Pre-Conference Talk by CHENG Yu Tong | A Pruning-based Question-Answering for Interactive Video Search: A Simple Baseline

Video search systems frequently struggle with imprecise user queries, which retrieve excessively broad and noisy sets of candidate videos and reduce the effectiveness of ranking functions. This issue is compounded by the sheer volume of visually and semantically similar videos in large datasets, forcing users to browse extensively and increasing cognitive load. While interactive questioning offers a promising strategy to progressively narrow down candidate videos, dynamically generating these questions remains a significant challenge. Relying on large language models for this task often introduces high computational latency and the risk of hallucination. In this talk, I will present a simple yet highly effective alternative grounded in Shannon’s information theory. Rather than generating questions from scratch, our method mathematically selects the most discriminative concepts to compose targeted questions, allowing for rapid and reliable pruning of the search space. I will discuss the practical implementation mechanics of the system, specifically detailing dynamic search result updates and top-K rank list sampling. Furthermore, the presentation will address how our approach accounts for realistic system and user constraints, including imperfectly indexed video content and noisy or misleading user feedback. Finally, I will share our experimental findings, which demonstrate the strong retrieval performance of this method on the AVSD and TRECVid benchmarks.

This is a Pre-Conference talk for 16th ACM International Conference on Multimedia Retrieval (ICMR 2026).