As technology continues to advance, the incorporation of artificial intelligence (AI) into our daily lives has become increasingly prevalent. One notable manifestation of this trend is the emergence of AI assistants, which serve a myriad of roles - from virtual personal assistants managing routine tasks, to social companion chatbots for entertainment, to digital consultants for recommendations. Yet, ensuring the trustworthiness of these AI assistants remains a critical challenge. They may produce irrelevant, non-factual, hallucinated, impolite, and even harmful responses, posing significant risks to their deployment in real-world applications. In this talk, I will share my research on building safe, user-centric AI assistants, with a particular focus on accurately modeling user intents and states. The underlying principle is that only by accurately capturing user needs can these AI assistants take right actions and provide reliable responses. Specifically, my research is centered around three key aspects: 1) user state tracking to maintain a structured and accurate representation of user needs; 2) user question rewriting to address ambiguities and ellipses in user inputs; and 3) user satisfaction estimation to evaluate user engagement and prevent unsatisfactory responses.

Speaker: Mr Fanghua YE
Date: 23 May 2024 (Thursday)
Time: 2:45pm – 3:45pm
Poster: Click here