Zero-Shot Cue-Grounded Topic Segmentation of Spoken Documents

Suhwan Choi; Myeongho Jeon; Myungjoo Kang

cs.CLSep 28, 2026

Zero-Shot Cue-Grounded Topic Segmentation of Spoken Documents

Authors: Suhwan Choi, Myeongho Jeon, Myungjoo Kang

Organizations: Seoul National University · KAIST

Abstract

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.

Figures & tables

Dataset	$n$	Length	Description
YTSeg [ 24 ]	1,448	12.9	Creator-authored chapters
ICSI [ 14 , 6 ]	75	241.0	Research meetings
AMI [ 2 ]	129	66.0	Multi-party meetings
MeetingBank [ 13 , 6 ]	28	72.6	City-council meetings
QMSum [ 28 , 6 ]	20	28.6	Parliamentary committees
SIM [ 6 ]	100	91.6	Spliced meeting excerpts

Table 1: Evaluated corpus subsets.

n

counts documents. For each document, we calculate the average ground-truth segment length in sentences. Length reports the median across documents.

		$P_{k}$ ( $\downarrow$ )							WindowDiff ( $\downarrow$ )							Boundary $F_{1}$ ( $\uparrow$ )
	Method	YT	IC	MB	SIM	AMI	QM	Avg.	YT	IC	MB	SIM	AMI	QM	Avg.	YT	IC	MB	SIM	AMI	QM	Avg.
Classical	TextTiling [ 12 ]	.513	.716	.648	.604	.611	.630	.620	.653	.996	.994	1.00	.816	.999	.910	.433	.036	.086	.057	.121	.129	.144
Classical	BERT-TT [ 26 ] †	.402	.473	.420	.378	.420	.407	.417	.409	.514	.464	.434	.442	.432	.449	.189	.055	.133	.197	.102	.188	.144
LLM-based	LumberChunker [ 3 ]	.378	.714	.641	.604	.540	.623	.583	.473	.983	.952	.994	.732	.972	.851	.539	.121	.182	.119	.291	.233	.247
	Mackenzie et al. [ 18 ]	.349	.533	.460	.540	.459	.511	.476	.421	.722	.692	.844	.599	.712	.665	.519	.182	.293	.141	.302	.226	.277
	TOC prompt [ 7 ]	.414	.516	.472	.574	.511	.495	.497	.483	.632	.648	.840	.606	.693	.650	.453	.098	.181	.090	.158	.322	.217

Table 2: Performance on the six benchmark corpora. Bold and underline mark the best and second-best values per column. † Training-free, but uses self-supervised BERT pretraining. ‡ Def-DTS uses only valid outputs completed within the output-token limit of backbones.

Guidance	Saved example
Segment count	3–5 segments
Structural phases	Prototype Discussion : Follows the introduction; covers design ergonomics, materials, and cost-saving measures.
Boundary signals	Introduction of new documents or evaluation criteria
Segment duration	120–300 seconds

Table 3: Stage-1 guidance supplied to Stage 2: an example from an AMI meeting. Phase and signal entries are excerpts; both ranges are shown in full.

	SegmentLLM			CGS (ours)
Backbone	$P_{k}$ ( $\downarrow$ )	WD ( $\downarrow$ )	$F_{1}$ ( $\uparrow$ )	$P_{k}$ ( $\downarrow$ )	WD ( $\downarrow$ )	$F_{1}$ ( $\uparrow$ )
Gemini-3.1-FL	.234	.285	.539	.177	.207	.557
GPT-5.6-Terra	.328	.415	.525	.181	.218	.550
Qwen3.6-27B	.241	.278	.495	.175	.208	.543
Gemma-4-26B	.411	.605	.363	.221	.296	.486
Qwen3.5-9B	.309	.334	.299	.230	.279	.464

Table 4: Performance by backbone, averaged equally over the six corpora. Table 2 averages LLM results over these backbones. Bold marks the better value.

Method	$P_{k}\!\downarrow$	Input (k) $\downarrow$	Output (k) $\downarrow$	Cost $\downarrow$
LumberChunker	.583	32.36	0.825	35.59
Mackenzie et al.	.476	46.25	0.233	53.03
TOC prompt	.497	14.94	0.306	17.17
Def-DTS	.340	16.37	42.221	317.48
SegmentLLM	.307	14.03	2.428	26.13
CGS	.208	19.63	0.603	21.87

Table 5:

P_{k}

from Table 2 and recorded token use (thousands per document). Per-document means are averaged across the six corpora and six backbones. Costs are calculated from recorded token usage at standard API rates and averaged over the two proprietary backbones (USD/1,000 documents). Bold and underline mark the best and second-best values.

Configuration	Flash-Lite	Qwen-27B
CGS	.177	.175
Stage 1 only ( Forced cues )	.197	.187
Stage 2 only	.306	.285
Cue indices only	.197	.180
Count-only fallback	.192	.183
Transcript-only fallback	.249	.218

Table 6: CGS ablations (

P_{k}\downarrow

, averaged equally over six corpora;

T{=}.3

). Bold marks the lowest displayed value.

Method	Correct $\uparrow$	Missed $\downarrow$	Incorrect $\downarrow$
TextTiling	61.6	38.4	1361.8
BERT-TT	13.3	86.7	71.6
LumberChunker	79.3	20.7	750.6
Mackenzie et al.	62.4	37.6	476.2
TOC prompt	41.4	58.6	262.4
Def-DTS	46.4	53.6	351.0

Table 7: Counts are normalized to 100 ground-truth boundaries. Incorrect predictions can exceed 100 when a method predicts too many boundaries. Correct counts predictions that match a ground-truth boundary within two sentences. Each boundary is matched at most once. Missed reports how many ground-truth boundaries are not detected. Incorrect counts wrong boundary predictions. Values average six corpora and, for LLMs, six backbones.

Corpus	Method	$P_{k}\downarrow$	WD $\downarrow$	$F_{1}\uparrow$
AMI	SegmentLLM	$.278$	$.359$	$.420$
	CGS	$\mathbf{.202}$	$\mathbf{.263}$	$\mathbf{.512}$
ICSI	SegmentLLM	$.310$	$.391$	$.355$
	CGS	$\mathbf{.244}$	$\mathbf{.307}$	$\mathbf{.456}$
YTSeg	SegmentLLM	$.277$	$.335$	$.577$
	CGS	$\mathbf{.253}$	$\mathbf{.286}$	$\mathbf{.592}$

Table 8: ASR segmentation performance, averaged over six backbones. Sentence divisions differ between the original and ASR transcripts.

Explore similar work

CobSeg: Coherence Boundary Modeling for Dialogue Topic Segmentation
May 29, 2026Sijin Sun, Liangbin Zhao, Jiaxiang Cai +3Discourse Coherence
Institute of High Performance Computing, Agency for Science, Technology and Technology · Shanghai Univeristy · Fudan University
Align, Integrate, and Fire: Efficient Token-Level Alignment for Zero-Shot SpeechLLMs
Sep 16, 2026Abderrahmane Issam, Yusuf Can Semerci, Jan Scholtes +1Speech Language Models Acoustic Latent Space
Department of Advanced Computing Sciences Maastricht University
TACO: Training-free Sound Prompted Segmentation via Semantically Constrained Audio-visual CO-factorization
Dec 2, 2024Hugo Malard, Michel Olvera, Stephane Lathuiliere +1Open-Vocabulary Segmentation Neural Audio
LTCI, Télécom Paris, Institut Polytechnique de Paris

May 29, 2026, cs.CLJ/K move · Enter open · S save

CobSeg: Coherence Boundary Modeling for Dialogue Topic Segmentation

Sijin Sun, Liangbin Zhao, Jiaxiang Cai, Ming Deng+2

Institute of High Performance Computing, Agency for Science, Technology and Technology · Shanghai Univeristy · Fudan University

Discourse Coherence

Open paper AskPDF arXiv

Dialogue topic segmentation is critical in many human-AI collaborative applications which requires identifying heterogeneous boundary cues, including lexical transitions near utterance edges and semantic discontinuities across utterances. Existing utterance models often dilute these local lexical signals. We propose CobSeg, a novel multi-branch architecture that separates coherence-level semantic continuity from lexical boundary transitions and recovers both through directional boundary prediction. CobSeg further uses boundary informativeness weighting to emphasize high-utility utterance positions, and incorporates a corpus-derived topic coherence cue with learned combination weights. While CobSeg is evaluated as a compact trainable segmenter under supervised gold-boundary training and a pseudo-label setting with automatically induced boundaries, it performs enhanced boundary prediction without LLM calls during inference. Across five benchmarks, it improves

P_k

and

W_d

particularly when local lexical cues are prominent: under gold supervision, it reduces

P_k

by 0.7 points and

W_d

by 0.6 points on VHF, and reaches

P_k

of 1.0 on DialSeg711; with induced boundaries, it reduces

P_k

by 14.8 points on VHF, by 1.5 points on DialSeg711, and by 1.1 points on TIAGE, outperforming prior non-LLM approaches.

Abstract

Explore similar work

CobSeg: Coherence Boundary Modeling for Dialogue Topic Segmentation

Align, Integrate, and Fire: Efficient Token-Level Alignment for Zero-Shot SpeechLLMs

TACO: Training-free Sound Prompted Segmentation via Semantically Constrained Audio-visual CO-factorization