Pachinko allocation model
Model Families AIMech:PachinkoAllocationModel DRAFT
Aliases: PAM
A topic-model family that generates each word by traversing a directed acyclic graph from a root through topic nodes to a word leaf, using probability distributions over each node's children.
Identity Notes & Concept Scope
Scoped to Li and McCallum's PAM construction, with Dirichlet child-distribution priors and the four-level experimental variant in ICML2006. Interior topics can have multiple parents, and paths may skip levels. A new path is selected for each token. This differs from a single topic path assigned to an entire document, a logistic-normal topic mixture, or a causal graph of semantic concepts. PAM is a source acronym within this scope; later hierarchical, nonparametric and scalable variants are not asserted as exact aliases. No normalized candidate identity matched all 442 pinned AIO IDs, labels and scoped synonyms or 432 preceding native IDs, names and aliases. Full-checkout searches included ignored and hidden files in both checkouts; six matching lines per checkout came from three unique file contents containing prior frontier leads, bibliography context and a cached publication-list mention. All candidate-matching excerpts were read; the long cached browser transcript was not reread in full. All 205 semantic label/synonym hits were byte-verified unchanged from prior reading. Eighteen annotated AIO CLASS nodes and outgoing edges comprise 17 verified prior inspections and one fresh reading. Twenty-seven contextual YAMLs comprise 25 unchanged prior readings, one fresh generated AIO reading and one final round102 native record verified against its fully read review inputs. No candidate record is present within this bounded pinned-snapshot and maintained-corpus search.
Claims & Grounded Evidence
2 Claims
Mechanism
#dag-path-generation
In the general Dirichlet formulation, each document draws a categorical child vector for each interior node. Each token starts at the root and repeatedly chooses a child until it reaches a word leaf. The finite rooted DAG and reachable word leaves ensure termination. In the demonstrated four-level variant, root-to-supertopic and supertopic-to-subtopic vectors vary per document, while subtopic-to-word vectors are drawn once and shared across the corpus. The latter sharing changes corpus marginalization: shared random emissions must be integrated jointly. The product of separately marginalized documents written for the fully document-local formulation is not automatically the marginal law of this shared-emission variant.
Mechanism
#collapsed-paths-and-moments
For fixed positive hyperparameters in the four-level variant, a token's collapsed supertopic/subtopic weight multiplies three leave-one-out predictive factors: the document root-to-supertopic count, the document supertopic-to-subtopic count, and the corpus subtopic-to-word count, each smoothed and normalized by its Dirichlet prior. The source jointly samples both topic levels. Its experimental fitting also changes supertopic hyperparameters using smoothed empirical moments and a concentration heuristic, rather than exact maximum likelihood or MAP. That changing-parameter procedure is not simply fixed-target posterior Gibbs sampling; neither finite samples nor moment updates establish convergence or optimality.
⚠️ Curated Limitations & Negative Results
- The experiments fix a fully connected four-level structure and topic counts. Learning arbitrary DAG structure or a nonparametric number of nodes is prospective in this paper. A DAG can encode nested and overlapping mixtures without establishing a universal representation theorem for every correlation structure.
- Dirichlet topic proportions are compositionally dependent, with negative off-diagonal covariances; the introductory LDA-independence contrast is not literal independence. Logistic-normal covariance parameters likewise do not mean higher-order dependence vanishes. PAM supplies a different mixing construction, not proof that CTM is exactly equivalent to a two-child DAG.
- Moment updates require policies for unvisited parent nodes, zero or invalid variances and nonpositive concentration estimates. Smoothing child means alone does not specify all edge cases. LDA reductions require compatible shared or fixed emissions; redrawing every word-emission vector independently per document is a different model.
- Evidence includes selected topics, a small human-judgment study, approximate held-out likelihood and five-way newsgroup classification. The prose and Table 2 disagree on majority-vote counts. The generated-document likelihood approximation is not exact marginal likelihood, and the reported results do not establish universal superiority. No author code or benchmark was executed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | This source-defined latent-variable model family is narrower than the pinned Model class. |
AIO:ProbabilisticTopicModel
Probabilistic Topic Model
|
related_match | The model supports probabilistic topic modeling; the pinned target is task-scoped, so related match preserves the model/task distinction. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Wei Li, Andrew McCallum
|
primary paper
full_text
|
ID: doi:10.1145/1143844.1143917
sha256:645882dd892a…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:24:39Z
Updated: 2026-10-12T00:24:39Z
Source File:
data/curated/PachinkoAllocationModel.yaml