Topic-Method Composition (TMC)

Updated 8 July 2026

Topic-Method Composition (TMC) is a compositional design pattern in machine learning that isolates reusable primitives, such as latent topics or expert updates, to enhance inference and prediction.
It employs a two-stage process where components are first learned independently and then combined via specific composition rules to fulfill tasks like interactive writing and topic-conditioned generation.
Applications span email assistance, spectral topic inference, and tangent-space model composition, demonstrating TMC’s versatility in merging topic models, experts, and specialist updates across domains.

Searching arXiv for the supplied ids and nearby metadata to ground the article in the cited papers. Topic-Method Composition (TMC) denotes, in a cluster of machine-learning and language-modeling works, a compositional strategy in which reusable topical, semantic, or specialist components are learned separately and then assembled for inference, generation, or interactive assistance. The literature considered here suggests that TMC is not a single canonical algorithm but a recurring design pattern: LDA-derived topic mixtures can guide email drafting, spectral topic models can be completed with document-level mixture inference, neural topic models can parameterize recurrent LLMs, and semantic networks can supervise topic discovery through graph-based optimization (Gella et al., 2015, Lee et al., 2017, Wang et al., 2017, Liu, 2021). Related work extends the same compositional intuition to independently trained specialists and to tangent-space model updates around a shared pretrained anchor (Granmo, 2023, Liu et al., 2023).

1. Conceptual structure and representative formulations

Across the cited work, TMC-like systems share a two-stage structure. First, they define a set of latent or explicit primitives: LDA topics, topic-topic moments, topic-specific expert networks, semantic topic sets, or specialized classifiers. Second, they define a composition rule that maps those primitives back to a concrete prediction task. In some cases the composition target is a discourse plan for the next sentence; in others it is a document-topic vector, a next-word distribution, or a final classifier output (Gella et al., 2015, Lee et al., 2017, Wang et al., 2017, Liu, 2021, Granmo, 2023, Liu et al., 2023).

Work	Primitive	Composition mechanism
Email response assistance	LDA latent topics	Predict whole-reply or next-sentence topic distributions
Spectral topic inference	Learned $B$ and $A$	Recover document mixtures with SPI or PADD
TCNLM	Document topic probabilities	Weight topic-specific RNN experts
UM-S-TM	Topic sets on a U-map	Minimize semantic center-of-mass distance
TMComposites	Specialized Tsetlin Machines	Normalize and sum specialist class sums
Tangent Model Composition	Tangent updates $\delta_i$	Add, scale, or subtract updates around a pretrained base

This variety is methodologically important. In the narrow topic-modeling literature, composition usually means assembling learned topic structure into document- or sentence-level predictions. In the broader model-composition literature, it denotes combining independently trained specialists without retraining. A plausible implication is that “composition” is the stable idea, whereas the nature of the components varies substantially across subfields.

2. Interactive writing and discourse planning

A direct TMC instantiation appears in email-assistance work that predicts latent topics rather than generating a full reply. The system first applies LDA to customer-agent email exchanges, uses the inferred topic distributions as “silver-standard” labels, and then trains supervised predictors for two tasks: predicting the topic distribution of the agent’s entire reply from the customer email, and predicting the topic distribution of the next sentence from the customer email, the agent’s current sentence, and sentence-order information (Gella et al., 2015).

The LDA component is standard. For a document $d$ , the topic mixture is sampled as $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ , with latent assignments $z_{dn} \sim \mathrm{Multinomial}(\theta_d)$ and words $w_{dn} \sim \mathrm{Multinomial}(\phi_{z_{dn}})$ . What is specific to the compositional setting is the downstream use of the inferred posterior vector $\hat{\theta}_d$ : it becomes a target for a classifier that maps sparse text features to topic scores. The composition unit is therefore not the entire email but a sequence of latent communicative functions.

The reported evaluation emphasizes top- $k$ usefulness rather than exact single-label accuracy. In 80% of the cases, the correct topic is present among the top five recommended topics out of fifty possible ones (Gella et al., 2015). That result is central to the paper’s interactive interpretation: the model need not deterministically choose the next sentence, but it can surface a compact candidate set that supports topical coherence, coverage, and incremental drafting. A common misconception is to treat this line of work as text generation; the paper instead frames it as composition assistance.

3. Topic-conditioned inference and generation

A second TMC line concerns what happens after global topic structure has been learned. In spectral topic models, training on word co-occurrence moments yields a word-topic matrix $B$ and a topic-topic matrix $A$ 0, but not the document-topic matrix $A$ 1. The paper on Prior-aware Dual Decomposition formalizes this missing step as a document-specific topic-composition problem. The corpus co-occurrence matrix is factorized as $A$ 2, and the goal is to recover each document mixture $A$ 3 from observed word counts $A$ 4 while respecting the learned prior information encoded in $A$ 5 (Lee et al., 2017).

The baseline SPI uses a probabilistic inverse $A$ 6 and computes $A$ 7, where $A$ 8. PADD adds prior matching and solves

$A$ 9

subject to

$\delta_i$ 0

The dual-decomposition formulation introduces a symmetric matrix $\delta_i$ 1 and yields independent per-document subproblems for fixed $\delta_i$ 2, with ADMM and Douglas-Rachford splitting used for the inner optimization. Empirically, SPI is strongest when topic correlations are weak, TLI degrades as $\delta_i$ 3 grows, and PADD substantially outperforms TLI while achieving topic compositions comparable to Gibbs sampling on semi-real and fully real data (Lee et al., 2017).

A complementary formulation appears in the Topic Compositional Neural LLM. There, a neural topic model infers a document-dependent topic vector $\delta_i$ 4, and the LLM is implemented as a Mixture-of-Experts in which each topic corresponds to an RNN expert. In the naive form,

$\delta_i$ 5

and the topic probabilities also parameterize effective recurrent matrices through

$\delta_i$ 6

The factorized implementation shares parameters while preserving topic specialization. On APNEWS, IMDB, and BNC, the model outperforms a pure RNN and other topic-guided LLMs; on APNEWS with the large model, perplexity reaches 47.74, compared with 58.89 for basic-LSTM, 50.17 for LDA+LSTM, 48.21 for TDLM, 48.97 for Topic-RNN, and 50.63 for LCLM (Wang et al., 2017). In this formulation, TMC is not post hoc inference but end-to-end topic-conditioned sequence modeling.

4. Semantic-network supervision and the Semantic Center of Mass

A non-probabilistic variant is provided by the Understanding Map Supervised Topic Model. UM-S-TM replaces distributions over words with sets of concepts embedded in an Understanding Map, a semantic network in which a concept links to another concept if the latter appears in the former’s definition. A document is represented as a bag of concepts with term frequencies as masses, and its topic is defined through the Semantic Center of Mass (SCOM) objective

$\delta_i$ 7

where $\delta_i$ 8 is a graph distance on the U-map and $\delta_i$ 9 is the candidate set of topics (Liu, 2021).

This construction yields a different notion of composition. The topic side is the candidate topic set on the graph; the method side is optimization over weighted semantic distances. The document-to-topic distance is further normalized as

$d$ 0

and the document can be vectorized by distances to all topics in a domain. The paper also extends this to sequential documents by averaging prefix-based vectors, thereby incorporating order information absent from a pure bag-of-words representation (Liu, 2021).

Three methods are proposed for discovering the SCOM. The first uses curve fitting over local champions indexed by topic length. The second adds an AIC/BIC-like penalty $d$ 1 to the distance. The third injects a constant noise term $d$ 2 into all squared distances to reduce sparsity bias. On the reported synthetic experiments, 68% of the 34 test cases show unanimous agreement among all three methods, 97% show agreement by at least two methods, and only one case yields complete disagreement (Liu, 2021). The paper contrasts this framework with LDA and pLSA by emphasizing that UM-S-TM is supervised by a semantic network, represents topics as concept sets, and represents documents as vectors of distances to topics rather than mixtures of latent distributions.

5. Extensions to specialist and tangent-space composition

Although not topic models in the narrow sense, two later works generalize the compositional idea to models themselves. TMComposites constructs a composite classifier from multiple specialized Tsetlin Machines, each trained independently under a different booleanization strategy: Histogram of Gradients, Adaptive Gaussian Thresholding, and Color Thermometers. Each specialist produces class sums

$d$ 3

which are normalized by

$d$ 4

and then aggregated as

$d$ 5

The collaboration is explicitly described as plug-and-play because the members can be connected at any time and in any combination without fine-tuning. The best reported composite, $d$ 6 Thresholding + $d$ 7 Color Thermometers + $d$ 8 Color Thermometers + HoG, achieves 75.1 on CIFAR-10, 52.2 on CIFAR-100, and 93.0 on Fashion-MNIST, corresponding to gains of +12, +9, and +2 points over individual Tsetlin Machines (Granmo, 2023).

Tangent Model Composition uses the same initials, but it denotes a distinct formalism. Here the components are tangent vectors around a pretrained model $d$ 9, with tangent model space

$\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 0

If component models $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 1 are trained independently, then weighted ensembling in tangent space collapses to a single composed tangent model: $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 2 The practical significance is that multiple components can be added, scaled, or subtracted at inference time with $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 3 inference cost rather than the $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 4 cost of evaluating $\theta_d \sim \mathrm{Dirichlet}(\alpha)$ 5 nonlinear models. The paper reports a 4.2% average accuracy improvement over ensembling non-linearly fine-tuned models at a 2.5× to 10× reduction of inference cost, and near-uniform outperformance of recent continual fine-tuning methods across task-incremental, class-incremental, and data-incremental settings on 13 experiments over 3 benchmark datasets, without a replay buffer (Liu et al., 2023).

These two works broaden the scope of TMC-like reasoning. Instead of composing topics, they compose specialist competencies or local parameter updates. This suggests that the compositional principle can migrate from representation space to model space.

6. Limitations, misconceptions, and acronym overlap

The cited literature places clear constraints on where TMC-style approaches are effective. In spectral topic inference, simple inversion is adequate when topics are mostly independent, but prior-aware inference becomes important when topic correlations are strong and documents are short (Lee et al., 2017). In tangent-space composition, the method is explicitly local to a pretrained embedding and is not suitable for arbitrary or distant tasks; the paper presents DomainNet as a failure mode in which non-linear fine-tuning outperforms tangent fine-tuning because the tasks are too dissimilar (Liu et al., 2023). In TMComposites, the collaboration rule is intentionally simple, the weighting beyond normalization is not learned, and the demonstration is confined to image classification, with better normalization and light optimization left open (Granmo, 2023). In interactive email assistance, the model predicts likely topics and supports drafting, rather than replacing the author with unconstrained text generation (Gella et al., 2015).

A related misconception is terminological. The literature considered here does not provide a universally standardized definition of “Topic-Method Composition”; instead, it presents multiple concrete formulations that share a compositional architecture. The same three-letter acronym is also used in unrelated domains. “TMC” can denote Tangent Model Composition in continual fine-tuning (Liu et al., 2023), transverse momentum conservation in Bayesian extractions of small-system collectivity (Guo et al., 23 Mar 2026), and it also appears in the astronomical source name TMC-1, as in the benchmark dark cloud used for reference chemical-composition studies (Gratier et al., 2016). Those usages are separate from the topic-model and model-composition literature discussed above.

Taken together, the research record portrays TMC less as a fixed algorithm than as a family of compositional methods. Its recurring premise is that a system can be improved by isolating reusable structure—topics, priors, semantic neighborhoods, experts, or tangent updates—and by defining an explicit mechanism for recombining that structure at the level of documents, sentences, tokens, or model outputs.