From Noise to Signal: When Outliers Seed New Topics

Evangelia Zve; Gauvain Bourgne; Benjamin Icard; Jean-Gabriel Ganascia

cs.CLMar 18, 2026

From Noise to Signal: When Outliers Seed New Topics

Authors: Evangelia Zve, Gauvain Bourgne, Benjamin Icard, Jean-Gabriel Ganascia

Organizations: Sorbonne Université, CNRS, LIP6, France · Infopro Digital, France

Abstract

Outliers in dynamic topic modeling are typically treated as noise, yet we show that some can serve as early signals of emerging topics. We introduce a temporal taxonomy of news-document trajectories that defines how documents relate to topic formation over time. It distinguishes anticipatory outliers, which precede the topics they later join, from documents that either reinforce existing topics or remain isolated. By capturing these trajectories, the taxonomy links weak-signal detection with temporal topic modeling and clarifies how individual articles anticipate, initiate, or drift within evolving clusters. We implement it in a cumulative clustering setting using document embeddings from eleven state-of-the-art language models and evaluate it retrospectively on HydroNewsFr, a French news corpus on the hydrogen economy. Inter-model agreement reveals a small, high-consensus subset of anticipatory outliers, increasing confidence in these labels. Qualitative case studies further illustrate these trajectories through concrete topic developments.

Figures & tables

Symbol	Definition
$T_{\!A}$	Document appearance time in the stream
$T_{\!T}$	Topic creation time
$T_{\!I}$	Document integration time
$T_{\!\text{ final}}$	Final time window in the corpus
$\theta_{\!\text{ delay}}$	Delay cutoff separating persistent and recent outliers

Table 1: Temporal notation used in the taxonomy

Symbol	Description	Condition
$\mathcal{T}$ (integrated into $\mathcal{T}$ opic)
$\mathcal{TO}$ ( $\mathcal{O}$ utlier integrated into $\mathcal{T}$ opic)
$\mathcal{TOA}$ ( $\mathcal{A}$ nticipatory $\mathcal{O}$ utlier integrated into $\mathcal{T}$ opic)
$\mathcal{TOA}_{\!\text{ first}}$	at topic creation	$T_{\!A}<T_{\!I}=T_{\!T}$
$\mathcal{TOA}_{\!\text{ late}}$	after topic creation	$T_{\!A}<T_{\!T}<T_{\!I}$
$\mathcal{TOD}_{\!\text{ late}}$ ( $\mathcal{D}$ rift $\mathcal{O}$ utlier integrated into $\mathcal{T}$ opic )		$T_{\!T}<T_{\!A}<T_{\!I}$

Table 2: Formal conditions for the taxonomy cases

Panel of Figure 2 — Figure 2: Overview of the seven taxonomy cases. Panels (a–g) show temporal relations between $T_{\!A}$ , $T_{\!T}$ , $T_{\!I}$ , $\theta_{\!\text{ delay}}$ , $T_{\!\text{ final}}$ . Colored belts indicate topic activity and outlier phases.

Figure 3: Temporal distribution of HydroNewsFr over the collection period (20 March–8 June 2025). Bars show daily publication counts; the stepped line and hatched area show cumulative totals.

Model	Dimensions	Language
sentence-camembert-base	768	French
Solon-embeddings-large-0.1	1024	French
paraphrase-..-MiniLM-L12-v2	384	Multilingual
paraphrase-..-mpnet-base-v2	768	Multilingual
LaBSE	768	Multilingual
multilingual-e5-large	1024	Multilingual

Table 3: The eleven embedding models.

Model	HDBSCAN							OPTICS
Model	2D	3D	5D	10D	20D	30D	40D	2D	3D	5D	10D	20D	30D	40D
bge-m3	0.621	0.611	0.589	0.618	0.583	0.602	0.615	0.634	0.615	0.610	0.592	0.637	0.602	0.638
sentence-camembert-base	0.582	0.579	0.568	0.562	0.574	0.559	0.561	0.612	0.565	0.585	0.551	0.600	0.571	0.578
gemini	0.601	0.581	0.575	0.576	0.574	0.578	0.565	0.633	0.605	0.619	0.598	0.626	0.621	0.602
multilingual-e5-large	0.624	0.604	0.616	0.603	0.617	0.601	0.608	0.635	0.624	0.612	0.597	0.620	0.620	0.611
mistral-embed	0.638	0.628	0.653	0.641	0.636	0.636	0.630	0.638	0.643	0.636	0.606	0.620	0.628	0.617

Table 4: Average silhouette scores per model and UMAP dimension. Best score per dimension is bolded.

Figure 4: Cumulative clustering over the first 9 time windows with mistral-embed and 2D UMAP . Colors indicate topics; newly assigned documents are larger and more opaque; black $\times$ denote outliers.

Figure 5: Empirical survival curve $S(t)=P(\Delta T>t)$ for integration delays, pooled across models and configurations. Dashed lines mark quantiles; the 90th percentile ( $p_{90}=26$ days) defines $\theta_{\text{delay}}$ .

Algorithm	Task	Mean Fleiss’ $\kappa$	Range $\theta_{\!\text{ align}}$	Model Range
HDBSCAN	Case	0.46	0.46–0.47	0.36–0.51
HDBSCAN	Delay	0.36	0.36–0.37	0.20–0.43
OPTICS	Case	0.37	0.37–0.37	0.34–0.41
OPTICS	Delay	0.30	0.30–0.31	0.23–0.33

Table 5: Mean Fleiss’

\kappa

on anticipatory outliers (

\mathcal{TOA}

) across all models and alignment thresholds

\theta_{\!\text{ align}}\in[0.2,0.7]

. “Range

\theta_{\!\text{ align}}

” shows variability across

\theta_{\!\text{ align}}

for fixed models, while “Model Range” shows variability across models for fixed

\theta_{\!\text{ align}}

Table 10

Figure 6: Distribution of taxonomy cases per model for HDBSCAN , UMAP 20D, and $\theta_{\!\text{ align}}=0.30$ .

Figure 7: Per-model share of anticipatory outliers ( $\mathcal{TOA}$ ) among all outlier cases ( $\mathcal{TO}$ , $\mathcal{O}$ )

Panel of Figure 8 — Figure 8: Proportion of articles for which at least $N$ models agree on (a) $\mathcal{T}$ and (b) $\mathcal{TOA}$ cases.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Model	$\theta_{\!\text{ align}}=0.2$		$\theta_{\!\text{ align}}=0.3$		$\theta_{\!\text{ align}}=0.4$		$\theta_{\!\text{ align}}=0.5$		$\theta_{\!\text{ align}}=0.6$		$\theta_{\!\text{ align}}=0.7$
	$\kappa$	$\kappa_{\textit{delay}}$	$\kappa$	$\kappa_{\textit{delay}}$	$\kappa$	$\kappa_{\textit{delay}}$	$\kappa$	$\kappa_{\textit{delay}}$	$\kappa$	$\kappa_{\textit{delay}}$	$\kappa$	$\kappa_{\textit{delay}}$
HDBSCAN
bge-m3	0.51	0.38	0.51	0.38	0.51	0.38	0.51	0.38	0.51	0.38	0.51	0.38
sentence-camembert-base	0.47	0.32	0.47	0.32	0.47	0.32	0.47	0.32	0.47	0.32	0.47	0.32
gemini	0.47	0.43	0.47	0.43	0.47	0.43	0.47	0.43	0.47	0.43	0.45	0.41
multilingual-e5-large	0.49	0.42	0.49	0.42	0.49	0.42	0.49	0.42	0.49	0.42	0.49	0.42

Appendix

Table 9: Per-model Fleiss’

\kappa

for agreement across UMAP dimensionalities, shown for alignment thresholds

\theta_{\!\text{ align}}\in[0.2,0.7]

Topic Modeling

Topic modeling discovers recurring themes and document-topic structure in text, helping organize, compare, and track large collections.

See all 58 papers

Predicting Emerging Topics from Outliers: A Prospective Study of Weak Signals in Embedding Space
Sep 24, 2026Evangelia Zve, Gauvain Bourgne, Jean-Gabriel GanasciaTopic Modeling Anomaly Detection
LIP6, Sorbonne Université, CNRS · Infopro Digital
Retrieving Floods without Floodlights: Topic Models as Binary Classifiers for Extreme Climate Events in German News
May 5, 2026Brielen Madureira, Mariana Madruga de Brito, Andreas NieklerTopic Modeling Text Classification
LeipzigLab - Climate Discourse, Leipzig University, Germany · Helmholtz Centre for Environmental Research - UFZ, Germany · Leipzig University, Germany +1
TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes
Feb 22, 2026Roman EggerTopic Modeling
Smartvisions.at & Modul University Vienna

Sep 24, 2026J/K move · Enter open · S save

Predicting Emerging Topics from Outliers: A Prospective Study of Weak Signals in Embedding Space

Evangelia Zve, Gauvain Bourgne, Jean-Gabriel Ganascia

LIP6, Sorbonne Université, CNRS · Infopro Digital

Topic Modeling Anomaly Detection

Open paper AskPDF arXiv

Some documents that embedding-based topic models initially classify as noise later become founding members of emerging topics. At publication time, however, they appear as scattered points in embedding space and are difficult to distinguish from ordinary noise without the benefit of hindsight. We study whether such anticipatory outliers can be predicted prospectively, using only information available when a document first appears. We derive labels from the subsequent trajectories of outlier documents, distinguishing those that anticipate new topics from those that reinforce existing topics or remain isolated, and estimate label confidence through agreement across multiple embedding models. On two French news corpora, anticipatory outliers prove predictable at publication time. Under cross-validation,

F_1

rises from about 0.77 over the full eligible population to above 0.90 on high-consensus subsets, and remains at 0.76-0.80 under a strictly chronological evaluation. Predictive performance is driven mainly by geometric features capturing each outlier's position in embedding space.

Abstract

Topic Modeling

Explore similar work

Predicting Emerging Topics from Outliers: A Prospective Study of Weak Signals in Embedding Space

Retrieving Floods without Floodlights: Topic Models as Binary Classifiers for Extreme Climate Events in German News

TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes