Kar Wai Lim, Wray Buntine, Changyou Chen, Lan Du (2016), International Journal of Approximate Reasoning 78, 172-191 (2016) — ARXIV-1609.06783
Key takeaways
- Hierarchies of Pitman-Yor processes can be assembled as modular “blocks” and fit with a single collapsed, blocked Gibbs sampler (extending Chen et al. 2011), making nonparametric topic models with auxiliary data practical; the resulting Twitter-Network Topic Model beats HDP-LDA and a nonparametric author-topic model on perplexity.
Seeded from the abstract and a skim, before the reading in NOTE-008. What follows is what the work says about itself.
The Dirichlet process and its generalization the Pitman-Yor process take distributions as parameters and can be stacked into hierarchical nonparametric Bayesian models. The article gives efficient inference methods for such hierarchies and applies them to latent-variable text models, i.e. topic models. It proposes a general design framework and then a specific nonparametric model for social-media text, using tweets as the case study. The nonparametric model is reported to outperform parametric baselines in goodness of fit and in applications.
Standing in the record
Filed by the reading-time triage of 2026-09-25: 2,386 seconds of active reading over 3 sessions in the papers-feed tracker. It was filed Deferred, unread. NOTE-008 is the close reading of 2026-09-25, and it placed the work: Active — worth reading as a clear, self-contained tutorial on the table-count (Chen et al. 2011) collapsed blocked Gibbs sampler for hierarchical PYPs with discrete and mixture bases. The TNTM experiments are thin: one corpus, clustering baselines copied from another paper, and assertion-level claims about kernels and labels. Cite it for the sampler exposition, not for the empirical superiority of nonparametric models.
When seeded, before the reading in NOTE-008, the entry gave a priority for a deeper reading: medium — the owner invested ~40 min (t=2386 s) and it is a good tutorial-style reference for HPYP Gibbs sampling, but the skim captures the claims and the application is dated.
What the seeding asked a deeper reading to check (NOTE-008 is that reading):
- Main reusable contribution is the modular CRP-count representation and sampler; check the table-count sampling details (§4, Appendix A) if implementing.
- Empirical comparisons are against 2013-era baselines on a small, filtered corpus (150 authors); results say little about modern neural topic or embedding methods.
- ML link: PYP discounts give power-law word statistics, relevant background for language modelling priors; otherwise historically rather than currently relevant to ML practice.
Access when seeded: arXiv abs page and full v1 PDF (41 pp., journal preprint) read via arxiv.org; DOI and journal reference confirmed on arXiv and via Crossref (IJAR vol. 78, pp. 172-191, Nov 2016; Crossref title uses “Pitman–Yor” with lower-case “topic modelling”). Keywords verbatim from the paper. Appendix derivations not checked.