topicwizard -- a Modern, Model-agnostic Framework for Topic Model Visualization and Interpretation

TL;DR

Topicwizard addresses the interpretability challenge of topic models by moving beyond top-$K$ word lists to interactive, model-agnostic visualizations that ground topics in the corpus. It unifies representations of topics, words, and documents via interactive views (topic representations, word embeddings, document organization, and group analyses) and leverages UMAP projections with formulas such as $s_t = \sum_{d=1}^D \Theta_{dt} \cdot |d|$ and $G_{ij} = \sum_{k=1}^D \Theta_{kj} \cdot I(g_k = i)$ to quantify importance and group relations. The framework is designed for broad compatibility (scikit-learn API, Gensim, BERTopic) and offers production-ready deployment (web app, Docker, Figures API), enabling practitioners to interpret and trust topic models across diverse methodologies. Empirically, topicwizard has seen substantial adoption (over 45,000 PyPI downloads), reflecting its practical impact for researchers and data analysts seeking richer, interactive topic-model interpretation.

Abstract

Topic models are statistical tools that allow their users to gain qualitative and quantitative insights into the contents of textual corpora without the need for close reading. They can be applied in a wide range of settings from discourse analysis, through pretraining data curation, to text filtering. Topic models are typically parameter-rich, complex models, and interpreting these parameters can be challenging for their users. It is typical practice for users to interpret topics based on the top 10 highest ranking terms on a given topic. This list-of-words approach, however, gives users a limited and biased picture of the content of topics. Thoughtful user interface design and visualizations can help users gain a more complete and accurate understanding of topic models' output. While some visualization utilities do exist for topic models, these are typically limited to a certain type of topic model. We introduce topicwizard, a framework for model-agnostic topic model interpretation, that provides intuitive and interactive tools that help users examine the complex semantic relations between documents, words and topics learned by topic models.

Figures (9)

• Figure 1: A Simplified Taxonomy of Topic Models
• Figure 2: Common Components Computed by Topic Models
• Figure 3: An overview of visualizations and pages in the topicwizard framework All visualizations were produced using KeyNMF keynmf
• Figure 4: Screenshot of the Topics page in the topicwizard Web Application
• Figure 5: Screenshot of the Words page in the topicwizard Web Application