Topic-to-Timestamp Alignment by Constrained Evidence Selection

Abstract: Meeting archives are difficult to search when users remember what was discussed but not when. We study topic-to-timestamp alignment: given a natural-language topic and a timestamped meeting transcript, the goal is to return the time at which the topic is discussed. A standard RAG setup can retrieve relevant transcript excerpts, but still asks the LLM to generate a timestamp, which can produce unsupported or invalid timecodes. We therefore recast timestamp prediction as constrained temporal candidate selection: the system retrieves timestamped transcript chunks, and the model selects the candidate that best grounds the topic instead of generating a timecode. On 420 topic-timestamp queries from 200 municipal meeting transcripts, this increases Recall@5 from 31.9% to 50.0%, reduces MAE from 837.0 seconds to 761.0 seconds with Mistral-7B-Instruct, and increases the number of parseable outputs from 373 to 419 of 420 queries. The results suggest that temporal grounding in long transcripts depends strongly on retrieval quality and output design, not only on the choice of the LLM.

• The paper presents a novel method that reframes timestamp prediction as selecting a discrete transcript chunk, effectively mitigating temporal hallucinations.
• It employs a hybrid sparse-dense retrieval and reciprocal rank fusion to identify relevant 90-second overlapping windows, improving recall and reducing errors.
• Experimental results show reduced mean absolute error and increased valid outputs, emphasizing the impact of retrieval quality over model scaling.

Topic-to-Timestamp Alignment by Constrained Evidence Selection

Problem Formulation and Motivation

The paper "Topic-to-Timestamp Alignment by Constrained Evidence Selection" (2606.20890) addresses the challenge of temporally grounding natural-language topic queries within long, timestamped meeting transcripts. Unlike prior work in meeting segmentation and summarization, the objective is to identify the precise moment a particular topic is discussed in the transcript. Conventional RAG architectures retrieve relevant spans and rely on LLMs to generate timestamps, but this process is susceptible to temporal hallucinations—LLMs frequently output plausible, but unsupported and often incorrect, timecodes. This limitation is pronounced due to the high precision required in timestamp prediction, where small numerical errors can significantly hinder usability.

Methodology

The paper recasts timestamp prediction as constrained temporal candidate selection. Instead of free-form timestamp generation, the model is presented with a set of temporally coherent transcript chunks and selects a discrete Chunk ID representing the discussion most relevant to the topic, decoupling retrieval and semantic grounding from numerical prediction.

Temporal Candidate Construction

Transcripts are partitioned into overlapping temporal windows (90 seconds with a 45-second overlap), maintaining conversational context per chunk. Each chunk retains start/end times and a synthetic identifier ([Cxx]).

Hybrid Sparse-Dense Retrieval

For each topic query, candidates are retrieved via hybrid methods: dense retrieval (BAAI/bge-large-en-v1.5) encapsulates semantic matches, while sparse BM25 retrieval preserves lexical anchors fundamental in procedural meeting language. Reciprocal Rank Fusion (RRF) [cormack2009rrf] is employed to combine ranked lists from both methods, enhancing coverage and promoting candidates supported by either retriever.

Constrained Chunk-ID Selection

The LLM receives the top-k candidate chunks and is required to select the candidate ID that best supports the topic. The output is then deterministically mapped to the start timestamp of the selected chunk. This approach mitigates temporal hallucinations, constraining the model to evidence-based output.

Experimental Evaluation

Dataset

A new benchmark based on 200 municipal meeting transcripts is introduced, comprising 600 topic-timestamp pairs, filtered to 420 queries to exclude repetitive procedural items. ASR via Whisper and diarization via pyannote ensure robust transcript annotations.

Systems Compared

• Baseline: Dense retrieval of 30-second windows, direct timestamp generation by the LLM.
• Proposed Pipeline: Hybrid RRF retrieval of 90-second overlapping windows, constrained Chunk-ID selection.

Models and Metrics

Evaluations use Mistral-7B-Instruct and several GPT-5.x variants. Metrics:

• Recall@5: Whether the gold timestamp is within top-5 retrieved chunks.
• Exact@30s: Predictions within ±30s of gold timestamp.
• MAE: Mean absolute error (seconds).
• Parsed: Proportion of queries yielding valid outputs.

Results

• Recall@5: Increased from 31.9% (baseline) to 50.0% (proposed) across both Mistral-7B-Instruct and GPT-5.4.
• MAE: Reduced from 837.0s to 761.0s (Mistral-7B-Instruct) and 1118.3s to 743.3s (GPT-5.4).
• Exact@30s: Improved moderately (30.0% Mistral-7B-Instruct, 34.3% GPT-5.4).
• Parseable Outputs: Increased substantially (373 to 419 for Mistral-7B-Instruct).

Notably, model scaling (GPT-5.4 vs Mistral-7B-Instruct) delivers minimal gains unless paired with improved retrieval and output constraint. The primary bottleneck is retrieval coverage, not model architecture. The transition to Chunk-ID selection enhances robustness and reduces outlier errors, but does not notably improve fine-grained boundary localization.

Analysis and Implications

Retrieval Quality as Critical Bottleneck

The hybrid RRF retrieval process dominates performance improvements. If the relevant segment is not among the candidates, downstream selection cannot recover. This finding underscores the necessity to prioritize retrieval optimization over LLM scale or free-form generation for temporally precise tasks.

Robustness and Error Reduction via Constrained Output

Constrained Chunk-ID selection eliminates invalid outputs and reduces large temporal errors, confirming the advantage of structured decision spaces in evidence-grounded tasks. While hallucinations are suppressed, boundary ambiguity persists due to gradual topic emergence in multiparty dialogue.

Generalizability and Dataset Constraints

The study demonstrates the effectiveness of evidence selection in municipal meetings, but results are conditioned on transcript and annotation quality. ASR and diarization errors propagate upstream and are not directly addressed. Furthermore, boundary ambiguity remains unresolved, as many relevant topics span multiple turns preceding official annotation points.

Future Directions

The paradigm of constrained evidence selection for temporal grounding invites further research:

• Generalization Across Domains: Extension to other languages, meeting types, and additional annotator agreement studies.
• Integration with Multimodal Cues: While this work is text-only, audio and video signals may further improve retrieval and boundary precision [paul2021text, zhang2023video].
• Fine-Grained Boundary Detection: Hierarchical chunking, multi-turn discourse modeling, and boundary refinement could enhance exact localization.
• Benchmark Development: Larger, multi-annotator datasets across diverse domains would provide stronger evaluation bases.

Conclusion

Topic-to-timestamp alignment is best cast as constrained evidence selection. The separation of retrieval from grounding, coupled with candidate-based output constraints, leads to tangible gains in recall and robustness, substantially reducing unsupported or invalid timestamp outputs. The findings indicate that retrieval strategies and output design outweigh model scale for temporal grounding in transcript archives. This insight has practical implications for search interfaces in long-form recordings and theoretical implications for evidence-driven generation tasks.