nucleation

LIT-1542: The nested Chinese restaurant process and Bayesian nonparametric inference of topic hierarchies

1 min read

Properties6
StatusActive — read 2026-10-26 (NOTE-1176) in full, from arXiv 0710.0845v3, the authors' revision; worth reading for the nested Chinese restaurant process as a prior on trees of unbounded depth and branching (THEORY-969), for hLDA's single-path model of a document and what that does to overlap, and for a careful simulated-recovery test (THEORY-1010). Read it knowing that §5.1's level conditional does not match §2.4's parameterization of the GEM and §6.2's hyperparameters are out of range as printed, that it has no identifiability result, that its comparison with LDA rests on harmonic-mean estimates on one corpus (THEORY-989, Proposed), and that the typeset JACM text was not compared.
Version2
Tagsprobabilistic-modeling, information-retrieval
Extended byLIT-002 — Nested Hierarchical Dirichlet Processes
Cited as “paper” byNOTE-1176 — Blei, Griffiths and Jordan — The nested Chinese restaurant process and hLDA
Cited as “source” byTHEORY-1010 — On ten corpora simulated from three-level hierarchical LDA itself (100 documents of 250 words, a vocabulary of 100, sparse topics with η = 0.005, γ = 1), the highest-probability state of the collapsed Gibbs sampler recovered the generating tree exactly in eight; in one of the other two a wrong tree had higher posterior probability than the true one, THEORY-969 — In hierarchical LDA every document draws all its words from the topics on one root-to-leaf path of a tree drawn from the nested Chinese restaurant process, so the prior is a nested sequence of partitions of the documents, exchangeable in them, and two documents share a topic only by sharing the whole path above it: no document can draw on two sibling subtrees, THEORY-989 — On the abstracts of the Journal of the ACM, the held-out likelihood of infinite-depth hierarchical LDA is higher than that of flat LDA at every fixed topic count in a range bracketing hLDA's own posterior topic count, about 140, though the authors expect flat LDA to win once the topic count is free

Graph View

Blei, D. M., Griffiths, T. L. & Jordan, M. I. (2010). The nested Chinese restaurant process and Bayesian nonparametric inference of topic hierarchies. Journal of the ACM 57(2): 1–30. doi:10.1145/1667053.1667056. arXiv:0710.0845.

Read in full on 2026-10-26 from arXiv 0710.0845v3 (NOTE-1176). The summary is checked against the reading.

Why this is here

Cited at T21, where the owner asked whether hLDA could infer higher-level perspectives from employees’ activity, and the reply took it as the starting point while warning that its topics are not perspectives. Until this reading the record held the nCRP only as the single-path limit that the nested HDP of LIT-002 (NOTE-023) relaxes, and the hierarchical Pitman–Yor topic models of LIT-013, which T21 did not use. It is one of the candidate methods of QUESTION-276.

Backlinks