Topic-to-Timestamp Alignment by Constrained Evidence Selection

The study presents a novel approach to topic-to-timestamp alignment in meeting transcripts, focusing on constrained temporal candidate selection rather than direct timestamp generation. By employing Mistral-7B-Instruct, the method improved Recall@5 from 31.9% to 50.0% and reduced mean absolute error (MAE) from 837.0 seconds to 761.0 seconds across 420 queries. This highlights the importance of retrieval quality and output design in enhancing temporal grounding in lengthy transcripts, offering valuable insights for practitioners developing LLMs for similar applications.