Computer Vision News - April 2023

12 Congrats, Doctor Alberto! Chatbots and conversational agents have recently become part of our daily lives. They are often conceived as passive agents that reply to users’ inquiries, but it is crucial to evaluate their ability to actively ask questions to solve problems. Asking questions is essential to resolve ambiguities, gather information about the world, and identify objects mentioned in the conversation, especiallywhen chatbots canprocess visual inputs about the surroundings. FromCognitive Science, it iswell known that asking informative questions requires a plethora of high-level cognitive skills, and modeling this feature in computational models represents an important challenge. In his work, Alberto takes referential guessing dialogue games as a testbed to evaluate the question-asking skills of visually-grounded Natural Language Generation models. Using existing surface-level metrics, he shows that machine-generated dialogues are often repetitive, and improving over these metrics does not help to achieve a higher task success Alberto Testoni has recently completed his PhD at the University of Trento, supervised by Raffaella Bernardi (with him in the photo). His research focused on investigating Natural Language Generation techniques in Visual Dialogue tasks. After demonstrating that existing metrics do not represent a good proxy of dialogue quality and visual grounding skills, he proposed to evaluate dialogues on a deeper level by taking into account the strategy across dialogue turns. He explored how cognitive studies on question- asking and information-seeking strategies can inspire the development of better decoding strategies. He is now a Postdoctoral Researcher at the University of Amsterdam. Congrats, Doctor Alberto!