Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Diagnosing and Improving Topic Models by Analyzing Posterior Variability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| rank topic words by mean/SD of posterior samples instead of mean probability | LDA topic top-10 words · Wiki | mean human rating: 3.09 -> 3.25 | Helped |
| rank topic words by mean/SD of posterior samples instead of mean probability | LDA topic top-10 words · News | mean human rating: 3.04 -> 3.12 | Helped |
| rank topic words by minimum posterior sample value instead of mean probability | LDA topic top-10 words · News | mean human rating: 3.04 -> 3.19 | Helped |
| rank topic words by minimum posterior sample value instead of mean probability | LDA topic top-10 words · Wiki | mean human rating: 3.09 -> 3.18 | Helped |
| adjusted word rankings vs mean ranking in pairwise human votes | LDA topic words · pairwise MTurk comparison | majority-vote ratio: 1.59x (1.88x at >=4 votes, 10.00x at 5 votes) | Helped |
| use cosine similarity instead of KL, Euclidean or Jaccard in topic stability (ablation) | topic stability metric · topic quality correlation | Spearman rho: 0.249 vs 0.016 (KL), 0.013 (Euclid), 0.108 (Jaccard) | Helped |
| topic stability vs NPMI as predictor of human topic quality | topic quality prediction · News | Spearman rho: 0.553 -> 0.248 | Hurt |
| topic stability vs coherence/NPMI as predictor of cross-model consistency | topic consistency prediction · News | Spearman rho: 0.456 -> 0.627 | Helped |
| Mean/SD vs Min word adjustments | word ranking · News and Wiki | mean human rating: News 3.12 vs 3.19; Wiki 3.25 vs 3.18 | Mixed |
Closest papers by meaning
- Provable Algorithms for Inference in Topic Models (ICML 2016 · similarity 0.89)
- Online Bayesian Moment Matching for Topic Modeling with Unknown Number of Topics (NeurIPS 2016 · similarity 0.87)
- Scalable inference of topic evolution via models for latent geometric structures (NeurIPS 2019 · similarity 0.87)
- Precision-Recall Balanced Topic Modelling (NeurIPS 2019 · similarity 0.86)
- Probabilistic Non-Negative Matrix Factorization and Its Robust Extensions for Topic Modeling (AAAI 2017 · similarity 0.86)
- Adaptive Elicitation of Latent Information Using Natural Language (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.