Probabilistic Latent Semantic Analysis

Thomas Hofmann

article1999UAI2,854 citations

Proposes Probabilistic Latent Semantic Analysis, replacing standard algebraic singular value decomposition with a statistical latent class mixture model fitted by tempered EM to substantially improve document modeling and information retrieval.

Listen

The article addresses challenges in analyzing text data where words can have multiple meanings (polysemy) or different words can refer to similar concepts (synonymy), limiting the effectiveness of keyword-based methods in information retrieval and related tasks. Standard Latent Semantic Analysis (LSA) attempts to address this by reducing high-dimensional word counts to a lower-dimensional space via singular value decomposition, but it lacks a strong statistical basis and can produce non-probabilistic or negative values that hinder interpretation.

The article sets out to develop and evaluate Probabilistic Latent Semantic Analysis (PLSA), a latent class model that treats co-occurrence data as arising from a mixture of unobserved topics, providing a proper generative probability distribution over documents and words.

The approach relies on the aspect model, a mixture decomposition fitted via the Expectation-Maximization algorithm, with a tempered variant (TEM) that incorporates annealing-like regularization to improve generalization on unseen data. Experiments used standard test collections such as MED (1033 documents), CRAN, CACM, and CISI, along with noun-adjective pairs from the LOB corpus, comparing perplexity and retrieval precision against LSA baselines.

PLSA reduced perplexity by factors of roughly 3.3 on MED data and 2.41 on LOB data, substantially outperforming LSA reductions of 1.9 and 2.08 respectively. In information retrieval tasks, PLSA delivered average precision gains of 44 percent on MED, 17 percent on CRAN, 28 percent on CACM, and 48 percent on CISI relative to a cosine term-frequency baseline, while LSA showed smaller or negative gains; combining multiple PLSA models further improved results by up to 58 percent.

These outcomes indicate that the probabilistic formulation yields more accurate topic decompositions, better handles word ambiguity through context-specific factor assignments, and supports reliable model selection, leading to measurable gains in retrieval quality and predictive modeling without relying on ad-hoc heuristics.

The findings support deploying PLSA for automated indexing and text analysis applications, with model combination offering a robust practical option; further work could explore online EM variants for larger collections.

Main limitations include the risk of local maxima in EM fitting and the need for held-out data to tune the tempering parameter, though results across multiple collections provide reasonable confidence in the reported advantages.

International Computer Science Institute University of California Berkeley latent factor latent variable machine learning Probabilistic Latent Semantic Analysis Aspect Model Expectation Maximization

arXiv: 1301.6705

Paper: Indexing By Latent Semantic Analysis, Scott Deerwester et al. (1990). This seminal paper introduces Latent Semantic Analysis via singular value decomposition, establishing the geometric latent-space paradigm that PLSA probabilistically reformulates.
Paper: Term-weighting approaches in automatic text retrieval, Gerard Salton et al. (1988). This classic work details term-weighting schemes and vector space retrieval models that serve as primary baselines and foundational representations in PLSA.
Paper: A comparison of event models for naive bayes text classification, Andrew McCallum et al. (1998). This paper establishes the multinomial generative model for text documents, providing the core statistical formulation that latent aspect models generalize.
Paper: Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval, S. E. Robertson et al. (1994). This work formulates probabilistic term-frequency weighting in text retrieval, setting the stage for statistical alternatives to geometric indexing.

Paper: Latent Dirichlet Allocation, David M. Blei et al. (2003). Latent Dirichlet Allocation introduces a fully generative Bayesian prior over topic distributions to overcome the document-level parameter growth and overfitting limitations of PLSA.
Paper: Unsupervised Learning by Probabilistic Latent Semantic Analysis, Thomas Hofmann (2001). This expanded journal treatise provides the complete theoretical foundation, tempered EM algorithms, and comprehensive experimental analysis for PLSA.
Paper: Reading Tea Leaves: How Humans Interpret Topic Models, Jonathan D. Chang et al. (2009). This study develops intrusion tasks to quantify how humans interpret the latent thematic spaces generated by topic models such as pLSI and LDA.
Paper: Exploring the Space of Topic Coherence Measures, Michael Röder et al. (2015). This paper establishes automated topic coherence metrics to evaluate the semantic quality and interpretability of topic modeling decompositions.
Paper: Probabilistic Matrix Factorization, Andriy Mnih et al. (2007). This work extends probabilistic latent-factor decomposition to dyadic collaborative filtering matrices under Gaussian likelihood assumptions.

Cover for Probabilistic Latent Semantic Analysis

Abstract

Probabilistic Latent Semantic Analysis is a novel statistical technique for the analysis of two-mode and co-occurrence data, which has applications in information retrieval and filtering, natural language processing, machine learning from text, and in related areas. Compared to standard Latent Semantic Analysis which stems from linear algebra and performs a Singular Value Decomposition of co-occurrence tables, the proposed method is based on a mixture decomposition derived from a latent class model. This results in a more principled approach which has a solid foundation in statistics. In order to avoid overfitting, we propose a widely applicable generalization of maximum likelihood model fitting by tempered EM. Our approach yields substantial and consistent improvements over Latent Semantic Analysis in a number of experiments.

Topics

latent factor latent variable machine learning maximum likelihood semantic information natural language understanding

Keywords

Probabilistic Latent Semantic Analysis Aspect Model Expectation Maximization Tempered Expectation Maximization Latent Semantic Analysis Singular Value Decomposition Mixture Model Perplexity Topic Decomposition Information Retrieval

1 Introduction
2 Latent Semantic Analysis
2.1 Count Data and Co-occurrence Tables
2.2 Latent Semantic Analysis by SVD
3 Probabilistic LSA
3.1 The Aspect Model
3.2 Model Fitting with the EM Algorithm
3.3 Probabilistic Latent Semantic Space
3.4 Topic Decomposition and Polysemy
3.5 Aspects versus Clusters
3.6 Model Fitting Revisited: Improving Generalization by Tempered EM
4 Experimental Results
4.1 Perplexity Evaluation
4.2 Information Retrieval
5 Conclusion
Acknowledgments
References

Knowls

Knowl 1 — The Aspect Model for Probabilistic Latent Semantic Analysis
model/method
The aspect model is a latent variable statistical model defined over dyadic co-occurrence data, such as a collection of text documents $\mathcal{D} = \{d_1, \dots, d_N\}$ and a vocabulary of terms $\mathcal{W} = \{w_1, \dots, w_M\}$ . Each observed occurrence of a word $w \in \mathcal{W}$ in a document $d \in \mathcal{D}$ is associated with an unobserved latent class variable $z \in \mathcal{Z} = {z_1, \dots, z_K}.

The model introduces a conditional independence assumption: document $d$ and word $w$ are conditionally independent given the latent variable $z$ . The joint probability distribution over $\mathcal{D} \times \mathcal{W}$ can be parameterized in an asymmetric form: $P(d, w) = P(d) P(w|d), \quad \text{where} \quad P(w|d) = \sum_{z \in \mathcal{Z}} P(w|z) P(z|d)$ Equivalently, the joint distribution can be parameterized in a symmetric form: $P(d, w) = \sum_{z \in \mathcal{Z}} P(z) P(d|z) P(w|z)$ Here, $P(d)$ is the prior probability of selecting document $d$ , $P(z)$ is the prior probability of latent class $z$ , $P(d|z)$ is the document distribution conditional on latent class $z$ , $P(w|z)$ is the class-conditional word distribution over $\mathcal{W}$ , and $P(z|d)$ represents the document-specific mixture proportions over the latent classes. Because the latent state space cardinality $K = |\mathcal{Z}|$ is chosen such that $K \ll \min(N, M)$ , the latent variable $z$ acts as an informational bottleneck.
Knowl 2 — Expectation-Maximization Algorithm for Fitting PLSA
algorithm
Parameters of the symmetric aspect model are estimated by maximum likelihood using the Expectation-Maximization (EM) algorithm. Given a co-occurrence matrix $N = (n(d, w))$ of observed counts of term $w \in \mathcal{W}$ in document $d \in \mathcal{D}$ , the EM algorithm alternates between computing posterior expectations of the latent variables and updating parameter distributions until convergence to a local maximum of the multinomial log-likelihood.
Input: Co-occurrence counts $n(d, w)$ for documents $d \in \mathcal{D}$ and terms $w \in \mathcal{W}$ , number of latent classes $K$
Output: Latent priors $P(z)$ , conditional document probabilities $P(d|z)$ , conditional term probabilities $P(w|z)$
Initialize non-negative parameters $P(z), P(d|z), P(w|z)$ subject to $\sum_z P(z) = 1$ , $\sum_d P(d|z) = 1$ , and $\sum_w P(w|z) = 1$
repeat
    for each document $d \in \mathcal{D}$ and term $w \in \mathcal{W}$ with $n(d, w) > 0$ do
        for each latent state $z \in \mathcal{Z}$ do
             $P(z|d, w) \leftarrow \frac{P(z)P(d|z)P(w|z)}{\sum_{z' \in \mathcal{Z}} P(z')P(d|z')P(w|z')}$
    for each latent state $z \in \mathcal{Z}$ do
        for each term $w \in \mathcal{W}$ do
             $P(w|z) \leftarrow \frac{\sum_{d \in \mathcal{D}} n(d, w) P(z|d, w)}{\sum_{w' \in \mathcal{W}} \sum_{d \in \mathcal{D}} n(d, w') P(z|d, w')}$
        for each document $d \in \mathcal{D}$ do
             $P(d|z) \leftarrow \frac{\sum_{w \in \mathcal{W}} n(d, w) P(z|d, w)}{\sum_{d' \in \mathcal{D}} \sum_{w \in \mathcal{W}} n(d', w) P(z|d', w)}$
         $P(z) \leftarrow \frac{\sum_{d \in \mathcal{D}} \sum_{w \in \mathcal{W}} n(d, w) P(z|d, w)}{\sum_{z' \in \mathcal{Z}} \sum_{d \in \mathcal{D}} \sum_{w \in \mathcal{W}} n(d, w) P(z'|d, w)}$
until log-likelihood converges
Knowl 3 — Matrix Formulation and Geometric Interpretation of PLSA vs. LSA
theoretical result
In matrix notation, the symmetric aspect model defining the joint probability matrix $\mathbf{P} \in \mathbb{R}^{N \times M}$ with entries $\mathbf{P}_{ij} = P(d_i, w_j)$ is expressed as: $\mathbf{P} = \hat{U} \hat{\Sigma} \hat{V}^T$ where $\hat{U} \in \mathbb{R}^{N \times K}$ has entries $\hat{U}_{ik} = P(d_i|z_k)$ , $\hat{V} \in \mathbb{R}^{M \times K}$ has entries $\hat{V}_{jk} = P(w_j|z_k)$ , and $\hat{\Sigma} = \text{diag}(P(z_1), \dots, P(z_K)) \in \mathbb{R}^{K \times K}$ .

This decomposition parallels Singular Value Decomposition (SVD) in Latent Semantic Analysis (LSA), $\mathbf{N} \approx U \Sigma V^T$ , with fundamental differences:
1. Distributional properties: $\hat{U}$ and $\hat{V}$ contain non-negative normalized conditional probabilities, and $\hat{\Sigma}$ contains prior mixture proportions replacing singular values. The matrix $\mathbf{P}$ defines a properly normalized probability distribution, whereas LSA approximations $\hat{N}$ can contain negative values and lack probabilistic normalization.
2. Optimization objective: LSA minimizes the $L_2$ (Frobenius) norm, corresponding to an implicit additive Gaussian noise assumption on counts. PLSA maximizes the multinomial likelihood function, which corresponds to minimizing the Kullback-Leibler divergence (cross-entropy) between the empirical distribution and the model distribution.
3. Geometric interpretation: Each factor $P(\cdot|z)$ represents a point on the $(M-1)$ -dimensional probability simplex over the vocabulary $\mathcal{W}$ . The convex hull of the $K$ factor vertices spans a sub-simplex of dimension at most $K - 1$ . Document distributions $P(w|d) = \sum_{z} P(w|z) P(z|d)$ are convex combinations of these factors, defining a continuous probabilistic latent semantic subspace.
Knowl 4 — Free Energy Objective and Tempered Posterior in Tempered EM
model/method
To prevent overfitting on co-occurrence counts and ensure generalization on unseen test data, maximum likelihood fitting for the aspect model is generalized via Tempered Expectation Maximization (TEM), which introduces an entropic regularization term.

The objective function is the parameterized Helmholtz free energy $\mathcal{F}_\beta$ : $\mathcal{F}_\beta = -\beta \sum_{d \in \mathcal{D}} \sum_{w \in \mathcal{W}} n(d, w) \sum_{z \in \mathcal{Z}} \tilde{P}(z; d, w) \log [P(d, w|z)P(z)] + \sum_{d \in \mathcal{D}} \sum_{w \in \mathcal{W}} n(d, w) \sum_{z \in \mathcal{Z}} \tilde{P}(z; d, w) \log \tilde{P}(z; d, w)$ where $n(d, w)$ denotes observed counts, $\tilde{P}(z; d, w)$ are variational posterior parameters defining a conditional distribution over latent classes $\mathcal{Z}$ , and $\beta \in (0, 1]$ is the inverse computational temperature.

Minimizing $\mathcal{F}_\beta$ with respect to the variational distribution $\tilde{P}$ yields the tempered posterior: $\tilde{P}(z; d, w) = \frac{[P(z)P(d|z)P(w|z)]^\beta}{\sum_{z' \in \mathcal{Z}} [P(z')P(d|z')P(w|z')]^\beta}$ At $\beta = 1$ , this coincides with the standard EM posterior. For $\beta < 1$ , the entropy penalty dampens posterior probabilities, shifting them closer to a uniform distribution and smoothing model predictions.
Knowl 5 — Inverse Annealing Algorithm for Model Fitting via Tempered EM
algorithm
Tempered Expectation Maximization (TEM) optimizes the aspect model using an inverse annealing schedule to control model complexity. Rather than starting hot and gradually cooling, it performs EM iterations at $\beta = 1$ with early stopping and subsequently decreases $\beta$ (heating the system) until performance on held-out validation data ceases to improve.
Input: Term-document count data $n(d, w)$ , held-out validation dataset, latent dimension $K$ , cooling factor $\eta \in (0, 1)$
Output: Fitted regularized parameters $P(z), P(d|z), P(w|z)$
Set inverse temperature $\beta \leftarrow 1.0$
Perform standard EM iterations with early stopping monitored on the held-out validation dataset
while true do
     $\beta \leftarrow \eta \cdot \beta$
    Perform one TEM iteration at inverse temperature $\beta$
    while performance on held-out validation dataset improves non-negligibly do
        Perform TEM iteration at the current inverse temperature $\beta$
    if decreasing $\beta$ fails to yield further validation improvements then
        break
return fitted model parameters
Knowl 6 — Probabilistic Latent Semantic Indexing and Model Combination
model/method
Probabilistic Latent Semantic Indexing (PLSI) applies the aspect model to information retrieval:
1. Latent Factor Representation: Documents $d$ are represented by their latent mixture weights $P(z|d)$ . Unseen queries $q$ are folded into the latent space by keeping trained term-factor distributions $P(w|z)$ fixed and calculating $P(z|q)$ using Tempered EM on query term counts.
2. Combined Similarity Function: Latent space cosine similarity $s_{\text{latent}}(d, q)$ between $P(z|d)$ and $P(z|q)$ is combined linearly with the baseline cosine term matching score $s_{\text{term}}(d, q)$ : $s_{\text{comb}}(d, q) = \lambda s_{\text{term}}(d, q) + (1 - \lambda) s_{\text{latent}}(d, q)$ where $\lambda \in [0, 1]$ is a combination weight.
3. Model Combination ( $\text{PLSI}^*$ ): Because statistical aspect models trained with different numbers of latent classes $K$ form non-nested decompositions that capture complementary topical structures, multiple models trained with varying dimensions $K$ (e.g., $K \in \{32, 48, 64, 80, 128\}$ ) are combined by uniformly averaging their similarity scores. This model averaging improves retrieval precision and eliminates the sensitivity associated with selecting a single latent dimension.
Knowl 7 — Distinction Between the Aspect Model and Distributional Clustering
theoretical result
The aspect model in PLSA is distinct from distributional document clustering models (unsupervised Naive Bayes models).

In document clustering, each document $d$ is assigned a single latent cluster variable $c(d) \in \mathcal{Z}$ , with conditional word probabilities: $P(w|d) = \sum_{z \in \mathcal{Z}} P(c(d) = z) P(w|z)$ By Bayes' rule, as document length increases, the cluster posterior $P(c(d) = z)$ concentrates its mass on a single cluster $z$ . Consequently, all words in the document must be drawn from that single cluster's distribution $P(w|z)$ , requiring each cluster to cover the full vocabulary of a subset of documents.

In contrast, the aspect model specifies: $P(w|d) = \sum_{z \in \mathcal{Z}} P(z|d) P(w|z)$ where $P(z|d)$ is a document-specific convex mixture over multiple latent aspects $z$ . Individual words within the same document are explained by different latent factors. This allows factors to specialize in specific topical sub-vocabularies and naturally accounts for polysemous words by assigning different occurrences of a word to distinct latent classes depending on surrounding document context.
Knowl 8 — Perplexity Reduction Performance of PLSA vs. LSA
empirical result
The predictive performance of PLSA and LSA was evaluated via test set perplexity (log-averaged inverse probability on unseen test words) across latent dimensions on two corpora:
- MED dataset (1033 medical abstracts): PLSA trained with Tempered EM (TEM) reduces perplexity from the unigram baseline of 3073 down to 936 (a $\approx 3.3\times$ reduction), whereas LSA achieves a reduction to 1647 (a $\approx 1.9\times$ reduction).
- LOB dataset (noun-adjective pairs from tagged LOB corpus): PLSA reduces perplexity from the unigram baseline of 1316 down to 547 (a $\approx 2.41\times$ reduction), compared to 632 for LSA (a $\approx 2.08\times$ reduction).
Comparing optimization strategies on the MED corpus shows that TEM fitting achieves substantially lower perplexity than standard EM with early stopping. While standard EM begins to degrade at higher dimensions due to inefficient parameter utilization, TEM maintains monotonically improving performance as the latent space dimensionality $K$ increases, continuing to generalize even when $K$ exceeds the rank of the co-occurrence matrix (evaluated up to $K = 2048$ ).

Knowl 9 — Information Retrieval Precision Comparison Across Document Collections

data/table

The retrieval performance of baseline cosine term frequency matching ( $\cos+\mathrm{tf}$ ), Latent Semantic Indexing (LSI), Probabilistic Latent Semantic Indexing (PLSI), and combined PLSI ( $\text{PLSI}^*$ ) was evaluated on four standard test collections using average precision at 9 recall levels (10% to 90%).

	MED		CRAN		CACM		CISI
Method	Prec. (%)	Impr. (%)	Prec. (%)	Impr. (%)	Prec. (%)	Impr. (%)	Prec. (%)	Impr. (%)
$\cos+\mathrm{tf}$	44.3	–	29.9	–	17.9	–	12.7	–
LSI	51.7	+16.7	28.7*	-4.0	16.0*	-11.6	12.8	+0.8
PLSI	63.9	+44.2	35.1	+17.4	22.9	+27.9	18.8	+48.0
PLSI*	66.3	+49.7	37.5	+25.4	26.8	+49.7	20.1	+58.3

The collections evaluated are:

MED: 1033 medical abstracts from the National Library of Medicine (baseline weight $\lambda = 0.5$ )
CRAN: 1400 aeronautics abstracts from the Cranfield Institute of Technology ( $\lambda = 0.5$ )
CACM: 3204 computer science abstracts from the CACM Journal ( $\lambda = 2/3$ )
CISI: 1460 library science abstracts from the Institute for Scientific Information ( $\lambda = 2/3$ )

An asterisk (*) indicates that LSI achieved no improvement over the baseline at any dimension, in which case the result at 256 dimensions with $\lambda = 2/3$ is shown.

Key results:

PLSI consistently improves over LSI across all four corpora, achieving substantial precision gains over the baseline (+17.4% to +48.0%), whereas LSI degrades baseline performance on CRAN (-4.0%) and CACM (-11.6%).
The multi-model combination $\text{PLSI}^*$ (averaging models trained with $K \in \{32, 48, 64, 80, 128\}$ ) achieves the highest precision across all datasets, delivering gains of +49.7% on MED, +25.4% on CRAN, +49.7% on CACM, and +58.3% on CISI relative to the $\cos+\mathrm{tf}$ baseline.

Coverage note — None was omitted; all core theoretical models, algorithms (EM, TEM), comparisons with LSA and clustering, perplexity experiments, and information retrieval benchmark results have been covered. Specific qualitative word stem tables from the exploratory CLUSTER dataset were omitted as purely illustrative examples.

References

1.J .R. Bellegarda. Exploiting both local and global constraints for multi-span statistical language modeling. In Proceedings of ICASSP'98, volume 2, pages 677-80, 1998.
2.N. Coccaro and D. Jurafsky. Towards better integration of semantic predictors in statistical language modeling. In Proceedings of ICSLP-98, 1998. to appear.
3.S. Deerwester, S. T. Dumais, G. W. Furnas, Landauer. T. K., and R. Harshman. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41, 1990.
4.A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. J. Royal Statist. Soc. B, 39: 1-38, 1977.
5.P.W. Foltz and S. T. Dumais. An analysis of information filtering methods. Communications of the ACM, 35(12):51-60, 1992.
6.T. Hofmann. Probabilistic latent semantic indexing. In Proceedings of SIGIR '99, 1999.
7.T. Hofmann, J. Puzicha, and M. I. Jordan. Unsupervised learning from dyadic data. In Advances in Neural Information Processing Systems, volume 11. MIT Press, 1999.
8.T.K. Landauer and S.T. Dumais. A solution to Plato's problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review, 104(2):211-240, 1997.
9.R.M. Neal and G .E. Hinton. A view of the EM algorithm that justifies incremental and other variants. In M.I. Jordan, editor, Learning in Graphical Models, pages 355-368. Kluwer Academic Publishers, 1998.
10.F.C.N. Pereira, N.Z. Tishby, and L. Lee. Distributional clustering of english words. In Proceedings of the ACL, pages 183-190, 1993.
11.K. Rose, E. Gurewitz, and G. Fox. A deterministic annealing approach to clustering. Pattern Recognition Letters, 11(11):589-594, 1990.
12.G. Salton and M. J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, 1983.
13.L. Saul and F. Pereira. Aggregate and mixed-order Markov models for statistical language processing. In Proceedings of the 2nd International Conference on Empirical Methods in Natural Language Processing, 1997.

Citation

MLA

Hofmann, T. “Probabilistic Latent Semantic Analysis”. Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, 1999, pp. 289–96, https://arxiv.org/abs/1301.6705.

APA

Hofmann, T. (1999). Probabilistic Latent Semantic Analysis. Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, 289–296. https://arxiv.org/abs/1301.6705

Chicago

Hofmann, T. 1999. “Probabilistic Latent Semantic Analysis”. Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, 289–96. https://arxiv.org/abs/1301.6705.

Harvard

Hofmann, T. (1999) “Probabilistic Latent Semantic Analysis”, Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, pp. 289–296. Available at: https://arxiv.org/abs/1301.6705.

Vancouver

1. Hofmann T (1999) Probabilistic Latent Semantic Analysis. In: Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence. pp 289–296

BibTeX

@inproceedings{hofmann1999probabilistic,
  title = {Probabilistic Latent Semantic Analysis},
  author = {Hofmann, Thomas},
  year = {1999},
  booktitle = {Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence},
  pages = {289-296},
  url = {https://arxiv.org/abs/1301.6705},
  eprint = {1301.6705}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Probabilistic Latent Semantic Analysis

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — The Aspect Model for Probabilistic Latent Semantic Analysis

Knowl 2 — Expectation-Maximization Algorithm for Fitting PLSA

Knowl 3 — Matrix Formulation and Geometric Interpretation of PLSA vs. LSA

Knowl 4 — Free Energy Objective and Tempered Posterior in Tempered EM

Knowl 5 — Inverse Annealing Algorithm for Model Fitting via Tempered EM

Knowl 6 — Probabilistic Latent Semantic Indexing and Model Combination

Knowl 7 — Distinction Between the Aspect Model and Distributional Clustering

Knowl 8 — Perplexity Reduction Performance of PLSA vs. LSA

Knowl 9 — Information Retrieval Precision Comparison Across Document Collections

References

Citation

BibTeX

Access the Paper