Labeled LDA: A supervised topic model for credit attribution in multi-labeled corpora

Daniel RamageDavid Leo Wright HallRamesh NallapatiChristopher D. Manning

article2009EMNLP1,480 citations

Introduces Labeled LDA, a supervised extension of Latent Dirichlet Allocation that constrains latent topics to observed document tags to directly learn word-tag correspondences and extract tag-specific snippets.

Listen

Modern digital collections, from collaborative web portals to enterprise knowledge bases, rely heavily on multi-tagging to organize content. However, human-provided tags typically apply with uneven specificity across different parts of a document rather than describing the entire text uniformly. Standard unsupervised topic models like Latent Dirichlet Allocation (LDA) fail to address this credit attribution challenge because they do not align learned latent themes with predefined human labels. Meanwhile, traditional classification algorithms such as Support Vector Machines (SVMs) lack the internal word-level attribution necessary to pinpoint precisely which portions of a document correspond to which tags.

The article introduces and evaluates Labeled LDA (L-LDA), a supervised probabilistic topic model designed to solve the credit attribution problem by constraining each document’s latent topics to match only its assigned user tags in a direct, one-to-one correspondence.

To evaluate the model, the researchers conducted empirical analyses across multiple tagging and classification benchmarks, including 4,000 tagged web pages from del.icio.us covering 20 distinct topics and multi-category datasets from the Yahoo directory. The approach uses collapsed Gibbs sampling for parameter learning and inference, comparing performance against standard unsupervised LDA and baseline one-versus-rest SVM classifiers across visualization, snippet extraction, and document-level multi-label classification tasks.

The analysis produced three primary findings. First, for tag-specific snippet extraction, human evaluators preferred L-LDA's extracted text passages over SVM outputs by more than three to one (72 preferred cases versus 21 out of 149 evaluated pairs, with 24 unanimous approvals versus only 2 for SVMs). Second, unlike standard LDA, which over-allocates topics to dominant themes and misses rare tags entirely, L-LDA guarantees full coverage and direct human interpretability for every assigned tag. Third, in multi-label classification on naturally multi-tagged web corpora, L-LDA outperformed SVM baselines, raising overall Micro-F1 accuracy from 39.33% to 52.12% while maintaining comparable Macro-F1 performance.

These findings indicate that directly integrating human label supervision into generative topic models delivers significant utility for enterprise search and document retrieval interfaces. By accurately mapping specific words and passages to defined tags, organizations can provide richer visual overviews, automated content summarization, and highly targeted snippet previews that reduce the time users spend searching for relevant information.

Organizations handling multi-labeled document repositories should consider deploying L-LDA for tag-focused passage extraction, contextual visualization, and classification. However, decision-makers should note that test-time classification in the current implementation relies on an unconstrained sampling approximation across all possible categories, and the model does not yet natively model label correlations. Prior to large-scale production deployment on complex multi-label classification pipelines, teams should validate performance on their domain-specific taxonomies and monitor ongoing developments in exact sampling algorithms and correlated topic modeling.

Stanford University electronic document multi-label classification social bookmarking Labeled LDA credit attribution multi-labeled corpora

Ramage et al (2009).pdf

Paper: Latent Dirichlet Allocation, David M. Blei et al. (2003). It introduces Latent Dirichlet Allocation, the core generative probabilistic framework that Labeled LDA directly constrains and builds upon for credit attribution.
Paper: Supervised Topic Models, David M. Blei et al. (2007). It establishes the foundational supervised topic modeling paradigm (sLDA), providing key background on incorporating document-level supervisory signals into LDA.
Paper: The Author-Topic Model for Authors and Documents, Michal Rosen-Zvi et al. (2004). It demonstrates how to constrain topic distributions to specific document metadata (authors), serving as conceptual precursor to mapping topics directly to user-defined document labels.
Paper: A kernel method for multi-labelled classification, A. Elisseeff et al. (2001). It formulates the problem of multi-label classification and ranking baselines against which generative multi-label models like Labeled LDA are benchmarked.
Paper: Probabilistic latent semantic indexing, Thomas Hofmann (1999). It introduces probabilistic latent semantic indexing, the foundational aspect model predecessor to LDA and topic-based credit attribution.

Paper: Latent Dirichlet allocation (LDA) and topic modeling: models, applications, a survey, Hamed Jelodar et al. (2017). It provides a broad retrospective survey of LDA-based extensions and topic modeling developments, contextualizing supervised and multi-label variations like Labeled LDA.
Paper: A Review on Multi-Label Learning Algorithms, Min-Ling Zhang et al. (2014). It provides a comprehensive review of multi-label learning algorithms, structuring problem-transformation and algorithm-adaptation methods that contextualize Labeled LDA's classification capabilities.
Paper: Collaborative topic modeling for recommending scientific articles, Chong Wang et al. (2011). It extends topic modeling into user-tagged bookmarking domains by pairing probabilistic topic representations with collaborative filtering for personalized recommendations.
Paper: Online Learning for Latent Dirichlet Allocation, Matthew D. Hoffman et al. (2010). It develops scalable online variational inference for LDA, addressing the computational bottlenecks present in batch inference for large-scale labeled document corpora.
Paper: Optimizing Semantic Coherence in Topic Models, David Mimno et al. (2011). It develops automated metrics and optimization for topic coherence, directly addressing topic interpretability issues observed in topic modeling representations.

Cover for Labeled LDA: A supervised topic model for credit attribution in multi-labeled corpora

Abstract

A significant portion of the world's text is tagged by readers on social bookmarking websites. Credit attribution is an inherent problem in these corpora because most pages have multiple tags, but the tags do not always apply with equal specificity across the whole document. Solving the credit attribution problem requires associating each word in a document with the most appropriate tags and vice versa. This paper introduces Labeled LDA, a topic model that constrains Latent Dirichlet Allocation by defining a one-to-one correspondence between LDA's latent topics and user tags. This allows Labeled LDA to directly learn word-tag correspondences. We demonstrate Labeled LDA's improved expressiveness over traditional LDA with visualizations of a corpus of tagged web pages from del.icio.us. Labeled LDA outperforms SVMs by more than 3 to 1 when extracting tag-specific document snippets. As a multi-label text classifier, our model is competitive with a discriminative baseline on a variety of datasets.

Topics

electronic document multi-label classification social bookmarking topic model visualization web content text classifiers

Keywords

Labeled LDA credit attribution multi-labeled corpora supervised topic model Latent Dirichlet Allocation multi-label text classification Gibbs sampler snippet extraction

1 Introduction
2 Labeled LDA
2.1 Learning and inference
2.2 Relationship to Naive Bayes
3 Credit attribution within tagged documents
4 Topic Visualization
5 Tagged document visualization
6 Snippet Extraction
7 Multilabeled Text Classification
7.1 Yahoo
7.2 Tagged Web Pages
8 Discussion
9 Conclusion
Acknowledgments
References

Knowls

Knowl 1 — Generative Process of Labeled LDA
model/method
Labeled Latent Dirichlet Allocation (L-LDA) is a supervised generative topic model for multi-labeled document collections that constrains Latent Dirichlet Allocation by defining a one-to-one correspondence between latent topics and human-provided labels. Let $K$ denote the total number of unique labels in the corpus and $V$ the vocabulary size.

Each document $d$ consists of a list of word tokens $\mathbf{w}^{(d)} = (w_1, \dots, w_{N_d})$ with $w_i \in \{1, \dots, V\}$ , and an observed binary label presence vector $\mathbf{\Lambda}^{(d)} = (l_1, \dots, l_K) \in \{0, 1\}^K$ . Let $\lambda^{(d)} = \{k \in \{1, \dots, K\} \mid l_k = 1\}$ be the index set of active labels for document $d$ , and let $M_d = |\lambda^{(d)}|$ be the number of active labels.

The generative process proceeds as follows:
1. For each topic/label $k \in \{1, \dots, K\}$ , draw a multinomial word distribution: $\beta_k \sim \text{Dirichlet}(\eta)$ where $\eta \in \mathbb{R}^V_{>0}$ is the symmetric or asymmetric Dirichlet word prior.
2. For each document $d \in \{1, \dots, D\}$ :
a. For each topic $k \in \{1, \dots, K\}$ , generate label indicator $\Lambda_k^{(d)} \sim \text{Bernoulli}(\Phi_k)$ , where $\Phi_k$ is the prior probability of label $k$ .

b. Construct an $M_d \times K$ label projection matrix $L^{(d)}$ whose entries are: $L_{ij}^{(d)} = \begin{cases} 1 & \text{if } \lambda_i^{(d)} = j \\ 0 & \text{otherwise} \end{cases}$ for row $i \in \{1, \dots, M_d\}$ and column $j \in \{1, \dots, K\}$ .

c. Compute the document-specific topic prior parameter vector $\alpha^{(d)} \in \mathbb{R}^{M_d}$ by projecting the global corpus topic prior $\alpha = (\alpha_1, \dots, \alpha_K)^T$ : $\alpha^{(d)} = L^{(d)} \alpha = (\alpha_{\lambda_1^{(d)}}, \dots, \alpha_{\lambda_{M_d}^{(d)}})^T$

d. Draw the document topic mixture proportion vector restricted to active labels: $\theta^{(d)} \sim \text{Dirichlet}(\alpha^{(d)})$

e. For each word token position $i \in \{1, \dots, N_d\}$ in document $d$ :
- Draw a topic assignment $z_i \in \lambda^{(d)}$ from the restricted mixture: $z_i \sim \text{Multinomial}(\theta^{(d)})$
- Draw word token $w_i \in \{1, \dots, V\}$ from the corresponding topic distribution: $w_i \sim \text{Multinomial}(\beta_{z_i})$
Knowl 2 — Collapsed Gibbs Sampling for Labeled LDA Training
algorithm
When document labels $\mathbf{\Lambda}^{(d)}$ are observed during training, the Bernoulli label priors $\Phi$ are d-separated from the rest of the network. Topic inference over word assignments $\mathbf{z}$ is performed using collapsed Gibbs sampling, where the latent topic $z_i$ at word position $i$ in document $d$ is restricted to the document's active label set $\lambda^{(d)} = \{k \mid \Lambda_k^{(d)} = 1\}$ .

The conditional posterior distribution for assigning topic $j \in \lambda^{(d)}$ to token position $i$ with observed word type $w_i$ , conditioned on all other assignments $\mathbf{z}_{-i}$ , is:

$P(z_i = j \mid \mathbf{z}_{-i}, \mathbf{w}, \mathbf{\Lambda}^{(d)}) \propto \frac{n_{-i, j}^{w_i} + \eta_{w_i}}{n_{-i, j}^{(\cdot)} + \sum_{v=1}^V \eta_v} \times \frac{n_{-i, j}^{(d)} + \alpha_j}{n_{-i, \cdot}^{(d)} + \sum_{k \in \lambda^{(d)}} \alpha_k}$

where $n_{-i, j}^{w_i}$ is the count of word type $w_i$ assigned to topic $j$ corpus-wide excluding current position $i$ , $n_{-i, j}^{(\cdot)} = \sum_{v=1}^V n_{-i, j}^v$ is the total count of all words assigned to topic $j$ excluding position $i$ , $n_{-i, j}^{(d)}$ is the count of words in document $d$ assigned to topic $j$ excluding position $i$ , and $n_{-i, \cdot}^{(d)} = \sum_{k \in \lambda^{(d)}} n_{-i, k}^{(d)}$ is the total number of words in document $d$ excluding position $i$ .
Input: Word tokens $\mathbf{w}^{(d)}$ and label indicator sets $\lambda^{(d)}$ for documents $d \in \{1, \dots, D\}$ ; Dirichlet hyperparameters $\alpha, \eta$ ; number of iterations $T$
Output: Word-topic counts $n_j^w$ , document-topic counts $n_j^{(d)}$ , and word topic assignments $\mathbf{z}$
Initialize $z_i^{(d)}$ uniformly at random from $\lambda^{(d)}$ for each document $d$ and token $i \in \{1, \dots, N_d\}$
Initialize count matrices $n_j^w$ and $n_j^{(d)}$ based on initial assignments
for iter = 1 to $T$ do
    for each document $d \in \{1, \dots, D\}$ do
        for each token index $i \in \{1, \dots, N_d\}$ do
             $w \leftarrow w_i^{(d)}$
             $j_{\text{old}} \leftarrow z_i^{(d)}$
            Decrement $n_{j_{\text{old}}}^w$ and $n_{j_{\text{old}}}^{(d)}$

            for each candidate label $j \in \lambda^{(d)}$ do
                 $P(j) \leftarrow \frac{n_j^w + \eta_w}{\sum_{v=1}^V (n_j^v + \eta_v)} \times \frac{n_j^{(d)} + \alpha_j}{\sum_{k \in \lambda^{(d)}} (n_k^{(d)} + \alpha_k)}$
            end for

            Normalize $P(j)$ across all $j \in \lambda^{(d)}$
            Sample new topic assignment $j_{\text{new}} \sim P(\cdot)$
             $z_i^{(d)} \leftarrow j_{\text{new}}$
            Increment $n_{j_{\text{new}}}^w$ and $n_{j_{\text{new}}}^{(d)}$
        end for
    end for
end for
Knowl 3 — Equivalence and Differences Between Labeled LDA and Multinomial Naive Bayes
theoretical result
Labeled LDA generalizes the generative event model of a Multinomial Naive Bayes classifier by introducing latent topic mixtures over observed labels:
1. Singly Labeled Documents ( $M_d = 1$ ): If every document $d$ in the corpus has exactly one label $l_d$ , the label set constraint forces $z_i = l_d$ for all word positions $i \in \{1, \dots, N_d\}$ . In this setting, collapsed Gibbs sampling increments the count of word $w_i$ for topic $l_d$ deterministically. The probability of document words $P(\mathbf{w}^{(d)} \mid l_d)$ under Labeled LDA is mathematically identical to the document likelihood under a standard Multinomial Naive Bayes event model trained on the same instances.
2. Multiply Labeled Documents ( $M_d > 1$ ): In a standard one-versus-rest Multinomial Naive Bayes model, a document labeled with multiple tags contributes a full count of 1 for each of its words to every label classifier independently. In contrast, Labeled LDA treats each document as a Dirichlet-distributed mixture over its observed labels $\lambda^{(d)}$ , distributing the credit mass of each individual word occurrence across only one topic assignment $z_i \in \lambda^{(d)}$ at a time via latent mixture sampling.
Knowl 4 — Unlabeled Document Inference in Labeled LDA for Multi-Label Classification
model/method
For multi-label document classification at test time, the true label set $\mathbf{\Lambda}^{(d)}$ is unobserved. Exact maximum a posteriori label set prediction requires evaluating posterior probabilities over all $2^K$ binary configurations of $\mathbf{\Lambda}^{(d)}$ , where the parameter vector $\alpha^{(d)}$ changes support with each subset.

To perform scalable test-time inference, the model is relaxed to standard Latent Dirichlet Allocation over the full $K$ -dimensional topic space using the vocabulary topic distributions $\beta_1, \dots, \beta_K$ learned during supervised training. Each test word $w_i$ can sample any topic $z_i \in \{1, \dots, K\}$ using standard LDA collapsed Gibbs sampling. The document's posterior topic distribution $\theta^{(d)} = (\theta_1^{(d)}, \dots, \theta_K^{(d)})$ is estimated by normalizing its sampled topic assignments:

$\theta_k^{(d)} = \frac{n_k^{(d)} + \alpha_k}{\sum_{j=1}^K (n_j^{(d)} + \alpha_j)}$

Predicted labels for the document are then determined by assigning label $k$ if its posterior topic probability $\theta_k^{(d)}$ exceeds a tuned global threshold.
Knowl 5 — Tag-Specific Snippet Extraction Methodology
experimental setup
Tag-specific snippet extraction aims to select a passage of a document that best reflects the document's content from the perspective of a particular assigned tag $k$ .

Given a document $d$ and a target tag $k \in \lambda^{(d)}$ , all candidate 15-word contiguous text windows are evaluated:
- Labeled LDA Scoring: L-LDA scores a candidate window $W = (w_1, \dots, w_{15})$ by the expected probability that the target tag $k$ generated each word token in the window: $\text{Score}_{\text{L-LDA}}(W, k) = \prod_{t=1}^{15} P(w_t \mid z_t = k) = \prod_{t=1}^{15} \beta_{k, w_t}$
- SVM Baseline Scoring: A collection of one-versus-rest linear Support Vector Machines (SVMs) is trained, one for each tag. Each 15-word window $W$ is represented as an independent bag-of-words pseudo-document and scored using the un-thresholded continuous decision value output by the target tag's SVM classifier.
For each method, the 15-word window achieving the maximum score for tag $k$ in document $d$ is extracted.

Knowl 6 — Human Evaluation of Tag-Specific Snippet Extraction

empirical result

Tag-specific snippet extraction was evaluated on 29 web pages from del.icio.us with at least two tags from a 20-tag vocabulary, forming 149 (document, tag) evaluation pairs. Three human judges rated 15-word snippets generated by Labeled LDA versus one-versus-rest SVMs in a randomized, blinded study.

Out of 149 total evaluations, L-LDA was judged superior in 72 cases, compared to 21 cases for SVM (with remaining cases judged as ties). The performance difference is statistically significant ( $p < 0.001$ , sign test).

On the subset of 51 document-tag pairs evaluated independently by all three judges, inter-annotator agreement reached Fleiss' $\kappa = 0.63$ . Within this 51-pair subset:

L-LDA was selected as superior by at least one annotator in 33 of 51 cases, and unanimously selected by all three annotators in 24 of 51 cases.
SVM was selected by at least one annotator in 10 of 51 cases, and unanimously selected in only 2 of 51 cases.

Knowl 7 — Multi-Label Classification Performance on Tagged del.icio.us Web Pages

empirical result

Multi-label document classification was evaluated on 4,000 web pages crawled from del.icio.us labeled across 20 medium-to-high frequency tags. The corpus contains an average of 781 non-stop words and 4 tags per document, with over 89% of documents possessing multiple labels.

Models were trained on 80% of the documents and evaluated on the remaining 20% across 10 random train/test splits. L-LDA parameters (classification threshold and Dirichlet scaling constants) were tuned via validation, and one-versus-rest SVMs used term frequency features and cost parameter $C = 10.0$ tuned via 4-fold cross-validation.

Labeled LDA achieves a comparable Macro-F1 score to one-versus-rest SVMs ( $39.85\%$ vs $39.00\%$ ) while outperforming SVMs on Micro-F1 by $12.79\%$ ( $52.12\%$ vs $39.33\%$ ). Both improvements are statistically significant under a two-tailed paired t-test at the 95% confidence level.

Knowl 8 — Multi-Label Text Classification Performance on Yahoo Directory Datasets

data/table

Multi-label classification performance was evaluated across 8 top-level categories from the Yahoo directory, where each document is assigned to one or more subcategories within the domain. Performance represents the mean (and standard deviation) over 10 independent runs.

Dataset	% Macro-F1		% Micro-F1
	L-LDA	SVM	L-LDA	SVM
Arts	30.70 (1.62)	23.23 (0.67)	39.81 (1.85)	48.42 (0.45)
Business	30.81 (0.75)	22.82 (1.60)	67.00 (1.29)	72.15 (0.62)
Computers	27.55 (1.98)	18.29 (1.53)	48.95 (0.76)	61.97 (0.54)
Education	33.78 (1.70)	36.03 (1.30)	41.19 (1.48)	59.45 (0.56)
Entertainment	39.42 (1.38)	43.22 (0.49)	47.71 (0.61)	62.89 (0.50)
Health	45.36 (2.00)	47.86 (1.72)	58.13 (0.43)	72.21 (0.26)
Recreation	37.63 (1.00)	33.77 (1.17)	43.71 (0.31)	59.15 (0.71)
Society	27.32 (1.24)	23.89 (0.74)	42.98 (0.28)	52.29 (0.67)

On the full Yahoo dataset—where only 33% of documents are multiply labeled—L-LDA outperforms one-versus-rest SVMs on Macro-F1 on 5 out of 8 categories (Arts, Business, Computers, Recreation, Society), but SVMs achieve higher Micro-F1 across all 8 datasets.

When evaluated specifically on the subset of Yahoo documents having more than one label in training, L-LDA beats SVMs on Macro-F1 on 4 datasets (1 SVM win, 3 ties) and on Micro-F1 on 4 datasets (with 4 ties), reflecting L-LDA's credit assignment advantage on multi-labeled documents.

Knowl 9 — Topic Specialization and Label Alignment in Labeled LDA versus Unsupervised LDA
empirical result
When trained on the 4,000-document del.icio.us corpus containing 20 distinct tags, standard unsupervised Latent Dirichlet Allocation (with $K = 20$ ) and Labeled LDA exhibit contrasting topic allocations:
- Unsupervised LDA: Allocates multiple redundant topic vectors to high-frequency semantic areas in the corpus (learning multiple distinct topics mapped by cosine similarity to web, culture, computer, reference, and politics), while failing to discover topics corresponding to lower-frequency tags (discovering zero topics for books, english, science, history, grammar, java, and philosophy).
- Labeled LDA: Explicitly constrains each of the 20 topics to correspond directly to one user tag in a one-to-one mapping, ensuring that low-frequency tags maintain dedicated word multinomial distributions $\beta_k$ and eliminating the need for post-hoc topic labeling.

Coverage note — No substantial contributed material was omitted.

References

1.D. M. Blei and J. Lafferty. 2006. Correlated Topic Models. NIPS, 18:147.
2.D. Blei and J McAuliffe. 2007. Supervised Topic Models. In NIPS, volume 21.
3.D. M. Blei, A.Y. Ng, and M.I. Jordan. 2003. Latent Dirichlet allocation. JMLR.
4.T. L. Griffiths and M. Steyvers. 2004. Finding scientific topics. PNAS, 1:5228–35.
5.P. Heymann, G. Koutrika, and H. Garcia-Molina. 2008. Can social bookmarking improve web search. In WSDM.
6.S. Ji, L. Tang, S. Yu, and J. Ye. 2008. Extracting shared subspace for multi-label classification. In KDD, pages 381–389, New York, NY, USA. ACM.
7.H. Kazawa, H. Taira T. Izumitani, and E. Maeda. 2004. Maximal margin labeling for multi-topic text categorization. In NIPS.
8.S. Lacoste-Julien, F. Sha, and M. I. Jordan. 2008. DiscLDA: Discriminative learning for dimensionality reduction and classification. In NIPS, volume 22.
9.D. D. Lewis, Y. Yang, T. G. Rose, G. Dietterich, F. Li, and F. Li. 2004. RCV1: A new benchmark collection for text categorization research. JMLR, 5:361–397.
10.Wei Li and Andrew McCallum. 2006. Pachinko allocation: Dag-structured mixture models of topic correlations. In International conference on Machine learning, pages 577–584.
11.A. McCallum and K. Nigam. 1998. A comparison of event models for naive bayes text classification. In AAAI-98 workshop on learning for text categorization, volume 7.
12.Q. Mei, X. Shen, and C Zhai. 2007. Automatic labeling of multinomial topic models. In KDD.
13.D. Ramage, P. Heymann, C. D. Manning, and H. Garcia-Molina. 2009. Clustering the tagged web. In WSDM.
14.N. Ueda and K. Saito. 2003. Parametric mixture models for multi-labeled text includes models that can be seen to fit within a dimensionality reduction framework. In NIPS.

Citation

MLA

Ramage, D., et al. “Labeled LDA”. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP '09, vol. 1, 2009, p. 248, https://doi.org/10.3115/1699510.1699543.

APA

Ramage, D., Hall, D., Nallapati, R., & Manning, C. D. (2009). Labeled LDA. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP '09, 1, 248. https://doi.org/10.3115/1699510.1699543

Chicago

Ramage, D., D. Hall, R. Nallapati, and C. D. Manning. 2009. “Labeled LDA”. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP '09 1: 248. https://doi.org/10.3115/1699510.1699543.

Harvard

Ramage, D. et al. (2009) “Labeled LDA”, Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP '09. Association for Computational Linguistics, p. 248. Available at: https://doi.org/10.3115/1699510.1699543.

Vancouver

1. Ramage D, Hall D, Nallapati R, Manning CD (2009) Labeled LDA. In: Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP '09. Association for Computational Linguistics, p 248

BibTeX

@inproceedings{Ramage_2009, series={EMNLP ’09}, title={Labeled LDA: a supervised topic model for credit attribution in multi-labeled corpora}, volume={1}, url={http://dx.doi.org/10.3115/1699510.1699543}, DOI={10.3115/1699510.1699543}, booktitle={Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing Volume 1 - EMNLP ’09}, publisher={Association for Computational Linguistics}, author={Ramage, Daniel and Hall, David and Nallapati, Ramesh and Manning, Christopher D.}, year={2009}, pages={248}, collection={EMNLP ’09} }

Metadata:Crossref

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-nc-sa/4.0/

Model	Best Snippet	Unanimous Agreement
Labeled LDA	72 / 149	24 / 51
SVM	21 / 149	2 / 51

Model	Macro-F1 (%)	Micro-F1 (%)
Labeled LDA	39.85 ( 0.989)	52.12 ( 0.434)
SVM	39.00 ( 0.423)	39.33 ( 0.574)

Labeled LDA: A supervised topic model for credit attribution in multi-labeled corpora

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Generative Process of Labeled LDA

Knowl 2 — Collapsed Gibbs Sampling for Labeled LDA Training

Knowl 3 — Equivalence and Differences Between Labeled LDA and Multinomial Naive Bayes

Knowl 4 — Unlabeled Document Inference in Labeled LDA for Multi-Label Classification

Knowl 5 — Tag-Specific Snippet Extraction Methodology