Distinguishing semantically similar queries in temporal video grounding via LLM-generated query

With the rapid growth of video content and increasing demands for semantic understanding in downstream applications, Video Temporal Grounding (VTG) and Highlight Detection (HD) have garnered significant attention. Both tasks aim to accurately localize relevant segments in long videos based on natural language queries and assess their semantic alignment with the query (i.e., saliency score). Although existing methods achieve promising results under conventional metrics, we observe that current benchmarks lack sufficiently challenging samples, and prevailing evaluation protocols fail to reflect model robustness and generalization under complex semantic conditions. To address these limitations, we propose a Natural Language Query Generation (NLQG) framework powered by Large Language Models (LLM). To enhance training effectiveness, our framework generates hard negative samples that are semantically similar but mismatched, as well as paraphrased queries that retain semantic equivalence. Additionally, we introduce new evaluation metrics designed to more accurately assess model robustness and generalization in semantically complex scenarios. Extensive experiments demonstrate that models trained with our approach not only maintain strong performance on standard metrics but also achieve substantial improvements in discriminative power and generalization under challenging conditions-thereby better aligning with the practical demands of real-world video understanding systems.