TriTopic
| Original author(s) | Roman Egger |
|---|---|
| Developer(s) | SmartVisions |
| Initial release | 2026 |
| Stable release | 2.6.0
/ October 2026 |
| Written in | Python |
| Engine | |
| Operating system | Cross-platform |
| Type | Topic modeling |
| License | MIT License |
| Website | www |
Search TriTopic on Amazon.
TriTopic is an open-source method and Python library for topic modeling of text collections. It was introduced by Roman Egger in 2026.[1] The method combines semantic text representations based on sentence embeddings with lexical information from tf–idf weighting and, optionally, document metadata in a single document graph. Topics are then identified through repeated graph-based community detection and consensus clustering.[2]
.svg)
Background
Topic modeling aims to discover recurring themes in large collections of text. Classical approaches include latent semantic analysis, non-negative matrix factorization and latent Dirichlet allocation (LDA).[3] More recent methods such as BERTopic and Top2Vec cluster documents based on transformer-based text embeddings.[4]
TriTopic belongs to the embedding- and graph-based approaches. Unlike probabilistic models such as LDA, it does not rely on a generative model but represents similarities between documents as a graph. The original publication motivates the method with three limitations it attributes to existing embedding-based approaches: results that vary between runs, a loss of lexical precision, and reliance on a single view of the data.[1]
Method
.svg)
Document representations
Documents are first encoded with a sentence embedding model, and the dimensionality of the embeddings is reduced with UMAP. In addition, a lexical representation based on tf–idf captures the actual vocabulary used. Optionally, metadata such as source or date can be included; according to the documentation, metadata only strengthens links between documents that are already similar in content and does not create new ones.[2]
.svg)
Graph construction
The representations are fused into a weighted document graph in which documents are nodes and similarities are edges. Neighbourhoods are determined using mutual k-nearest neighbors and shared nearest neighbors, and links that appear in more than one view receive a higher weight.[1]
Consensus clustering
Communities in the graph are detected with the Leiden algorithm.[5] Because the result depends on random initialisation, TriTopic runs the clustering several times and builds a consensus graph in which document pairs that are frequently assigned to the same community receive stronger links. This is intended to make the results more reproducible.[1]
Number of topics and refinement
The number of topics can be specified or determined automatically. In automatic mode, several resolution parameters are tested and compared by the coherence of the resulting keywords. The document representations are then iteratively moved towards the centre of their topic and re-clustered until consecutive partitions agree.[2]
Topic representation
Topics are described by characteristic keywords and representative documents; keyword selection takes into account both how frequent a term is within a topic and how many of the topic's documents contain it.[2] The publication also describes an archetype-based topic representation that considers boundary cases in addition to central documents. Unlike density-based clustering methods such as HDBSCAN, every document is assigned to a topic.[1]
Further features
The library supports partially guided topic modeling, in which known categories are supplied as seed topics while further topics can still emerge from the data. It also provides functions for analysing how topics change over time, hierarchical topic structures, and analysis tools aimed at empirical research, such as topic stability under resampling, thematic saturation and group comparisons.[2]
Large language models are not required to form the topics. They can optionally be used after clustering to name and describe topics.[2]
Evaluation
In the original publication, TriTopic was compared with LDA, non-negative matrix factorization and BERTopic on the 20 Newsgroups, BBC News, AG News and arXiv datasets. The author reports a mean normalized mutual information (NMI) of 0.575 for TriTopic, compared with 0.513 for BERTopic, 0.416 for NMF and 0.299 for LDA.[1] No independent comparative evaluations have been published so far.
Applications and limitations
The documentation lists open-ended survey responses, interviews, customer reviews, news articles, social media posts and support tickets as typical applications, for collections ranging from a few hundred to tens of thousands of texts. It describes the method as less suitable for very short text snippets, for long documents covering many themes (which should be split into sections first) and for corpora of millions of documents.[2]
See also
References
- ↑ 1.0 1.1 1.2 1.3 1.4 1.5 Egger, Roman (2026-02-22). "TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes". arXiv:2602.19079 [cs.CL].
- ↑ 2.0 2.1 2.2 2.3 2.4 2.5 2.6 "TriTopic – Graph-Based Topic Modeling". SmartVisions. Retrieved 2026-10-18.
- ↑ Egger, Roman (2022). "Topic Modelling: Modelling Hidden Semantic Structures in Textual Data". In Egger, Roman. Applied Data Science in Tourism: Interdisciplinary Approaches, Methodologies, and Applications. Tourism on the Verge. Cham: Springer. pp. 375–403. doi:10.1007/978-3-030-88389-8_18. ISBN 978-3-030-88388-1. Search this book on
- ↑ Egger, Roman; Yu, Joanne (2022). "A Topic Modeling Comparison Between LDA, NMF, Top2Vec, and BERTopic to Demystify Twitter Posts". Frontiers in Sociology. 7. doi:10.3389/fsoc.2022.886498. Unknown parameter
|article-number=ignored (help) - ↑ Traag, Vincent A.; Waltman, Ludo; van Eck, Nees Jan (2019). "From Louvain to Leiden: guaranteeing well-connected communities". Scientific Reports. 9. doi:10.1038/s41598-019-41695-z. Unknown parameter
|article-number=ignored (help)
External links
This article "TriTopic" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:TriTopic. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
