Pith. sign in

REVIEW 4 major objections 7 minor 87 references

E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper provides the first practical algorithms with provable guarantees for inferring LDA topic-word assignments, achieving a 1−1/e approximation in logarithmic parallel time.

desk verdict A genuinely useful submodular reformulation of sparsity-constrained topic assignment with solid guarantees on the surrogate, packaged with an overstated claim to solve LDA's MAP problem. read the letter →

arxiv 2506.07747 v1 pith:VHFYXMQU submitted 2025-06-06 cs.LG cs.CLstat.ML

classification cs.LGcs.CLstat.ML
keywords latentDirichletallocationtopicmodelsMAPinferencesubmodularmaximizationgreedyalgorithmparallelalgorithmsinterpretabilitycausalwithtext
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Latent Dirichlet Allocation is the workhorse topic model of the social sciences, but every practical algorithm for its core inference problem—assigning each word in each document to the topic that probably generated it—has come without approximation guarantees, because the problem is NP-hard. This paper argues that the difficulty disappears once the problem is viewed in the right limit: when the Dirichlet sparsity prior is taken to infinity and replaced by a hard per-document topic budget, the MAP assignment objective becomes a monotone submodular set function over topic–document links. Maxima of such functions are known to be approximable within 1−1/e of optimal by simple greedy algorithms, and the paper supplies fast serial and parallel implementations whose per-iteration cost is logarithmic in the number of documents. It further shows that the candidate topics in the greedy search can be generated from standard coherence functions, so every returned topic is formally labeled by a keyword. If the guarantees hold, researchers get something they did not have before: interpretable, near-optimal, out-of-sample LDA assignments that are compatible with downstream causal inference.

What carries the argument

The central object is the set function f(E) = Σ_{d∈D} Σ_{i=1}^{|d|} max_{τ∈E_d} log φ_τ[w_{d,i}] over topic–document links, with a placeholder topic added to each document so the function is defined everywhere. Its two structural properties do the work: monotonicity (adding links never hurts) and submodularity (the marginal value of a link decreases as the set of links grows), which together trigger the classical greedy 1−1/e guarantee. The implementation machinery is the memoization of per-word log-probabilities and a max-heap of per-document best candidate links, which reduce each greedy iteration to O(log |D| + ℓ|Φ|) and make the parallel adaptive-sampling version simulate each submodular query in O(ℓ) or O(ℓ|Φ|) time.

What would settle it

Generate a synthetic corpus from a two-topic LDA where each topic is a uniform mixture over a distinct set of 100 vocabulary words, so that no single-keyword candidate topic matches either true topic; run E-LDA with the standard keyword-based candidate set and compare its log-posterior to the brute-force MAP optimum. If the achieved value falls below 1−1/e of the optimum, or the recovered assignments diverge qualitatively from the true mixture structure, the central guarantee is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the MAP topic-word assignment problem of LDA—previously known to be NP-hard even for one document—becomes a tractable combinatorial optimization problem when formulated in the infinite-Dirichlet limit under an explicit cardinality constraint. In that limit (Theorem 1), the posterior maximization is exactly the maximization of a sum of per-word log-likelihoods. Introducing the set of topic–document links E and defining f(E) = Σ_{d∈D} Σ_{i=1}^{|d|} max_{τ∈E_d} log φ_τ[w_{d,i}] makes the objective a monotone submodular function (up to a placeholder-topic normalization), so the standard greedy algorithm returns a deterministic 1−1/e approximation to the constrained optimum (Theorem 2). The algorithmic contribution is to make greedy practical: by exploiting the conditional independence of documents, each marginal value depends only on the word-level log probabilities of a single document, so marginal updates can be memoized and organized in a heap, reducing the per-iteration cost from cubic in |D| to O(log |D| + ℓ|Φ|). With the same techniques, the parallel adaptive-sampling algorithm yields a 1−1/e−4ε approximation in O($ε^{{-2}}$ log(|Φ||D|) · $ℓ^{2}$) adaptive rounds. The paper also constructs candidate topic sets from coherence functions such as Exp-UMass and co-occurrence counts, yielding topics that are formally associated with a keyword, and shows that the out-of-sample version of the objective restores the SUTVA independence required for causal inference.

Load-bearing premise

The guarantee holds only if the true topics are already contained in the pre-specified candidate set, and only if the data-generating process is close to the infinite-Dirichlet limit in which each document draws from few topics; when true topics are diffuse mixtures of many words, the returned assignments can be arbitrarily far from the true MAP assignments.

Editorial extensions

If this is right

  • For the first time, LDA document-topic assignments can be computed with a deterministic 1−1/e approximation guarantee under an explicit sparsity constraint, in serial time that is logarithmic per iteration.
  • A single run of FastGreedy-E-LDA provides near-optimal solutions for every average sparsity from 1 to κ topics per document, so practitioners can compare sparse to dense summaries without refitting.
  • Because candidate topics are generated from standard coherence functions, every topic in the solution is formally associated with a keyword, ending the manual 'reading tea leaves' step of topic-model practice.
  • The out-of-sample version of the objective preserves the SUTVA independence needed for causal inference frameworks that treat topics as treatments, giving those designs provable approximation guarantees.
  • The techniques extend to matroid and packing constraints, allowing researchers to encode domain-specific structure into the topic model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One step beyond the paper: if the infinite-Dirichlet idealization holds approximately, the greedy selection of topic–document links acts as a sparse encoder for documents, which could be tested by measuring how well training-set assignments transfer to held-out documents relative to Gibbs assignments.
  • One step beyond the paper: the keyword-labeled candidate sets suggest a cheap interpretability audit—compare the top-ranked words of each selected topic to its keyword label and report disagreements as a diagnostic.
  • One step beyond the paper: in corpora where true topics are broad mixtures rather than keyword-specific profiles, single-keyword candidate sets may underperform; enriching Φ with mixture topics would preserve submodularity and may close that gap.
  • One step beyond the paper: because the objective is a sum of per-document maxima, the out-of-sample assignment routine can process documents independently, which matters for privacy-preserving or federated text analysis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes E-LDA, a combinatorial approach to inferring per-word topic assignments in LDA. The authors define a sparsity-constrained objective f(E) that sums, for each document and word, the maximum log-probability over a candidate set of topics linked to the document. They prove this objective is monotone submodular, so standard Greedy yields a 1−1/e approximation; they then provide a fast serial algorithm (FastGreedy-E-LDA) and a low-adaptivity parallel algorithm (FAST-based) with the same guarantee. They also propose generating the candidate set of topics from coherence functions so that each selected topic is associated with a known keyword, and they argue that out-of-sample inference can preserve the independence assumptions needed for causal inference. Experiments on three datasets compare E-LDA's objective values and topic coherence against Gibbs LDA variants, neural topic models, and BERTopic.

Significance. If the claims are properly scoped, the algorithmic contribution is solid: a natural sparsity-constrained objective for topic-word assignment is shown to be monotone submodular, the fast algorithms are nontrivial and the code is publicly available, and the experiments span several baselines and datasets. The per-document and out-of-sample guarantees are potentially useful for social-science workflows that require document-level topic labels. However, the current statement of the contribution overreaches: the guarantees apply to a surrogate α→∞ sparsity-constrained likelihood, not to the finite-α LDA MAP problem, and the coherence evaluation is partly circular because the candidate topics and the evaluation metric are both based on co-occurrence statistics. These issues materially affect the strength of the paper's central claims, though they are addressable through reframing and additional evaluation.

major comments (4)
  1. [Section 3, Theorems 1–2] The theoretical guarantees are for the sparsity-constrained surrogate f(E) in eq. 3, which is the α→∞ limit of the LDA log-posterior plus an added cardinality constraint |E| ≤ κ|D|. They are not guarantees for the LDA MAP problem of Sontag and Roy (2011), i.e., maximization of eq. 1 at finite α, and no finite-α version of the guarantee is supplied. The abstract and introduction nevertheless claim 'near-optimal posterior probability' and 'solve the LDA MAP topic-word assignment problem'; these claims are unsupported. The paper should be reframed as solving a new, well-defined sparsity-constrained likelihood problem, with the α→∞ correspondence stated as motivation rather than as an equivalence.
  2. [Section 7 and Theorem 2] The 1−1/e approximation holds only with respect to a pre-specified candidate set Φ. If the true latent topics are not members of Φ—for example, if a true topic is a mixture of many keywords rather than a single keyword's co-occurrence profile—the returned assignments can be arbitrarily far from the true MAP assignments. The paper's construction of Φ from coherence functions makes this assumption substantive, not a technicality. Please state this assumption explicitly in the main results and discuss its consequences; the current text only notes it in passing in Section 7.
  3. [Section 9 and Appendix Q] The semantic-quality evaluation is partially circular. Candidate topics are generated by co-occurrence-based functions (Co-occurrence, Exp-UMass), and the evaluation metric (UMass coherence) is also a co-occurrence statistic; topics selected to maximize the likelihood under these generators will tend to score high on UMass coherence even if they are not semantically meaningful. The control with the unmodified UMass generator does not resolve this, because all three generators share the same co-occurrence-based ranking of words (as the paper notes in Appendix Q); the control only shows that the exponential transform changes which topics are selected. To support the 'higher semantic quality' claim, the paper should include a human evaluation or validate with a coherence measure less directly tied to the generating functions (e.g., NPMI or embedding-based coherence).
  4. [Theorem 2, first bullet] The phrase 'the log posterior value of the (optimal) LDA MAP topic-word assignments (eq. 3)' mislabels eq. 3. Equation 3 is the sparsity-constrained surrogate objective defined in this paper, not the LDA posterior of eq. 1. This conflation recurs in the abstract and in Section 4 and should be corrected throughout.
minor comments (7)
  1. [Section 5, FastInitialize] The parameter ε in the initialization P ← log(ε)·D is not defined in the main text; please clarify its value and role.
  2. [Section 5, Update] The operation 'σd∗ ← Φ ⊙ [D[d∗, :]; . . .; D[d∗, :]]' is under-specified; please state the dimensions and the intended elementwise semantics.
  3. [Appendix B] The displayed asymptotic property contains a typo: the term should be a difference of log Γ values, not log Γ(kq+r)/log(Γ(kq)) − r log(kq); the subsequent derivation is correct but the displayed statement is not.
  4. [Theorem 1] The limit is stated only as α→∞; please specify that α is scalar or that ατ = α for all τ, since the proof assumes a common α.
  5. [Theorem 2] The per-iteration complexity O(log|D| + ℓ|Φ|) uses ℓ without a definition; please define ℓ as the maximum document length (as in Appendix F).
  6. [Figure 1] The figure contains duplicated paragraphs of text from Sections 3 and 4 in the rendered version; please clean the figure so that it only shows the intended panels.
  7. [Section 9] The claim that E-LDA CO-OCCURRENCE's mean coherence outperforms the best single topic's coherence of all Gibbs baselines would be easier to assess if the figure plotted the distribution (e.g., box plots) of baseline topic-level coherence, rather than only means.

Circularity Check

1 steps flagged · score 4.0 of 10

Theoretical guarantee chain is largely self-contained; the main circularity is a metric-aligned coherence evaluation, while the alpha-infinity surrogate claim is a scope overclaim rather than a circular reduction.

  1. self definitional [Section 7 (Co-occurrence topic generator) and Section 9 / Appendix P, Eq. 19 (UMass coherence evaluation)]
    "Co-occurrence: ... the dot product counts word co-occurrences. So, we generate topics via: φ^co-occurrence_{w*}[v] ∝ exp(|D_{w*,v}| + ε). ... We measure topics' semantic quality via standard UMass coherence. ... for a given topic τ, the normalized coherence is: C(τ,h*) = ... log |D_{τ_h,τ_l}| + ε / |D_{τ_l}|."

    The topic generator and the evaluation metric are both computed from the same corpus co-occurrence counts |D_{w,v}|. Co-occurrence topics rank words by exp(|D_{w*,v}|), so a selected topic's top words are exactly those that co-occur with the label word w*. UMass coherence then rewards topics whose top words co-occur, using the analogous ratio log |D_{τ_h,τ_l}|/|D_{τ_l}|. Thus the reported 'higher semantic quality' of E-LDA Co-occurrence in Fig. 2a is substantially inherited from the alignment between the generator and the metric, not from an independent semantic judgment. The paper's unmodified-UMass control shows that not every generator benefits, but it does not remove the construction-level alignment for the exponential and co-occurrence generators.

full rationale

The core theoretical derivation — Theorem 1's α→∞ limit, submodularity of the objective f, the deterministic 1−1/e greedy guarantee, and the memoization-based speedups — is internally consistent and does not reduce to its inputs. The use of Breuer et al. (2020) for the parallel FAST algorithm is legitimate self-citation: that prior theorem is a general submodular-maximization result with independent content, and the present paper contributes the E-LDA-specific query simulation (Theorem 4). The central circularity concern is the experimental semantic-quality claim: candidate topics are generated from co-occurrence statistics while the evaluation uses UMass coherence, which is itself co-occurrence-based. This makes the coherence advantage partially definitional, although the control with the unmodified UMass generator mitigates the concern. Separately, the abstract's claim of 'near-optimal posterior probability' for LDA goes beyond what is proved, since the guarantees apply to the α→∞ likelihood with an added hard cardinality constraint rather than to the finite-α LDA posterior; however, the paper states this limit explicitly and this is a scope overclaim, not a circular reduction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the sparsity-constrained objective being a meaningful proxy for LDA inference, on the candidate set containing the true topics, and on the α→∞ limit. No new physical or conceptual entities are introduced; the placeholder topic is a technical device for initialization.

free parameters (1)
  • sparsity bound κ = user-chosen (e.g., 4)
    The average number of topics per document is a user-specified constraint that directly controls the solution. It is not fitted from data, but the approximation guarantee is relative to this parameter.
assumptions (3)
  • domain assumption The exchange of argmax and limit as α→∞ is valid for the LDA posterior (Theorem 1).
    The paper assumes that for large α the sparsity term becomes independent of the assignment, so the argmax converges to the likelihood maximizer. This is plausible but requires uniform convergence; the proof in App. B is only a sketch.
  • ad hoc to paper The true latent topics belong to the candidate set Φ generated from coherence functions.
    The greedy guarantee is relative to Φ. If the true topics are not expressible as a single keyword's co-occurrence distribution, the returned solution may be arbitrarily bad. The paper states this as a requirement but gives no method to verify it.
  • standard math The LDA generative model with bag-of-words and Dirichlet priors is the correct model for the data.
    All guarantees are built on the LDA likelihood. This is a standard assumption in topic modeling and is not questioned by the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time." pith.science (2026). https://pith.science/paper/VHFYXMQU

@misc{pith2026250607747,
  author       = {Pith},
  title        = {Pith review of: E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHFYXMQU}},
  note         = {Machine review of arXiv:2506.07747}
}
read the original abstract

In this paper, we provide the first practical algorithms with provable guarantees for the problem of inferring the topics assigned to each document in an LDA topic model. This is the primary inference problem for many applications of topic models in social science, data exploration, and causal inference settings. We obtain this result by showing a novel non-gradient-based, combinatorial approach to estimating topic models. This yields algorithms that converge to near-optimal posterior probability in logarithmic parallel computation time (adaptivity) -- exponentially faster than any known LDA algorithm. We also show that our approach can provide interpretability guarantees such that each learned topic is formally associated with a known keyword. Finally, we show that unlike alternatives, our approach can maintain the independence assumptions necessary to use the learned topic model for downstream causal inference methods that allow researchers to study topics as treatments. In terms of practical performance, our approach consistently returns solutions of higher semantic quality than solutions from state-of-the-art LDA algorithms, neural topic models, and LLM-based topic models across a diverse range of text datasets and evaluation parameters.

Figures

Figures reproduced from arXiv: 2506.07747 by the authors.

Figure 3
Figure 3. Convergence of E-LDA’s objective values ˙f using the (top row) EXP-UMASS topic generator and the (bottom row) CO-OCCURRENCE topic generator. Q. Deferred discussion of unmodified UMASS topic generator We include the UMASS topic generator both to build intuition, and also because its poor post-hoc coherence scores provide a clear illustration that other topic generators are not obtaining high post-hoc coherence merely… view at source ↗
Figure 4
Figure 4. We reprint the main Gibbs baseline from Experiments Set 2 (Fig [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 69 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A spectral algorithm for latent dirichlet allocation

    Anima Anandkumar, Dean P Foster, Daniel J Hsu, Sham M Kakade, and Yi-Kai Liu. A spectral algorithm for latent dirichlet allocation. Advances in neural information processing systems, 25, 2012

  3. [3]

    Learning topic models--going beyond svd

    Sanjeev Arora, Rong Ge, and Ankur Moitra. Learning topic models--going beyond svd. In 2012 IEEE 53rd annual symposium on foundations of computer science, pages 1--10. IEEE, 2012

  4. [4]

    A practical algorithm for topic modeling with provable guarantees

    Sanjeev Arora, Rong Ge, Yonatan Halpern, David Mimno, Ankur Moitra, David Sontag, Yichen Wu, and Michael Zhu. A practical algorithm for topic modeling with provable guarantees. In International conference on machine learning, pages 280--288. PMLR, 2013

  5. [5]

    Computing a nonnegative matrix factorization---provably

    Sanjeev Arora, Rong Ge, Ravi Kannan, and Ankur Moitra. Computing a nonnegative matrix factorization---provably. SIAM Journal on Computing, 45 0 (4): 0 1582--1611, 2016 a

  6. [6]

    Provable algorithms for inference in topic models

    Sanjeev Arora, Rong Ge, Frederic Koehler, Tengyu Ma, and Ankur Moitra. Provable algorithms for inference in topic models. In International Conference on Machine Learning, pages 2859--2867. PMLR, 2016 b

  7. [7]

    Learning topic models provably and efficiently

    Sanjeev Arora, Rong Ge, Yoni Halpern, David Mimno, Ankur Moitra, David Sontag, Yichen Wu, and Michael Zhu. Learning topic models provably and efficiently. volume 61, pages 85--93. ACM, 2018

  8. [8]

    Text algorithms in economics

    Elliott Ash and Stephen Hansen. Text algorithms in economics. Annual Review of Economics, 15 0 (1): 0 659--688, 2023

Show all 87 references
  1. [9]

    Naive feature selection: Sparsity in naive bayes

    Armin Askari, Alexandre d’Aspremont, and Laurent El Ghaoui. Naive feature selection: Sparsity in naive bayes. In International Conference on Artificial Intelligence and Statistics, pages 1813--1822. PMLR, 2020

  2. [10]

    The adaptive complexity of maximizing a submodular function

    Eric Balkanski and Yaron Singer. The adaptive complexity of maximizing a submodular function. In Proceedings of the 50th annual ACM SIGACT symposium on theory of computing, pages 1138--1151, 2018

  3. [11]

    An optimal approximation for submodular maximization under a matroid constraint in the adaptive complexity model

    Eric Balkanski, Aviad Rubinstein, and Yaron Singer. An optimal approximation for submodular maximization under a matroid constraint in the adaptive complexity model. In Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing, pages 66--77, 2019

  4. [12]

    Who leads? who follows? measuring issue attention and agenda setting by legislators and the mass public using social media data

    Pablo Barber \'a , Andreu Casas, Jonathan Nagler, Patrick J Egan, Richard Bonneau, John T Jost, and Joshua A Tucker. Who leads? who follows? measuring issue attention and agenda setting by legislators and the mass public using social media data. American Political Science Revi...

  5. [13]

    Inference for regression with variables generated by ai or machine learning

    Laura Battaglia, Timothy Christensen, Stephen Hansen, and Szymon Sacher. Inference for regression with variables generated by ai or machine learning. arXiv preprint arXiv:2402.15585, 2024

  6. [14]

    Revolving door lobbyists

    Jordi Blanes i Vidal, Mirko Draca, and Christian Fons-Rosen. Revolving door lobbyists. The American Economic Review, 102 0 (7): 0 3731, 2012

  7. [15]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of machine Learning research, 3 0 (Jan): 0 993--1022, 2003

  8. [16]

    The fast algorithm for submodular maximization

    Adam Breuer, Eric Balkanski, and Yaron Singer. The fast algorithm for submodular maximization. In International Conference on Machine Learning, pages 1134--1143. PMLR, 2020

  9. [17]

    Business news and business cycles

    Leland Bybee, Bryan Kelly, Asaf Manela, and Dacheng Xiu. Business news and business cycles. The Journal of Finance, 79 0 (5): 0 3105--3147, 2024

  10. [18]

    From pork to policy: The rise of programmatic campaigning in Japanese elections

    Amy Catalinac. From pork to policy: The rise of programmatic campaigning in Japanese elections. The Journal of Politics, 78 0 (1): 0 1--18, 2016

  11. [19]

    Reading tea leaves: How humans interpret topic models

    Jonathan Chang, Sean Gerrish, Chong Wang, Jordan Boyd-Graber, and David Blei. Reading tea leaves: How humans interpret topic models. Advances in neural information processing systems, 22, 2009

  12. [20]

    Submodular function maximization in parallel via the multilinear relaxation

    Chandra Chekuri and Kent Quanrud. Submodular function maximization in parallel via the multilinear relaxation. SODA, 2019

  13. [21]

    Marginal likelihood from the gibbs output

    Siddhartha Chib. Marginal likelihood from the gibbs output. Journal of the American statistical association, 90 0 (432): 0 1313--1321, 1995

  14. [22]

    Topic modeling in embedding spaces

    Adji B Dieng, Francisco JR Ruiz, and David M Blei. Topic modeling in embedding spaces. Transactions of the Association for Computational Linguistics, 8: 0 439--453, 2020

  15. [23]

    Pitch perfect: Vocal pitch and the emotional intensity of congressional speech

    Bryce J Dietrich, Matthew Hayes, and Diana Z O’brien. Pitch perfect: Vocal pitch and the emotional intensity of congressional speech. American Political Science Review, 113 0 (4): 0 941--962, 2019

  16. [24]

    Persuasion through slanted language: Evidence from the media coverage of immigration

    Milena Djourelova. Persuasion through slanted language: Evidence from the media coverage of immigration. American economic review, 113 0 (3): 0 800--835, 2023

  17. [25]

    Graph-sparse lda: a topic model with structured sparsity

    Finale Doshi-Velez, Byron C Wallace, and Ryan Adams. Graph-sparse lda: a topic model with structured sparsity. In Twenty-Ninth AAAI conference on artificial intelligence, 2015

  18. [26]

    How to make causal inferences using texts

    Naoki Egami, Christian J Fong, Justin Grimmer, Margaret E Roberts, and Brandon M Stewart. How to make causal inferences using texts. Science Advances, 8 0 (42): 0 eabg2652, 2022

  19. [27]

    Submodular maximization with matroid and packing constraints in parallel

    Alina Ene, Huy L Nguyen, and Adrian Vladu. Submodular maximization with matroid and packing constraints in parallel. STOC, 2019

  20. [28]

    Keyword assisted topic models

    Shusei Eshima, Kosuke Imai, and Tomoya Sasaki. Keyword assisted topic models. arXiv preprint arXiv:2004.05964, 2020

  21. [29]

    Submodular maximization with optimal approximation, adaptivity and query complexity

    Matthew Fahrbach, Vahab Mirrokni, and Morteza Zadimoghaddam. Submodular maximization with optimal approximation, adaptivity and query complexity. SODA, 2019

  22. [30]

    Causalm: Causal model explanation through counterfactual language models

    Amir Feder, Nadav Oved, Uri Shalit, and Roi Reichart. Causalm: Causal model explanation through counterfactual language models. Computational Linguistics, 47 0 (2): 0 333--386, 2021

  23. [31]

    Causal inference in natural language processing: Estimation, prediction, interpretation and beyond

    Amir Feder, Katherine A Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E Roberts, et al. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. Transac...

  24. [32]

    Discovery of treatments from text corpora

    Christian Fong and Justin Grimmer. Discovery of treatments from text corpora. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1600--1609, 2016

  25. [33]

    Causal inference with latent treatments

    Christian Fong and Justin Grimmer. Causal inference with latent treatments. American Journal of Political Science, 67 0 (2): 0 374--389, 2023

  26. [34]

    A network approach to topic models

    Martin Gerlach, Tiago P Peixoto, and Eduardo G Altmann. A network approach to topic models. Science advances, 4 0 (7): 0 eaaq1360, 2018

  27. [35]

    Budgeted nonparametric learning from data streams

    Ryan Gomes and Andreas Krause. Budgeted nonparametric learning from data streams. In ICML, volume 1, page 3. Citeseer, 2010

  28. [36]

    Finding scientific topics

    Thomas L Griffiths and Mark Steyvers. Finding scientific topics. Proceedings of the National academy of Sciences, 101 0 (suppl 1): 0 5228--5235, 2004

  29. [37]

    Text as data: A new framework for machine learning and the social sciences

    Justin Grimmer, Margaret E Roberts, and Brandon M Stewart. Text as data: A new framework for machine learning and the social sciences. Princeton University Press, 2022

  30. [38]

    Bertopic: Neural topic modeling with a class-based tf-idf procedure

    Maarten Grootendorst. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794, 2022

  31. [39]

    Near-optimal sensor placements in gaussian processes

    Carlos Guestrin, Andreas Krause, and Ajit Paul Singh. Near-optimal sensor placements in gaussian processes. In Proceedings of the 22nd international conference on Machine learning, pages 265--272, 2005

  32. [40]

    Studying the history of ideas using topic models

    David Hall, Dan Jurafsky, and Christopher D Manning. Studying the history of ideas using topic models. In Proceedings of the 2008 conference on empirical methods in natural language processing, pages 363--371, 2008

  33. [41]

    A causal lens for controllable text generation

    Zhiting Hu and Li Erran Li. A causal lens for controllable text generation. Advances in Neural Information Processing Systems, 34: 0 24941--24955, 2021

  34. [42]

    A memoization framework for scaling submodular optimization to large scale problems

    Rishabh Iyer and Jeffrey Bilmes. A memoization framework for scaling submodular optimization to large scale problems. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 2340--2349. PMLR, 2019

  35. [43]

    Recent advances in text analysis

    Zheng Tracy Ke, Pengsheng Ji, Jiashun Jin, and Wanshan Li. Recent advances in text analysis. Annual Review of Statistics and Its Application, 11 0 (1): 0 annurev, 2024

  36. [44]

    Scalable greedy feature selection via weak submodularity

    Rajiv Khanna, Ethan Elenberg, Alex Dimakis, Sahand Negahban, and Joydeep Ghosh. Scalable greedy feature selection via weak submodularity. In Artificial Intelligence and Statistics, pages 1560--1568. PMLR, 2017

  37. [45]

    Newsweeder: Learning to filter netnews

    Ken Lang. Newsweeder: Learning to filter netnews. In Machine learning proceedings 1995, pages 331--339. Elsevier, 1995

  38. [46]

    Machine reading tea leaves: Automatically evaluating topic coherence and topic model quality

    Jey Han Lau, David Newman, and Timothy Baldwin. Machine reading tea leaves: Automatically evaluating topic coherence and topic model quality. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 530--539, 2014

  39. [47]

    Scaling politically meaningful dimensions using texts and votes

    Benjamin E Lauderdale and Tom S Clark. Scaling politically meaningful dimensions using texts and votes. American Journal of Political Science, 58 0 (3): 0 754--771, 2014

  40. [48]

    Reuters-21578 text categorization test collection

    David Lewis. Reuters-21578 text categorization test collection. Distribution 1.0, AT&T Labs-Research, 1997

  41. [49]

    Pachinko allocation: Dag-structured mixture models of topic correlations

    Wei Li and Andrew McCallum. Pachinko allocation: Dag-structured mixture models of topic correlations. In Proceedings of the 23rd international conference on Machine learning, pages 577--584, 2006

  42. [50]

    A class of submodular functions for document summarization

    Hui Lin and Jeff Bilmes. A class of submodular functions for document summarization. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 510--520, 2011

  43. [51]

    Leveraging sparsity for efficient submodular data summarization

    Erik Lindgren, Shanshan Wu, and Alexandros G Dimakis. Leveraging sparsity for efficient submodular data summarization. Advances in Neural Information Processing Systems, 29, 2016

  44. [52]

    The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery

    Zachary C Lipton. The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue, 16 0 (3): 0 31--57, 2018

  45. [53]

    Entropy-rate clustering: Cluster analysis via maximizing a submodular function subject to a matroid constraint

    Ming-Yu Liu, Oncel Tuzel, Srikumar Ramalingam, and Rama Chellappa. Entropy-rate clustering: Cluster analysis via maximizing a submodular function subject to a matroid constraint. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36 0 (1): 0 99--112, 2013

  46. [54]

    Artificial intelligence act

    Tambiama Madiega. Artificial intelligence act. European Parliament: European Parliamentary Research Service, 2021

  47. [55]

    Local news and national politics

    Gregory J Martin and Joshua McCrain. Local news and national politics. American Political Science Review, 113 0 (2): 0 372--384, 2019

  48. [56]

    Mallet: A machine learning for languagetoolkit

    Andrew Kachites McCallum. Mallet: A machine learning for languagetoolkit. http://mallet. cs. umass. edu

  49. [57]

    Optimizing semantic coherence in topic models

    David Mimno, Hanna Wallach, Edmund Talley, Miriam Leenders, and Andrew McCallum. Optimizing semantic coherence in topic models. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 262--272, 2011

  50. [58]

    Accelerated greedy algorithms for maximizing submodular set functions

    Michel Minoux. Accelerated greedy algorithms for maximizing submodular set functions. In Optimization techniques, pages 234--243. Springer, 1978

  51. [59]

    Distributed submodular maximization: Identifying representative elements in massive data

    Baharan Mirzasoleiman, Amin Karbasi, Rik Sarkar, and Andreas Krause. Distributed submodular maximization: Identifying representative elements in massive data. In Advances in Neural Information Processing Systems, pages 2049--2057, 2013

  52. [60]

    Lazier than lazy greedy

    Baharan Mirzasoleiman, Ashwinkumar Badanidiyuru, Amin Karbasi, Jan Vondr \'a k, and Andreas Krause. Lazier than lazy greedy. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015

  53. [61]

    Fast constrained submodular maximization: Personalized data summarization

    Baharan Mirzasoleiman, Ashwinkumar Badanidiyuru, and Amin Karbasi. Fast constrained submodular maximization: Personalized data summarization. In ICML, pages 1358--1367, 2016

  54. [62]

    Causal inference on outcomes learned from text

    Iman Modarressi, Jann Spiess, and Amar Venugopal. Causal inference on outcomes learned from text. arXiv preprint arXiv:2503.00725, 2025

  55. [63]

    Reading between the lines: Prediction of political violence using newspaper text

    Hannes Mueller and Christopher Rauh. Reading between the lines: Prediction of political violence using newspaper text. American Political Science Review, 112 0 (2): 0 358--375, 2018

  56. [64]

    Best algorithms for approximating the maximum of a submodular set function

    George L Nemhauser and Laurence A Wolsey. Best algorithms for approximating the maximum of a submodular set function. Mathematics of operations research, 3 0 (3): 0 177--188, 1978

  57. [65]

    An analysis of approximations for maximizing submodular set functions—i

    George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. An analysis of approximations for maximizing submodular set functions—i. Mathematical Programming, 14 0 (1): 0 265--294, 1978

  58. [66]

    Approximate bayesian inference with the weighted likelihood bootstrap

    Michael A Newton and Adrian E Raftery. Approximate bayesian inference with the weighted likelihood bootstrap. Journal of the Royal Statistical Society Series B: Statistical Methodology, 56 0 (1): 0 3--26, 1994

  59. [67]

    Concealing corruption: How chinese officials distort upward reporting of online grievances

    Jennifer Pan and Kaiping Chen. Concealing corruption: How chinese officials distort upward reporting of online grievances. The American Political Science Review, 112 0 (3): 0 602--620, 2018

  60. [68]

    Causality

    Judea Pearl. Causality. Cambridge university press, 2009

  61. [69]

    On the unreasonable effectiveness of the greedy algorithm: Greedy adapts to sharpness

    Sebastian Pokutta, Mohit Singh, and Alfredo Torrico. On the unreasonable effectiveness of the greedy algorithm: Greedy adapts to sharpness. In International Conference on Machine Learning, pages 7772--7782. PMLR, 2020

  62. [70]

    Fast parallel algorithms for statistical subset selection problems

    Sharon Qian and Yaron Singer. Fast parallel algorithms for statistical subset selection problems. In Advances in Neural Information Processing Systems, pages 5073--5082, 2019

  63. [71]

    Exploring the space of topic coherence measures

    Michael R \"o der, Andreas Both, and Alexander Hinneburg. Exploring the space of topic coherence measures. In Proceedings of the eighth ACM international conference on Web search and data mining, pages 399--408, 2015

  64. [72]

    Estimating causal effects of treatments in randomized and nonrandomized studies

    Donald B Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology, 66 0 (5): 0 688, 1974

  65. [73]

    Randomization analysis of experimental data: The fisher randomization test comment

    Donald B Rubin. Randomization analysis of experimental data: The fisher randomization test comment. Journal of the American statistical association, 75 0 (371): 0 591--593, 1980

  66. [74]

    Complexity of inference in latent dirichlet allocation

    David Sontag and Dan Roy. Complexity of inference in latent dirichlet allocation. Advances in neural information processing systems, 24, 2011

  67. [75]

    Autoencoding variational inference for topic models

    Akash Srivastava and Charles Sutton. Autoencoding variational inference for topic models. arXiv preprint arXiv:1703.01488, 2017

  68. [76]

    Silvia Terragni, Elisabetta Fersini, Bruno Giovanni Galuzzi, Pietro Tropeano, and Antonio Candelieri. Octis: Comparing and optimizing topic models is simple! In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System ...

  69. [77]

    Get out the vote: determining support or opposition from congressional floor-debate transcripts

    Matt Thomas, Bo Pang, and Lillian Lee. Get out the vote: determining support or opposition from congressional floor-debate transcripts. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 327--335, 2006

  70. [78]

    Rising power on the mind

    Dustin Tingley. Rising power on the mind. International Organization, pages S165--S188, 2017

  71. [79]

    Adapting text embeddings for causal inference

    Victor Veitch, Dhanya Sridhar, and David Blei. Adapting text embeddings for causal inference. In Conference on Uncertainty in Artificial Intelligence, pages 919--928. PMLR, 2020

  72. [80]

    Rethinking lda: Why priors matter

    Hanna Wallach, David Mimno, and Andrew McCallum. Rethinking lda: Why priors matter. Advances in neural information processing systems, 22, 2009 a

  73. [81]

    Evaluation methods for topic models

    Hanna M Wallach, Iain Murray, Ruslan Salakhutdinov, and David Mimno. Evaluation methods for topic models. In Proceedings of the 26th annual international conference on machine learning, pages 1105--1112, 2009 b

  74. [82]

    Structured topic models for language

    Hanna Megan Wallach. Structured topic models for language. PhD thesis, University of Cambridge Cambridge, UK, 2008

  75. [83]

    Effective neural topic modeling with embedding clustering regularization

    Xiaobao Wu, Xinshuai Dong, Thong Thanh Nguyen, and Anh Tuan Luu. Effective neural topic modeling with embedding clustering regularization. In International Conference on Machine Learning, pages 37335--37357. PMLR, 2023

  76. [84]

    A survey on neural topic models: methods, applications, and challenges

    Xiaobao Wu, Thong Nguyen, and Anh Tuan Luu. A survey on neural topic models: methods, applications, and challenges. Artificial Intelligence Review, 57 0 (2): 0 18, 2024 a

  77. [85]

    Fastopic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model

    Xiaobao Wu, Thong Nguyen, Delvin Zhang, William Yang Wang, and Anh Tuan Luu. Fastopic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model. Advances in Neural Information Processing Systems, 37: 0 84447--84481, 2024 b

  78. [86]

    Towards the topmost: A topic modeling system toolkit

    Xiaobao Wu, Fengjun Pan, and Luu Anh Tuan. Towards the topmost: A topic modeling system toolkit. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31--41, 2024 c

  79. [87]

    Topics, concepts, and measurement: A crowdsourced procedure for validating topics as measures

    Luwei Ying, Jacob M Montgomery, and Brandon M Stewart. Topics, concepts, and measurement: A crowdsourced procedure for validating topics as measures. Political Analysis, 30 0 (4): 0 570--589, 2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.