. We will develop an architecture which blends an explicitly designed dialogue manager module, with the effective neural encoder/decoder approaches of today. A symbolic dialogue manager will provide control over the outputs from the system and will help explain the system's decisions. Additionally, it will inherently ameliorate the need for a large dataset, since explicit expert insights can be used rather than emerge as latent variables. At the same time, we will explore how modern meaning representations combined with data augmentation models can be used to develop a larger corpus. The motivation for this research is culturally competent patient education, which is vital to better health outcomes especially in minority groups.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.