Dr. Jinho Choi gave a seminar to the Linguistics Colloquium at Seoul National University (SNU). Thanks to Dr. Hyopil Shin for the invitation.
Towards Human-like Conversational AI: A Cognition-oriented Framework
The remarkable prowess of large language models (LLMs) has spurred a surge of interest in end-to-end models for dialogue systems. While these models excel in performance, they overlook the intricate cognitive processes inherent in human conversation, relying solely on pattern matching. In this talk, I propose a Cognition-oriented Framework for Conversational AI, designed to emulate human cognition in conversations, with two novel models targeting commonsense reasoning and dialogue state generation. First, I present ConvoSense, a comprehensive dataset comprising over 500,000 inferences across 10,000 dialogues. Leveraging ConvoSense, we develop an inference model and adapt it as a cognition module to augment open-chat dialogue quality, surpassing the performance of established LLMs such as GPT-4. Next, I introduce DS5K, a diverse dataset featuring 5,000 dialogues across 1,000 domains, and a dialogue state tracking model, showing competitive performance to previous approaches using 13 times larger LLMs. Leveraging DS5K, we develop a dialogue state generation model, serving as another cognition module, to extract pertinent information from dialogue contexts to construct a memory structure, enabling Conversational AI to effectively engage in long-term, multi-sessional conversations. Finally, I present a diagnostic model capable of conducting clinical interviews with trauma patients, facilitating personalized treatment decisions with minimal resources. This innovative approach exemplifies the potential of AI to revolutionize mental healthcare by delivering cost-effective, data-driven interventions.