High-granularity topic modelling extracts hundreds of niche topics from a corpus, yet this regime is rarely studied and is more commonly framed as a failure mode: "over-clustering," in which coherent themes fragment into redundant subtopics. We argue that this framing conflates genuine redundancy with topics that are semantically similar but meaningfully distinct, and that the central challenge in high-granularity modelling is separating the two. Recent post-hoc refinement methods address topic redundancy through either embedding-based similarity thresholding or LLM-guided merging. We show that neither suffices in high-granularity settings and propose embedding-gated iterative merge (EGIM), a hybrid approach that iteratively generates context-rich topic representations, uses embedding similarity to propose candidate merges across the topic space, then applies LLM judgement to accept only genuinely redundant pairs while preserving niche distinctions. Across two patent corpora and one legal corpus, LLM judgement rejects 83% of embedding-identified candidates as non-redundant. EGIM improves discriminability on merged topics by 25-55% over a prior iterative LLM baseline while better preserving the original topic distribution.
Figures & tables
K
UMass
Coh.
Intr.
Div.
20
−1.76
2.34
.66
.89
40
−1.97
2.36
.67
.85
80
−2.09
2.32
.64
.80
160
−2.28
2.26
.62
.76
320
−2.48
2.19
.59
.72
Table 1: Topic quality across granularity levels, averaged over all configurations in the Stage A sweep. Coh. = LLM coherence rating (1–3); Intr. = word-intrusion accuracy (chance = .17); Div. = topic diversity.
LASER
LEGAL
SPAD
Original
.533
.355
.369
EGIM
.537 (101%)
.349 (98%)
.346 (94%)
LLM-Iter
.469 (88%)
.275 (78%)
.127 (34%)
Table 2: Topic dispersion after refinement (higher is better). Percentages show relative change against the original model. EGIM wins 18/18 configurations.
Corpus
EGIM
LLM-Iter
Δ
LASER
.130
.103
+25%
LEGAL
.104
.079
+32%
SPAD
.096
.062
+55%
Table 3: Topic discriminability on merged topics (higher is better). EGIM wins 16/18 configurations.
LIDAR and LADAR Systems
+ Weapon-Locating Ladar
+ Raman/Fluorescence Explosive Detection
+ Laser Range Finding
+ Semi-Active Laser Guidance
+ Laser Wavelength Selection
Result: Laser Targeting & Communication
Table 4: Semantic snowballing under LLM-Iter (LASER corpus). One topic absorbs five unrelated others.
Figure 1: Embedding similarity (cosine) distributions for merged and declined pairs across 3,491 candidates. Dashed lines indicate class means. Heavy overlap confirms that no threshold reliably separates merge from decline decisions.
Topic A
Topic B
Sim
Declined (distinct concepts)
Contract Interp. Principles
Statutory Interp. Principles
.893
Misleading Conduct
Unconscionable Conduct
.854
Merged (true duplicates)
Tax Assessment Objections
Taxation Objection Appeals
.861
Extension of Time Applications
Extension of Time for Appeals
.875
Table 5: Similarity alone cannot determine merge decisions (LEGAL corpus). Distinct concepts may exceed the similarity of true duplicates.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Corpus
Context line
LASER
You are labeling topics from a corpus of laser technology patents filed by defense companies.
SPAD
You are labeling topics from a corpus of patents related to Single-Photon Avalanche Diode (SPAD) technology.
LEGAL
You are labeling topics from a corpus of Australian Federal Court legal cases.
Appendix
Table 6: Corpus-specific context lines injected into topic representation and merge/keep prompts.
Decision
Guidance (LASER corpus example)
MERGE
• Same core technology, even if applied differently (e.g., “WDM Networks” and “WDM for Data Transmission” → both WDM) • One topic is a specific technique/variant of the other (e.g., “Grating-Based Beam Combining” is a type of “Coherent Beam Combining”) • Same physical principle with different implementations • A domain expert would file these under one category for competitive analysis
KEEP
• Fundamentally different physical principles (e.g., “Thermal Cooling” vs “Optical Sensing”) • Completely different application domains (e.g., “Medical Imaging” vs “Weapons Targeting”) • Topics serve genuinely different analytical purposes • A domain expert would want these tracked separately
Appendix
Table 7: Domain-specific merge/keep guidance injected into the judgment prompt (LASER corpus shown). SPAD and LEGAL corpora use equivalent guidance with domain-appropriate terminology and examples.
Component
Parameter
Type
Range
UMAP
n_neighbors
int
[5,50]
n_components
cat
{5,10,15,20}
min_dist
float
[0.0,0.5]
HDBSCAN
min_cluster_size
int
[2,200]
min_samples
int
[2,60]
cluster_sel_epsilon
float
[0.0,0.5]
Appendix
Table 8: Hyperparameter search space for Stage A optimisation.
“…Emphasize identifying repetitive patterns, formulaic language, and structural boilerplate …”
6
Qwen3-Embedding-0.6B
“…Focus on novel substantive content, key concepts, and unique information …”
Appendix
Table 9: Embedding configurations for the three LLM-assisted cleaning variants.
Corpus
Metric
Original
EGIM
LLM-Iter
laser
cv
.710
.710
.710
NPMI
.154
.156
.155
UMass
−2.46
−2.46
−2.47
LLM Coherence
2.38
2.40
2.38
LLM Intrusion
.701
.716
.703
Diversity
.724
.731
.730
Appendix
Table 13: Full model quality across all metrics and methods. Corpus rows show means over 6 preprocessing variants; grand-mean rows average the three corpus means. Parenthesised percentages show change relative to the original model. All changes remain within ± 3%.
Figure 2: Merge and decline decisions across refinement rounds, aggregated over 18 configurations. Productive merging is concentrated in early rounds; later rounds are dominated by decline decisions.
Figure 3: Embedding similarity distributions for merged and declined pairs, separated by corpus. Dashed lines indicate class means. LASER shows the cleanest separation (AUC = 0.89); SPAD’s denser similarity structure produces heavy overlap (AUC = 0.75).
Topic A (label + summary)
Topic B (label + summary)
Sim.
Dec.
Beam Expansion and Collimation. ⋆ Use of multi-lens arrangements and beam expanders to manipulate laser beam divergence and collimation.
Laser Beam Expanders with Liquid Lenses. Beam expanders utilising liquid lens elements, often made with Cargille liquids, to achieve specific optical properties.
.860
merge
Optical Reflection and Refraction. ⋆ Use of mirrors and prisms for manipulating laser beams through reflection and refraction in optical systems.
Turning Mirror in Laser Cladding. Use of a turning mirror in laser cladding systems to redirect beams and protect components during welding operations.
.838
decline
Optical Beam Splitters. ⋆ Use of beam splitters to divide laser beams into multiple segments for various applications, including measurement and signal processing.
Optical Signal Distribution in Sensor Arrays. Methods and systems for distributing optical signals to sensor arrays using couplers, distribution buses, and return buses.
.799
decline
Narrow Bandpass Optical Filters. ⋆ Narrow bandpass filters designed to selectively transmit specific wavelengths while blocking others, often used in laser applications.
Metamaterial Filters in IR Systems. Use of metamaterial filters in active infrared systems for spectral filtering and handling various laser wavelengths and incident angles.
.789
decline
Optical Comb Filtering and Sampling. ⋆ Use of optical comb filters and samplers to process and integrate modulated spectral copies for frequency analysis and signal processing.
Optical and Femtosecond Frequency Combs. Generation, stabilisation, and application of optical frequency combs, consisting of multiple phase-locked frequency tones.
.787
decline
Diffraction Grating Systems. ⋆ Use of diffraction gratings to manipulate and control beam dispersion and spatial chirping in optical systems.
In-Plane and Bevel Angles in Waveguides. Manipulation of in-plane and bevel angles in optical waveguides to control photon trajectories and enhance light guidance.
.779
decline
Appendix
Table 14: EGIM decisions on the six chain topics from Figure 4 ( laser corpus). Each starred ( ⋆ ) topic is paired with its highest-similarity partner under EGIM; none of the six are paired with each other, and all are preserved as distinct.
Topic A (label + summary)
Topic B (label + summary)
Sim.
Dec.
Copyright Infringement Analysis. ⋆ Legal analysis of copyright infringement, focusing on the concepts of substantial similarity, originality, and the reproduction of works.
Copyright Infringement Dispute. ⋆ A legal dispute involving allegations of copyright infringement related to architectural plans created by Atrium and used by the Gaunt brothers and Beaumonde.
.805
merge
Damages for IP Infringement ⋆ : not selected for any merge by EGIM; preserved as a distinct topic.
Appendix
Table 15: EGIM decisions on the three chain topics from Figure 5 ( legal corpus). EGIM merges the two infringement-analysis topics but preserves the damages/remedies topic that LLM-Iter absorbed.
Topic A (label + summary)
Topic B (label + summary)
Sim.
Dec.
Photodiode Array Structures. ⋆ Design and arrangement of photodiode arrays, including avalanche photodiodes, within semiconductor substrates for light detection.
Photodiode Array Configurations. ⋆ Various configurations and applications of photodiode arrays, including avalanche photodiodes, for precise photon detection.
.907
merge
Photodetector–Substrate Integration ⋆ and HA and VA Photodetector Arrays ⋆ : not selected for any merge by EGIM; preserved as distinct topics.
Appendix
Table 16: EGIM decisions on the four chain topics from Figure 6 ( spad corpus). EGIM merges the two near-duplicate array topics but preserves the substrate-integration topic and the specific HA/VA geometry topic that LLM-Iter absorbed.
Recent advancements in topic modeling have split into two main paths: efficient embedding-based methods that offer scalability but are often outperformed on standard clustering metrics, and Large Language Model (LLM) approaches that achieve state-of-the-art clustering performance but at the cost of prohibitive computational expense and poor scalability. This paper introduces TopicMerge, a novel multi-stage pipeline that bridges these approaches to achieve both state-of-the-art clustering performance and computational efficiency. TopicMerge first employs an efficient clustering pipeline to deliberately over-segment a document corpus into a large set of fine-grained, semantically pure micro-clusters, which are then given initial semantic labels. It then introduces a novel, iterative merging algorithm where an LLM, guided by local neighborhood context, arbitrates on consolidating candidate pairs. These approved merges are then selectively executed and semantically relabeled by the LLM, hierarchically building broader, more coherent topics. Extensive experiments on the Bills and Wiki datasets show that TopicMerge achieves state-of-the-art clustering performance while reducing required LLM calls by over 92% compared to document-level methods, offering a practical and scalable solution.
Matthew Raffel, Lizhong Chen, Sai Munikoti +1
Oregon State University · Pacific Northwest National Laboratory
Traditional topic modeling assigns a single topic to each document. In practice, however, many real-world documents, such as product reviews or open-ended survey responses, contain multiple distinct topics. This mismatch often leads to topic contamination, where unrelated themes are merged into a single topic, making it difficult to identify documents that truly focus on a specific subject. We address this issue by introducing segment-based topic allocation (SBTA), a reformulation of topic modeling that assigns topics not to entire documents, but to segments: short, coherent spans of text that each express a single theme. By modeling topical structure at the segment level, our approach yields cleaner and more interpretable topics and better supports analysis of multi-theme documents. To support systematic evaluation, we construct a SemEval-STM, a new dataset inspired by aspect-based sentiment analysis. Documents are first decomposed into topical segments using large language models (LLMs), followed by human refinement to ensure segment quality. We also propose a segment-level extension of the word intrusion task, enabling human evaluation of topical coherence at the granularity where topics are actually assigned. Across multiple models and evaluation metrics, we show that SBTA improves clustering quality and interpretability. Overall, this work provides a practical, scalable framework for fine-grained topic analysis in heterogeneous text corpora where documents naturally span multiple topics. URL: https://huggingface.co/datasets/LG-AI-Research/SemEval-STM
Hoonsang Yoon, Takyoung Kim, Wonkee Lee +3
1LG AI Research · University of Illinois Urbana-Champaign
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.