Uploaded June 2023 | Updated September 2026, 4 days ago
Neural dialogue systems have made impressive advancements in natural language understanding and generation, yet challenges in reliability and continual improvement persist. In this talk, I will present my work focused on improving the reliability of dialogue systems while emphasizing continual enhancement and improved control. Specifically, I will discuss two key components: DialGuide, a framework for aligning dialogue model behavior with developer guidelines, and target-guided response generation, which enables dialogue systems to smoothly transition conversations towards specific targets. DialGuide provides natural language rules and guidelines to control dialogue model behavior, ensuring safe and engaging responses
that align with developer guidelines. Target-guided response generation, which empowers dialogue systems to guide conversations towards desired targets, leverages common-sense knowledge and data augmentation techniques to create smooth transitions from the dialogue context to the target sentence. I will also discuss future directions that encompass feedback and interaction, alignment and constitutional AI, and improving pragmatics to further advance the capabilities of next-generation dialogue systems.
Prakhar is a PhD student at Language Technologies Institute, Carnegie Mellon University. His research focuses on improving reliability of natural language generation and dialogue systems, including improving the control, safety and evaluation of generation systems. He completed his undergraduate education in Computer Science at Indian Institute of Technology, Roorkee. Before joining CMU, he worked as a research
associate at Adobe Research India.
Neural dialogue systems have made impressive advancements in natural language understanding and generation, yet challenges in reliability and continual improvement persist. In this talk, I will present my work focused on improving the reliability of dialogue systems while emphasizing continual enhancement and improved control. Specifically, I will discuss two key components: DialGuide, a framework for aligning dialogue model behavior with developer guidelines, and target-guided response generation, which enables dialogue systems to smoothly transition conversations towards specific targets. DialGuide provides natural language rules and guidelines to control dialogue model behavior, ensuring safe and engaging responses
that align with developer guidelines. Target-guided response generation, which empowers dialogue systems to guide conversations towards desired targets, leverages common-sense knowledge and data augmentation techniques to create smooth transitions from the dialogue context to the target sentence. I will also discuss future directions that encompass feedback and interaction, alignment and constitutional AI, and improving pragmatics to further advance the capabilities of next-generation dialogue systems.
Prakhar is a PhD student at Language Technologies Institute, Carnegie Mellon University. His research focuses on improving reliability of natural language generation and dialogue systems, including improving the control, safety and evaluation of generation systems. He completed his undergraduate education in Computer Science at Indian Institute of Technology, Roorkee. Before joining CMU, he worked as a research
associate at Adobe Research India.
