Composite HMM-topic model
Model Families AIMech:CompositeHMMTopicModel DRAFT
A generative sequence model combining an HMM class sequence with a document topic mixture, where one designated class emits words through latent topics and each other class emits words from its own categorical distribution.
Identity Notes & Concept Scope
Scoped to the composite model of Griffiths, Steyvers, Blei and Tenenbaum's Integrating Topics and Syntax. The maintained name describes that source-defined construction; no additional acronym is asserted as a source synonym. All tokens follow class transitions, while only the semantic class uses the document-topic mixture. This differs from independent topic models, generic HMMs, a product of HMM experts, or modern mixture-of-experts language models. No normalized candidate identity matched all 442 pinned AIO IDs, labels and scoped synonyms or 426 preceding native IDs, names and aliases. Both full-checkout searches included ignored and hidden files; 13 matching lines in each checkout came from four unique file contents containing prior frontier leads, contextual distinctions and a cached bibliography mention. Matching concept excerpts were inspected; the long cached web transcript was not reread in full. Semantic search returned 204 label/synonym hits: 200 verified unchanged from prior reading and four newly read. Eighteen annotated AIO CLASS nodes and outgoing edges comprise 16 verified unchanged and two newly read; 30 contextual YAMLs comprise 25 unchanged prior readings, two newly read AIO records and three final records verified against the round101 scientific-review inputs. No candidate identity is present within this bounded pinned-snapshot and maintained-corpus search.
Claims & Grounded Evidence
2 Claims
Mechanism
#class-dependent-emission
A document draws a Dirichlet topic vector. An HMM generates the token-class sequence. If a token occupies the designated semantic class, a topic selected from the document vector determines its word distribution; otherwise its class-specific word distribution generates the token. The Bayesian formulation puts symmetric Dirichlet priors on topic emissions, nonsemantic class emissions and transition rows. The source draws a topic at every position, but topics unused by nonsemantic emissions can be marginalized. The displayed model is first order; experiments also use higher-order transitions and a dedicated sentence-marker state. Initial and boundary conventions, class count, topic count and vocabulary are part of a complete specification.
Mechanism
#collapsed-auxiliary-inference
The source alternates topic and class updates with parameters integrated out. Its document-topic counts include semantic tokens only, and the first-order class update accounts for both neighboring transitions, including corrections when they share a transition row or cell. A coherent interpretation first marginalizes unused topics, then introduces a temporary topic for the currently updated nonsemantic token from the leave-one-out document predictive law. Updating that topic and then the class allows the common document-topic factor to cancel in the class conditional; discard the temporary topic if the resulting class is nonsemantic. This is an explicit curator interpretation of the printed count convention, checked in a small enumerated model, not a claim about unseen author code. Retaining all inactive topic draws in a fully augmented sampler would instead require document counts over all those draws.
⚠️ Curated Limitations & Negative Results
- The distinction between syntactic classes and semantic topics is induced by this model and corpus. It need not coincide with linguistic ground truth, supervised part-of-speech tags or full grammar. Fixed class/topic counts and finite-state local transitions restrict its scope.
- The source LDA comparison uses a semantic class plus a dedicated sentence-marker class with the stated boundary handling; arbitrary two-state HMMs are not equivalent to LDA. With one topic, the semantic emission has its own prior, so the resulting HMM comparison must preserve those hyperparameters.
- A finite auxiliary-update invariance calculation supports a compatible semantic-only interpretation; it is not a proof that historical software followed that convention or that finite chains mixed. The displayed first-order transition formula requires additional factors for higher-order contexts.
- Historical results use particular Brown, TASA and NIPS preparations. Adjusted Rand agreement is not supervised tag accuracy, and the reported harmonic-mean evidence estimate is not exact marginal likelihood or guaranteed stable. The document-classification experiment reports LDA somewhat better. No universal advantage or empirical reproduction is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | The maintained identity is a specified probability-model family, narrower than the pinned Model class. |
AIO:ProbabilisticTopicModel
Probabilistic Topic Model
|
related_match | The model supports probabilistic topic modeling; the pinned target is classified as a machine-learning task, so related match preserves the model/task distinction. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Thomas L. Griffiths, Mark Steyvers, David M. Blei, Joshua B. Tenenbaum
|
primary paper
full_text
|
sha256:8878657c94e1…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:00:04Z
Updated: 2026-10-12T00:00:04Z
Source File:
data/curated/CompositeHMMTopicModel.yaml