Dialogue-based Structured Conversational Artificial Intelligence

Systems powered by conversational Artificial Intelligence (AI) have seen a significant increase in uptake in recent years. Smart speakers and devices such as Amazon's Alexa, Apple's Siri, Google's Home and Microsoft's Cortana have become ubiquitous, as end-users look to use their voice to perform tasks and retrieve information. The recent widespread launch of platforms such as ChatGPT has further piqued public interest in more advanced conversational systems that, ostensibly, exhibit understanding and are capable of sustaining longer form conversations. Structured models of dialogue provide a framework for how an interaction (such as a conversation) between two or more participants can and should proceed. For example, a certain model might require that a "question" must always be followed by an "answer". These models have been shown to underpin novel conversational applications in focused domains, such as law, mediation and health care. There have however remained several barriers to such applications moving out of the lab and into widespread use. First, it is necessary to implement these dialogue models for new domains from scratch. This is a challenging, multi-stage problem that requires analysis by experts in dialogue structure, followed by implementation in languages with a necessarily complex syntax and steep learning curve. Second, structured models of dialogue provide an idealised account of interaction that assumes all parties will strictly adhere to the rules and as such imposes constraints to ensure this. One of these constraints is a requirement to identify exactly which type of interaction is being made (e.g., a "question" or "assertion"). While this is necessary to ensure expected dialogue flow, it reduces realism for human users, and as such has an effect on user experience. This project aims to create for the first time theories, tools and techniques that will support widespread use of structured models of dialogue in underpinning conversational systems. The project will facilitate automated implementation of such models from natural language transcripts of real examples, thus removing the need for expert annotators and developers, as well as developing new methods for more natural user interactions. In achieving the aim, the project has four objectives. The first objective is to develop theories, tools and techniques for automatically identifying the structure of a conversation from natural language transcripts. This will support the second objective, which will build the tools required to turn these identified structures into implemented structured models of dialogue. The third objective is to develop methods of supporting natural language interactions, making applications more user-friendly and intuitive. Finally, the fourth objective is to implement an open framework for conversational AI development using the project's results. This will be supported by case study applications with industry and academic partners. The project will support the development of new and innovative conversational applications, that will have benefits for academics, systems developers, commercial users, and members of the public. Academically, the project aims to support further research into the structure and composition of dialogue. For systems developers, the project will provide an open framework for the rapid development of realistic conversational systems, based on real interactions. Commercially, the more engaging paradigms provided by these applications will drive uptake and, subsequently, business growth. Finally, members of the public will benefit from applications designed to address a variety of challenges in domains such as health care, coaching, dispute mediation and law.