CobwebTM: Probabilistic Concept Formation for Lifelong and Hierarchical Topic Modeling

Karthik Singaravadivelan; Anant Gupta; Zekun Wang; Christopher J. MacLellan

cs.CLApr 15, 2026

CobwebTM: Probabilistic Concept Formation for Lifelong and Hierarchical Topic Modeling

Authors: Karthik Singaravadivelan, Anant Gupta, Zekun Wang, Christopher J. MacLellan

Organizations: College of Computing, Georgia Institute of Technology, Atlanta, GA 30332 USA

Abstract

Topic modeling seeks to uncover latent semantic structure in text corpora with minimal supervision. Neural approaches achieve strong performance but require extensive tuning and struggle with lifelong learning due to catastrophic forgetting and fixed capacity, while classical probabilistic models lack flexibility and adaptability to streaming data. We introduce CobwebTM, a low-parameter lifelong hierarchical topic model based on incremental probabilistic concept formation. By adapting the Cobweb algorithm to continuous document embeddings, CobwebTM constructs semantic hierarchies online, enabling unsupervised topic discovery, dynamic topic creation, and hierarchical organization without predefining the number of topics. Across diverse datasets, CobwebTM achieves strong topic coherence, stable topics over time, and high-quality hierarchies, demonstrating that incremental symbolic concept formation combined with pretrained representations is an efficient approach to topic modeling.

Figures & tables

Panel of Figure 1 — Figure 1: A visualization of three levels of the hierarchy induced by CobwebTM . For each node, we display the top five representative words extracting using the c-tf-idf procedure described in Section 3.2.1 . Words that appear in multiple nodes at the same level are underlined to highlight shared semantic content across sibling topics.

Method	${C_{v}}_{f}$	$\Delta C_{v}$ %	ARI	TCD	ISIM f
CobwebTM (ours)	0.741	110.82	0.915	0.000	0.191
DBSTREAM	0.602	71.55	0.594	0.202	0.200
LifelongDocNADE	0.281	$-11.61$	0.479	0.369	0.194
MiniBatchKMeans	0.413	$-0.89$	0.952	0.236	0.215
OnlineLDA	0.317	7.06	0.948	0.033	0.216
Refit-HDBSCAN	0.673	7.35	0.944	0.053	0.196

Table 1: Lifelong topic modeling results on TweetNER . Best in bold, second-best underlined.

Method	${C_{v}}_{f}$	$\Delta C_{v}$ %	ARI	TCD	ISIM f
CobwebTM (ours)	0.613	26.09	0.997	0.000	0.207
DBSTREAM	0.457	$-3.67$	0.012	0.155	0.188
LifelongDocNADE	0.184	$-75.07$	0.122	0.684	0.206
MiniBatchKMeans	0.423	$-9.67$	0.955	0.245	0.176
OnlineLDA	0.520	10.71	0.897	0.029	0.200
Refit-HDBSCAN	0.425	1.81	0.735	0.050	0.157

Table 2: Lifelong topic modeling results on Stack Overflow Dataset. Best in bold, second-best underlined.

Method	${C_{v}}_{f}$	$\Delta C_{v}$ %	ARI	TCD	ISIM f
CobwebTM (ours)	0.796	57.51	0.984	0.000	0.208
DBSTREAM	0.418	$-10.29$	0.408	0.193	0.191
LifelongDocNADE	0.238	$-49.49$	0.127	0.462	0.196
MiniBatchKMeans	0.646	40.37	0.962	0.203	0.196
OnlineLDA	0.422	0.99	0.882	0.053	0.206
Refit-HDBSCAN	0.657	19.11	0.891	0.067	0.193

Table 3: Lifelong topic modeling results on Spatiotemporal News Dataset. Best in bold, second-best underlined.

Figure 2:

C_{v}

comparison on the StackOverflow, Spatiotemporal News, and TweetNER dataset.

Figure 3: ARI comparison on the StackOverflow, Spatiotemporal News, and TweetNER dataset.

Figure 4: TCD comparison on the StackOverflow, Spatiotemporal News, and TweetNER dataset.

Figure 5: Intruder Similarity (ISIM) comparison on the StackOverflow, Spatiotemporal News, and TweetNER dataset.

Dataset	20 News Groups			AG News			Stack Exchange
Method	NPMI	PCC	SD	NPMI	PCC	SD	NPMI	PCC	SD
BERTopic-HDBSCAN	0.161	0.045	0.600	0.101	0.031	0.913	0.130	0.036	0.891
BERTopic-KMeans	0.169	0.091	0.875	0.067	0.005	0.856	0.125	0.041	0.940
TraCo	0.155	-0.038	0.941	-0.107	-0.043	0.862	-0.040	-0.156	0.913
BoxTM	0.047	0.055	0.993	0.002	0.005	0.998	0.019	0.022	0.996
CobwebTM (ours)	0.206	0.141	0.958	0.108	0.027	0.942	0.131	0.073	0.959

Table 4: Comparison of hierarchical topic-modeling metrics across datasets.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	StackExchange
Model	None	16	128	512
COBWEBTM (ours)	0.6131	0.6331	0.6227	0.6383
BERTopic (MiniBatchKMeans)	0.4234	0.4377	0.4202	0.4393
BERTopic Refit (HDBSCAN)	0.4249	0.4829	0.4666	0.4713
BERTopic Refit (KMeans)	0.5507	0.5068	0.5110	0.5123
Lifelong NTM	0.1840	0.1621	0.1868	0.2057

Appendix

Table 5: Final C v by UMAP dimensionality on StackExchange.

Dataset	TweetNER
Model	None	16	128	512
COBWEBTM (ours)	0.7414	0.7249	0.7581	0.7048
BERTopic (MiniBatchKMeans)	0.4129	0.3531	0.3987	0.3807
BERTopic Refit (HDBSCAN)	0.6733	0.6501	0.6524	0.4558
BERTopic Refit (KMeans)	0.3694	0.4131	0.4094	0.4111
Lifelong NTM	0.2811	0.3042	0.3446	0.1684

Appendix

Table 6: Final C v by UMAP dimensionality on TweetNER.

Dataset	Spatiotemporal
Model	None	16	128	512
COBWEBTM (ours)	0.7960	0.6948	0.6837	0.6284
BERTopic (MiniBatchKMeans)	0.6460	0.6068	0.6403	0.6322
BERTopic Refit (HDBSCAN)	0.6572	0.5103	0.5190	0.5105
BERTopic Refit (KMeans)	0.6136	0.5568	0.5471	0.5516
Lifelong NTM	0.2381	0.1211	0.2331	0.1191

Appendix

Table 7: Final C v by UMAP dimensionality on Spatiotemporal dataset.

Figure 6: An ablation study to compare the lifelong results of using different embedding models for CobwebTM .

Dataset	20 News Groups			AG News			Stack Exchange
Model	NPMI	PCC	SD	NPMI	PCC	SD	NPMI	PCC	SD
MiniLM-L6	0.205	0.130	0.935	0.104	0.023	0.924	0.127	0.070	0.952
MiniLM-L12	0.196	0.119	0.952	0.109	0.031	0.959	0.130	0.055	0.862
RoBERTa (ours)	0.206	0.141	0.958	0.108	0.027	0.942	0.131	0.073	0.959

Appendix

Table 8: An ablation study to compare the hierarchical results of using different embedding models for CobwebTM .

Panel of Figure 7 — Figure 7: An ablation study to compare the lifelong results of CobwebTM to other methods across all datasets with a reduced initial batch size of 500 documents.

Dataset	TweetNER (Holdout: person )
Model	Before Injection C v	After Injection C v
COBWEBTM (ours)	0.6425	1.0000
BERTopic (MiniBatchKMeans)	0.3812	0.0000
BERTopic Refit (HDBSCAN)	0.4845	1.0000
BERTopic Refit (KMeans)	0.3940	1.0000
Lifelong NTM	0.6120	0.0000

Appendix

Table 9: Holdout topic injection experiment on TweetNER dataset.

Figure 8: An ablation study to compare the amount of each Cobweb operation per-batch for all three datasets in CobwebTM .

Dataset	StackExchange (batch_size=10)
Model	Total Time (s)	Final Batch (s)
COBWEBTM (ours)	382.94	1.37
BERTopic (MiniBatchKMeans)	362.81	1.52
BERTopic Refit (HDBSCAN)	16525.93	87.79
BERTopic Refit (KMeans)	10749.74	50.93
Lifelong NTM	657.44	2.62

Appendix

Table 10: Runtime comparison on StackExchange dataset.

Dataset	20k StackExchange		50k StackExchange
Model	Start C v	Final C v	Start C v	Final C v
COBWEBTM (ours)	0.4932	0.6573	0.5010	0.6441
BERTopic (MiniBatchKMeans)	0.4810	0.4270	0.4958	0.4476
BERTopic Refit (HDBSCAN)
BERTopic Refit (KMeans)
Lifelong NTM	0.7133	0.5522	0.7752	0.0000

Appendix

Table 11: Coherence C v comparison on 20k and 50k StackExchange datasets.

Topic Modeling

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

See all 58 papers

TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes
Feb 22, 2026Roman EggerTopic Modeling
Smartvisions.at & Modul University Vienna
Improving Topic Modeling by Distilling Soft Labels from Language Models
Feb 20, 2026Raymond Li, Amirhossein Abaskohi, Chuyuan Li +2Topic Modeling
University of British Columbia, Vancouver, Canada.
Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders
Jul 31, 2025Carolina Zheng, Nicolas Beltran-Velez, Sweta Karlekar +5Topic Modeling
Columbia University · Independent · Google Research

Feb 22, 2026J/K move · Enter open · S save

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

Roman Egger

Smartvisions.at & Modul University Vienna

Topic Modeling

Open paper AskPDF arXiv

Topic modeling extracts latent themes from large text collections, but leading approaches like BERTopic face critical limitations: stochastic instability, loss of lexical precision ("Embedding Blur"), and reliance on a single data perspective. We present TriTopic, a framework that addresses these weaknesses through a tri-modal graph fusing semantic embeddings, TF-IDF, and metadata. Three core innovations drive its performance: hybrid graph construction via Mutual kNN and Shared Nearest Neighbors to eliminate noise and combat the curse of dimensionality; Consensus Leiden Clustering for reproducible, stable partitions; and Iterative Refinement that sharpens embeddings through dynamic centroid-pulling. TriTopic also replaces the "average document" concept with archetype-based topic representations defined by boundary cases rather than centers alone. In benchmarks across 20 Newsgroups, BBC News, AG News, and Arxiv, TriTopic achieves the highest NMI on every dataset (mean NMI 0.575 vs. 0.513 for BERTopic, 0.416 for NMF, 0.299 for LDA), guarantees 100% corpus coverage with 0% outliers, and is available as an open-source PyPI library.

Abstract

Topic Modeling

Explore similar work

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

Improving Topic Modeling by Distilling Soft Labels from Language Models

Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders