Open Domain Statistical Spoken Dialogue Systems

口语对话系统包含了建立有效的人机界面所需的技术，这些人机界面主要依赖于语音。到目前为止，它们大多被部署在基于电话的呼叫中心应用程序中，如银行、账单查询和旅行信息，它们是使用手工制定的规则构建的。最近苹果Siri和Google Now的推出使基于语音的界面成为主流。这些虚拟个人助理(VPA)提供了彻底改变我们与机器交互的方式的潜力，它们为正确控制和管理新兴的物联网开辟了道路-这是一个快速增长的智能设备网络，缺乏任何形式的传统用户界面。然而，目前的个人助理是使用与有限域口语对话系统相同的技术构建的。它们不能维持对话，除非在选定的有限领域内，它们已被明确编程处理。最近关于统计SD的工作表明，这种系统不仅有可能在其设计的领域内适应和提高性能，而且该系统还可能自动扩展其覆盖范围，以包括迄今为止尚未见过的新概念。这表明，应当有可能在开发有限领域统计数字统计标准方面取得的进展的基础上，设计一种全新形式的口语对话系统(以及VPA)，该系统能够扩展和适应使用，以涵盖更广泛的对话主题。这类系统的设计是本研究方案的重点，其核心思想是将最新的统计对话技术集成到一个覆盖范围广的知识图(如Free Base)中，该图不仅包含实体的本体信息，还包含可应用于这些实体的操作(如查找航班信息、预订酒店房间、购买电子书等)。要实现一个能够解释和响应每个用户请求的单一整体口语对话系统是不可行的。因此，本文提出了一种新的分布式系统体系结构，该体系结构具有三个关键特征：1.语义解码器、对话管理器和响应生成器这三个基本组件分布在知识图上，而不是简单地试图扩大现有系统的覆盖面。本质上，图中的每个节点都有能力识别它被引用的时间，并有能力做出适当的响应。当用户说话时，所有的语义解码器都在听，根据解码器输出的激活级别，话题跟踪器识别哪个概念是焦点，并激活其对话策略。所有组件都是统计的，使它们能够使用非监督调整在线自动调整。通过确保类层次结构中的顶级节点具有经过良好训练的组件来管理数据稀疏性。最初，较低级别的更专业的概念只是从它们的超类继承所需的统计模型。随着系统与用户的交互和收集的数据越来越多，底层组件获得足够的数据来训练自己的专用统计模型，最终结果是一个不断在线学习的系统。它一开始是一种有限和生硬的对话风格，但使用得越多，就变得越流利，随着用户探索新的主题，系统会学习适应和扩展其处理这些新主题的能力。由于许多用户可以同时使用该系统，因此学习可以快速并能够适应底层数据的实时更新，所有这些都是虚拟个人助理必须真正有用的特征。

Spoken Dialogue Systems (SDS) encompass the technologies required to build effective man-machine interfaces which depend primarily on voice. To date they have mostly been deployed in telephone-based call centre applications such as banking, billing queries and travel information and they are built using hand-crafted rules.The recent introduction of Apple Siri and Google Now has moved voice-based interfaces into the main-stream. These virtual personal assistants (VPAs) offer the potential to revolutionise the way we interact with machines, and they open the way to properly control and manage the emerging Internet of Things - the rapidly growing network of smart devices which lack any form of conventional user interface. However, current personal assistants are built using the same technology as limited domain spoken dialogue systems. They are not capable of sustaining conversational dialogues except within the selected limited domains which they have been explicitly programmed to handle.Very recent work on statistical SDS has demonstrated that it is not only possible for such a system to adapt and improve performance within the domain for which it has been designed but it is also possible for the system to automatically extend its coverage to include new, hitherto unseen concepts. This suggests that it should be possible to build on the progress achieved in the development of limited domain statistical SDS to design a radically new form of spoken dialogue system (and hence VPA) which is able to extend and adapt with use to cover an ever-wider range of conversational topics. The design of such a system is the focus of this research proposal.The key idea is to integrate the latest statistical dialogue technology into a wide coverage knowledge graph (such as freebase) which contains not only ontological information about entities but also the operations that can be applied to those entities (e.g. find flight information, book a hotel room, buy an ebook, etc. ).The implementation of a single monolithic spoken dialogue system capable of interpreting and responding to every conceivable user request is simply not practicable. Hence, rather than simply trying to broaden the coverage of existing SDS, a novel distributed system architecture is proposed with three key features:1. the three essential components of an SDS (semantic decoder, dialogue manager and response generator) are distributed across the knowledge-graph. In essence, every node in the graph has the capability to recognise when it is being referred to and have the capability to respond appropriately.2. when the user speaks, all semantic decoders are listening, based on the activation levels of the decoder outputs, a topic tracker identifies which concept is in focus and activates its dialogue policy.3. all components are statistical enabling them to be adapted automatically on-line using unsupervised adaptation. Data sparsity is managed by ensuring that the top level nodes in the class hierarchy have well-trained components. Initially, lower level more specialised concepts simply inherit the required statistical models from their super-classes. As the system interacts with users and more data is collected, lower level components acquire sufficient data to train their own dedicated statistical models.The end result is a system that continually learns on-line. It starts with a limited and stilted conversational style, but the more it is used, the more fluent it becomes, and as users explore new topics, the system learns to adapt and extend its capability to handle those new topics. Since many users can be using the system simultaneously, learning can be fast and capable of accommodating live updates of the underlying data, all of which are characteristics that a virtual personal assistant must have to be genuinely useful.