Human-Centric Topic Modeling with Goal-Prompted Contrastive Learning and Optimal Transport

Rui Wang; Yi Zheng; Dongxin Wang; Haiping Huang; Yuanzhi Yao; Yuxiang Zhou; Jialin Yu; Philip Torr

cs.AIApr 14, 2026

Human-Centric Topic Modeling with Goal-Prompted Contrastive Learning and Optimal Transport

Authors: Rui Wang, Yi Zheng, Dongxin Wang, Haiping Huang, Yuanzhi Yao, Yuxiang Zhou, Jialin Yu, Philip Torr

Organizations: School of Computer Science, Nanjing University of Posts and Telecommunications · School of Computer Science and Information Engineering, Hefei University of Technology · School of Electronic Engineering and Computer Science, Queen Mary University of London · Department of Engineering Science, University of Oxford

Abstract

Existing topic modeling methods, from LDA to recent neural and LLM-based approaches, which focus mainly on statistical coherence, often produce redundant or off-target topics that miss the user's underlying intent. We introduce Human-centric Topic Modeling, Human-TM), a novel task formulation that integrates a human-provided goal directly into the topic modeling process to produce interpretable, diverse and goal-oriented topics. To tackle this challenge, we propose the Goal-prompted Contrastive Topic Model with Optimal Transport (GCTM-OT), which first uses LLM-based prompting to extract goal candidates from documents, then incorporates these into semantic-aware contrastive learning via optimal transport for topic discovery. Experimental results on three public subreddit datasets show that GCTM-OT outperforms state-of-the-art baselines in topic coherence and diversity while significantly improving alignment with human-provided goals, paving the way for more human-centric topic discovery systems.

Figures & tables

Figure 1. Key differences between Traditional Topic Modeling and Human-Centric Topic Modeling (with human’s goal).

Figure 2. The framework of GCTM-OT (a) and details of Goal-oriented Text Representation Mechanism (b).

Figure 3. Details of Goal Summarization Prompt

\mathcal{P}_{s}

Figure 4. Details of Text Augmentation Prompt

\mathcal{P}_{a}

Datasets	#Doc	#Vocab	Avg_Len
Bothering	1,483	7,700	56.48
TeslaModel3	1,991	5,634	32.82
AskAcademia	3,348	8,122	36.16

Table 1. The statistics of processed datasets.

Figure 5. The comparison of topic coherence values vs. different topic number settings [20, 30, 40, 50].

Dataset	Bothering				TeslaModel3				AskAcademia
Model	$\bm{C_{P}}$	$\bm{C_{A}}$	$\bm{NPMI}$	$\bm{UCI}$	$\bm{C_{P}}$	$\bm{C_{A}}$	$\bm{NPMI}$	$\bm{UCI}$	$\bm{C_{P}}$	$\bm{C_{A}}$	$\bm{NPMI}$	$\bm{UCI}$
LDA	0.1214	0.1286	0.0047	-0.3226	0.1143	0.1451	-0.0056	-0.6976	0.0997	0.1422	-0.0114	-0.6964
BAT	0.0127	0.1324	-0.0022	-0.8499	-0.0246	0.1441	-0.0058	-0.7193	-0.0786	0.1439	-0.0031	-0.6011
CTMNeg	-0.1319	0.1189	-0.0575	-1.9725	-0.1510	0.1195	-0.0611	-2.0211	-0.0156	0.1319	-0.0267	-1.2121
vONT	0.1535	0.1353	0.0020	-0.5920	0.0995	0.1363	-0.0093	-0.8079	0.0883	0.1187	-0.0018	-0.3956
CWTM	0.0175	0.1132	-0.0107	-0.6486	0.0417	0.1320	-0.0265	-1.0433	0.1039	0.1414	-0.0042	-0.6252

Table 2. Average topic coherence values with four topic number settings [20, 30, 40, 50].

Dataset	Human vs. Human			Human vs. LLM
Dataset	$\bm{JS}$	$\bm{SMC}$	$\bm{DSC}$	$\bm{JS}$	$\bm{SMC}$	$\bm{DSC}$
Bothering	0.7579	0.7700	0.8623	0.7738	0.7750	0.8722
TeslaModel3	0.9192	0.9200	0.9579	0.8800	0.8800	0.9357
AskAcademia	0.9600	0.9600	0.9796	0.9196	0.9200	0.9581

Table 5. Coefficients between two experts’ annotation results.

Model	Topic Summary	Topic Words
LDA	—	completely, walk , girlfriend, hang , answer, finally, happy, honestly text couple
	—	pain, issue, tired, doctor, knee, sister , sleep, clean, house, tear
	—	money, college, afford, bill, rent, apply, house, plan, student, boss
LLM-ITL	—	loved, relationship, past, old, text, feel, love, emotional, life, talk
	—	doctor, hospital, mental, cancer, symptom, helping, degree, currently, training, surgery
	—	money, work, need, wanted, open, using, month, job, know, asking

Table 6. Topic Examples extracted from the ‘Bothering’ dataset, italic means out-of-topic words.

Parameter	Setting	$\bm{C_{P}}$	$\bm{C_{A}}$	$\bm{NPMI}$	$\bm{UCI}$	$\bm{UT}$
CAST		0.2575	0.1592	0.0258	-0.3314	0.6300
$\delta$	0.2	0.3161	0.2001	0.0408	-0.1491	0.9700
	0.4	0.3431	0.1891	0.0414	-0.2578	0.9650
	0.6	0.3518	0.2084	0.0524	0.0568	0.9800
	0.8	0.3371	0.1962	0.0416	-0.1541	0.9850
	1	0.2959	0.1957	0.0329	-0.2770	0.9450

Table 7. Comparison results of hyperparameter analysis.

Model		$\bm{C_{P}}$	$\bm{C_{A}}$	$\bm{NPMI}$	$\bm{UCI}$	$\bm{UT}$
CAST		0.2575	0.1592	0.0258	-0.3314	0.6300
$\mathcal{T}$	GCTM-OT (RoBERTa)	0.3460	0.1944	0.0489	0.0414	0.9900
	GCTM-OT(BERT)	0.2760	0.1673	0.0381	0.0639	0.9450
	GCTM-OT(Sentence-BERT)	0.2760	0.1668	0.0371	-0.0707	0.9500
	GCTM-OT(MiniLM)	0.3359	0.1939	0.0506	0.0312	0.9600
	GCTM-OT	0.3518	0.2084	0.0524	0.0568	0.9800

Table 8. Comparison results of ablation studies.

Topic Modeling

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

See all 57 papers

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
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.
A Manifold-Aware Topic Modeling Approach via Rank-Based Prototypes
Aug 29, 2026Thiago César Castilho Almeida, Daniel Carlos Guimarães PedronetteTopic Modeling
Department of Statistics, Applied Mathematics and Computing State University of São Paulo (UNESP), Rio Claro, Brazil

Jul 31, 2025J/K move · Enter open · S save

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

Carolina Zheng, Nicolas Beltran-Velez, Sweta Karlekar, Claudia Shi+4

Columbia University · Independent · Google Research

Topic Modeling

Open paper AskPDF arXiv

Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.

Abstract

Topic Modeling

Explore similar work

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

Improving Topic Modeling by Distilling Soft Labels from Language Models

A Manifold-Aware Topic Modeling Approach via Rank-Based Prototypes