Group-topic model
Model Families AIMech:GroupTopicModel DRAFT
Aliases: GT model
A probabilistic model family jointly assigning topics to relation-bearing events and topic-specific latent groups to entities, using event text and event-specific group-pair agreement probabilities.
Identity Notes & Concept Scope
Scoped to Wang, Mohanty and McCallum's NIPS2005 group-topic model. Each entity has a separate group assignment for each topic; each event chooses one topic. Agreement probabilities are indexed by event and group pair, as shown by the event plate, rather than being one global rate for a topic/group pair. This differs from author-topic or ART models that condition on observed participants without a pairwise agreement likelihood, and from a token-wise document topic mixture. The conference year follows official proceedings, not the author PDF filename. No normalized candidate identity matched all 442 pinned AIO IDs, labels and scoped synonyms or 438 preceding native IDs, names and aliases. Full-checkout searches included ignored and hidden files in both checkouts; eleven matching lines per checkout came from four unique prior discovery-ledger contents with pending leads and bibliography context. All matching excerpts were read. All 205 semantic label/synonym hits and 21 annotated AIO nodes with outgoing edges were byte-verified against prior readings. Twenty-nine contextual YAMLs comprise 28 unchanged prior readings and one fully read final round103 native target verified against its saved review inputs. No candidate record is present within this bounded pinned-snapshot and maintained-corpus search.
Claims & Grounded Evidence
3 Claims
Mechanism
#event-topic-group-generation
Each topic has a Dirichlet-distributed group mixture and word vector. Every entity draws a group separately for each topic. An event chooses one topic uniformly, generates its word attributes from that topic and generates observed pairwise agreement indicators using Beta-distributed event/group-pair probabilities. Entity group assignments are shared across events assigned to the same topic, whereas the agreement-rate variables are event-specific. The worked relation is symmetric; a precise implementation must declare its participating pairs, missing observations and self-pair policy.
Mechanism
#collapsed-and-tempered-inference
Integrating group mixtures, word vectors and event block rates yields group updates combining a leave-entity-out group factor with Beta-integrated likelihood ratios for affected incident relations in matching-topic events. Event-topic updates combine a leave-event-out Dirichlet-multinomial word factor with the whole event's integrated agreement likelihood under the candidate topic groups. The printed group formula multiplies a topic indicator inside the product over all events, which would zero weights when an unrelated event is present; the prose-consistent interpretation restricts the product to matching events. The experiments raise the text factor to powers five and five hundred for Senate and UN data, respectively, changing the target relative to the untempered model.
Limitation
#derived-agreement-scope
Conditionally independent pairwise agreement observations define a model of the derived relation network, not automatically a coherent generative law for underlying individual votes. For three entities in one block, three pairwise disagreements have integrated probability 1/4 under a uniform Beta block prior, although no binary vote assignment can realize that triangle. This does not invalidate a model scoped to observed agreement networks; it limits interpretation as a complete individual-vote model.
⚠️ Curated Limitations & Negative Results
- Fixed topic and group counts, symmetric binary agreement and event-specific Beta rates define the inspected construction. General asymmetric or continuous relations are suggested without a complete evaluated extension. The text's i<=j notation does not settle every practical self-pair or missing-participant policy.
- The printed collapsed formulas require consistent removal of old assignments and the correct scopes for entity, word and relation counts. Modality powers define a tempered posterior weighting, not exact inference under the original untempered generative process. No behavior of uninspected author software is inferred from a display error.
- The selected Senate and UN studies compare joint fitting with a sequential unigram-mixture/blockstructure baseline using fixed, somewhat arbitrarily chosen counts. Agreement Index results and selected political group interpretations do not demonstrate causal influence, validated ideological truth or general future prediction.
- Finite calculations check 4096 group/event conditionals for untempered and power-five synthetic examples with explicitly chosen distinct unordered pairs. They do not reproduce the original preprocessing, power-five-hundred benchmark or author implementation.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | This source-defined probabilistic model family is narrower than the pinned Model class. |
AIO:ProbabilisticTopicModel
Probabilistic Topic Model
|
related_match | The family 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
|
Xuerui Wang, Natasha Mohanty, Andrew McCallum
|
primary paper
full_text
|
sha256:f01d6c2c039f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:49:25Z
Updated: 2026-10-12T00:49:25Z
Source File:
data/curated/GroupTopicModel.yaml