Source-linked AI summary
Zero-Shot Cue-Grounded Topic Segmentation of Spoken Documents
Suhwan Choi, Myeongho Jeon, Myungjoo Kang
TL;DR
Existing LLM segmenters can struggle to match the wide variation in spoken-document topic granularity. CGS grounds boundaries in topic-opening cues and uses structure-guided semantic fallback when those cues are insufficient. Across six corpora and six backbones, it outperforms evaluated baselines on all metrics and retains an advantage on noisy ASR transcripts.
Problem
Spoken documents vary widely in topic granularity, while existing LLM segmenters may produce sections too short or too long for a target corpus and do not explicitly ground boundaries in topic-opening cues.
Source: §1 · p. 1Method
CGS uses indexed topic-opening phrases as boundaries when cues suffice and otherwise applies semantic segmentation guided by inferred document structure.
Source: §1 · p. 2Results
Takeaways & Limitations
CGS offers a practical approach to spoken-document segmentation without task-specific training, combining accurate segmentation with low API cost.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Topic segmentation structures spoken documents into coherent sections, facilitating navigation and downstream understanding. The appropriate granularity can vary substantially, ranging from broad thematic shifts to fine-grained subtopics. Existing LLM-based segmenters, however, often struggle to adapt to this variation, causing them to either merge distinct subtopics or over-segment coherent themes. To address this, we introduce Cue-Grounded Segmentation (CGS), a training-free framework that operates without any task-specific supervision. CGS first identifies phrases that explicitly signal the start of a new topic and uses their sentence positions as segment boundaries. When such cues are insufficient, it falls back to semantic segmentation, guided by the document structure inferred during cue extraction. Across six benchmarks and six LLM backbones, CGS consistently outperforms existing baselines, remains robust to noisy ASR transcripts, and achieves these gains with low API cost on proprietary models.
1. INTRODUCTION
Spoken documents lack explicit section headings, while their topic structures vary widely and existing segmenters can mistake minor shifts for new topics or miss meaningful changes. CGS grounds boundaries in topic-opening cues and uses structure-guided semantic segmentation when those cues are insufficient.
1. INTRODUCTION
- Existing LLM segmenters predict boundaries without explicitly grounding them in topic-opening cues, risking over-segmentation of minor shifts or under-segmentation of meaningful topic changes.Spoken documents rarely provide headings, so boundaries must be inferred from discourse.Source: §1 · p. 1
- Ground-truth median segment lengths range from 12.9 sentences in chaptered YouTube videos to 241 in research-group meetings, an 18.6-fold difference.Despite this variation, existing LLM segmenters often produce sections that are too short or too long for the target corpus.Source: §1 · pp. 1–2, Figure 1 · p. 1
- CGS uses topic-opening phrases as sentence-position boundaries when cues are informative, then falls back to semantic segmentation guided by inferred document structure.This design is intended to adapt section lengths to the spacing between topic openings.Source: §1 · p. 2
- Across six benchmarks and six LLM backbones, CGS achieves the best average performance on all metrics, with low API cost on proprietary models.Ablations report benefits from combining cue extraction with semantic fallback and supplying segment-count guidance.Source: §1 · p. 2
2. METHODOLOGY
CGS first extracts indexed topic-opening cues and judges their reliability; when cues require fallback, it supplies inferred structure and a segment-count range to semantic segmentation. The direct route uses valid cue indices as boundaries, while fallback predicts topic starts from transcript content.
2. METHODOLOGY
- CGS uses two calls to the same LLM with shared prompts across datasets, without dataset-specific tuning or labeled examples.It first extracts transition cues and falls back to semantic segmentation guided by inferred document structure when cues are insufficient.Source: §2 · p. 2
2.1. Stage 1: Indexed Transition Evidence
- Stage 1 quotes topic-opening phrases with sentence indices, labels cue reliability, and requests fallback for unreliable cues or an explicit fallback decision.It also predicts a segment-count range and, when fallback is required, supplies phases, boundary signals, and duration estimates.Source: §2.1 · p. 2
- On the direct route, CGS discards out-of-range cue indices and uses the remaining cue positions, together with sentence 0, as boundaries.The cue-derived boundary set is Bcue = {0} ∪ {i : (i, q) ∈ C}.Source: §2.1 · p. 2
2.2. Stage 2: Conditioned Semantic Fallback
- Stage 2 reads the full transcript with Stage-1 structural guidance and predicts topic starts from semantic content, without receiving the Stage-1 cue list.The suggested segment-count range guides rather than strictly constrains predictions, and neighboring boundaries at most two sentences apart are merged.Source: §2.2 · pp. 2–3
3. EXPERIMENTAL SETUP
The evaluation uses six benchmarks with fixed prompts and settings, and compares classical, representation-based, and LLM segmentation baselines. It reports Pk, WindowDiff, and Boundary F1, with results averaged across documents and corpora as described.
3. EXPERIMENTAL SETUP
- The six benchmarks cover 1,800 documents; supplied transcripts are indexed by sentence or utterance, and evaluated configurations keep prompts and settings fixed across corpora.No model is fine-tuned for segmentation or given labeled examples from target corpora.Source: §3 · p. 3
- Evaluation uses Pk, WindowDiff, and Boundary F1, with two-sentence tolerance for a correct boundary match and macro scores averaged equally across corpora.Pk and WindowDiff use k = max(1, floor of half the document’s mean ground-truth segment length).Source: §3 · p. 3
- Baselines include lexical-cohesion and contextual-embedding methods plus LLM approaches based on table-of-contents planning, recursive splitting, chunking, grouping, or utterance intent.Def-DTS generates intermediate outputs for every utterance, which can lead to substantial output-token use on long transcripts.Source: §3 · p. 3
- The study evaluates six LLM backbones, and Table 4 reports performance by backbone averaged equally over the six corpora.The evaluated models include proprietary API models and open-weight models.Source: §3 · p. 3, Table 4 · p. 3
4. RESULTS AND ANALYSIS
CGS outperforms the evaluated baselines across corpora and backbones, while combining low segmentation error with low API cost. Ablations, boundary counts, and ASR tests characterize its component contributions and robustness.
Overall performance and cost
- Across six backbones, CGS outperforms SEGMENTLLM on all three macro metrics for each backbone and all six published baselines on every corpus under all three metrics.CGS completes segmentation after Stage 1 for 80–93% of documents.Source: §4 · p. 3, Table 5 · p. 4
- CGS combines the lowest segmentation error with the second-lowest API cost and generates about one-quarter as many output tokens as SEGMENTLLM.On proprietary backbones, output tokens cost six times more than input tokens, allowing output savings to offset additional input cost.Source: §4 · pp. 3–4, Table 5 · p. 4
CGS ablations
- CGS performs best among the ablation variants on both evaluated backbones, including comparisons with single-stage and unquoted-cue variants.Stage 1-only forces cue boundaries without semantic fallback; Stage 2-only omits Stage 1, while cue-indices-only removes the quoted phrases.Source: §4 · p. 4, Table 6 · p. 4
Stage-1 guidance ablation
- Retaining the suggested segment-count range provides most of the improvement over transcript-only fallback; structural guidance further reduces Pk, especially on Qwen-27B.Count-only fallback omits phase descriptions, boundary signals, and duration estimates; transcript-only fallback also omits the count range.Source: §4 · p. 4
Boundary errors
- CGS detects about as many ground-truth boundaries as SEGMENTLLM while making the fewest incorrect predictions overall.Table 7 normalizes counts to 100 ground-truth boundaries; incorrect predictions can exceed 100.Source: §4 · p. 4, Table 7 · p. 4
ASR transcripts
- On ASR transcripts, CGS outperforms SEGMENTLLM on all metrics for all three corpora, averaged across six backbones.The evaluation uses released Whisper-Large transcripts for YTSeg and Whisper large-v3 transcripts for AMI and ICSI.Source: §4 · p. 4, Table 8 · p. 4
5. CONCLUSION
The conclusion presents CGS as a practical approach to spoken-document segmentation, combining quoted topic-opening cues with selective structure-guided semantic fallback. It reports baseline advantages across six corpora and six backbones, benefits from segment-count guidance, and retained advantage on noisy ASR transcripts.
Conclusion
- CGS combines quoted topic-opening cues with selective structure-guided semantic fallback and outperforms evaluated baselines across six corpora with each of six backbones.The conclusion also reports benefits from segment-count guidance, robustness to noisy ASR transcripts, and low API cost.Source: §5 · p. 4
7. COMPLIANCE WITH ETHICAL STANDARDS
The study uses publicly released research corpora under their respective terms and collects no new human-subject data.
Compliance with Ethical Standards
- The study uses publicly released research corpora under their respective terms and collects no new human-subject data.Source: §7 · p. 5