Traditional topic modeling, exemplified by Latent Dirichlet Allocation (LDA), operates on a purely unsupervised premise: if words frequently co-occur, they must belong to the same topic. However, this statistical assumption often fails to align with human intuition. Unsupervised models frequently produce "garbage topics" composed of high-frequency but semantically vacuous words, or they conflate distinct concepts that happen to share a vocabulary.
To bridge this gap, researchers have developed a suite of techniques to "steer" the discovery process. By incorporating human knowledge—in the form of seed words, category labels, or downstream tasks—these models ensure that latent themes are both mathematically coherent and practically useful.
Anchoring Topics with Seed Words
The most direct way to guide a model is through seed words. Instead of letting the model discover clusters from scratch, a user provides a few representative terms for the topics they care about.
In TopClus, Yu Meng and colleagues demonstrate how to move this process into the latent space of Pretrained Language Models (PLMs). By learning a spherical projection that clusters document representations around these seeds, the model generates topics that are far more targeted than "bag-of-words" approaches. Similarly, SeedTopicMine uses an ensemble ranking process to refine these seeds, ensuring that the discovered terms align with specific user interests rather than broad, overlapping descriptors.
The intuition here is anchoring. By fixing the position of certain "must-have" words in the latent space, the model is forced to organize the remaining vocabulary around these known landmarks. This is often implemented by modifying the underlying probability distributions. For instance, Yulan He and Deyu Zhou showed that modifying the Gibbs sampling process to enforce hard constraints on word-topic assignments allows models to match the performance of fully supervised systems using only a translated lexicon as a guide.
Joint Learning and the Predictive Objective
Another powerful mechanism for guidance is supervision via downstream tasks. If we know that certain documents belong to specific categories (e.g., "Positive" vs. "Negative" reviews), we can force the topic model to find themes that explain those labels.
This is the core idea behind Supervised LDA (sLDA) and its variants. Jun Zhu and Amr Ahmed introduced MedLDA, which integrates the "max-margin" principle of Support Vector Machines directly into the topic discovery process. Instead of just maximizing the likelihood of the text, MedLDA optimizes for predictive power. This "tight coupling" ensures that the discovered topics are discriminative; they don't just describe the data, they distinguish between classes.
In the neural era, this joint optimization is handled through multi-task loss functions. Babak Esmaeili proposed VALTA, a Variational Autoencoder (VAE) that balances text reconstruction with rating prediction. By tying latent topic discovery to predictive performance, the model is incentivized to ignore "noise" and focus on features that explain the variance in user ratings.
Constraints and Topic Sprinkling
When full supervision isn't available, researchers use weak supervision to nudge the model. One creative approach is Topic Sprinkling, introduced by Swapnil Hingmire. In TS-LDA, artificial tokens representing class labels are "sprinkled" into the training documents. During the generative process, these tokens act as fixed anchors. Because the model must account for these artificial words, it naturally aligns its latent topics with the desired classification structure.
For more complex data, such as product reviews where features like "battery life" might be discussed without being named, constrained topic modeling is used. Wei Wang and Hua Xu utilized indicator functions to enforce "must-link" and "cannot-link" relationships between words. This ensures that if the model sees "lasted all day," it is statistically pressured to associate it with the "battery" topic rather than a generic "quality" cluster.
Why Guidance Works
The success of these methods stems from a fundamental shift in how we view latent variables. In a purely unsupervised model, the latent space is a "black box" optimized for data compression. By adding seeds or labels, we transform that space into a structured representation that mirrors human ontology.
As Xian-Ling Mao demonstrated with SSHLDA, using labels to anchor the upper levels of a hierarchy prevents "semantic drift"—the tendency for unsupervised models to aggregate unrelated terms into broad, meaningless root nodes. Whether through Anchored CorEx's information-theoretic approach or GSSNMF's matrix factorization, the goal remains the same: to provide the model with enough "hints" to ensure that the patterns it finds are the patterns we actually care about.