Pith. sign in

REVIEW 4 major objections 4 minor 37 references

Understanding Cross-Domain Adaptation in Low-Resource Topic Modeling

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

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that cross-domain transfer in low-resource topic modeling is governed by a finite-sample generalization bound, and that a model built around that bound—DALTA—consistently beats existing topic models.

desk verdict A useful empirical transfer method for low-resource topic modeling, but the theoretical bound that is supposed to justify it does not hold as stated and needs major rework. read the letter →

arxiv 2506.07453 v1 pith:MNHI5O2T submitted 2025-06-09 cs.CL

Pritom Saha Akash , Kevin Chen-Chuan Chang This is my paper
classification cs.CL
keywords low-resourcetopicmodelingdomainadaptationgeneralizationboundneuralmodelsadversarialalignmentlatentspacediscrepancycoherencetransferlearning
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

The paper tries to establish when and why a topic model trained on a high-resource source corpus can help infer topics from a low-resource target corpus without drowning the target in irrelevant source content. It derives a finite-sample generalization bound stating that the target-domain reconstruction error is controlled by four measurable quantities: reconstruction quality on both domains, the discrepancy between source and target latent distributions, disagreement between the two domains' optimal decoders, and how strongly the latent space is regularized. If that bound is right, it turns cross-domain transfer in topic modeling from a heuristic into an explicit optimization target. The paper then builds DALTA (Domain-Aligned Latent Topic Adaptation), which shares one encoder across domains, uses separate decoders, and adversarially aligns latent representations, and reports consistent gains in topic coherence, stability, and transferability on low-resource datasets.

What carries the argument

The load-bearing object is Theorem 1's bound, assembled from two ingredients. The first is a standard domain-adaptation error decomposition (Lemma 1 in the paper) that splits target error into source error, an H-divergence term $d_{\mathcal H}(D_S(Z),D_T(Z))$, and a term measuring disagreement between the optimal source and target reconstruction functions $|f_S-f_T|$. The second is a PAC-Bayesian reconstruction guarantee for variational autoencoders (Lemma 2) that converts finite-sample empirical errors into expected errors, contributing the KL term and the sample-complexity term. DALTA is the architecture that converts these bound terms into losses: a shared encoder $q_\phi$ with adversarial domain discrimination minimizes the H-divergence, separate decoders $p_{\theta_S}$ and $p_{\theta_T}$ minimize the reconstruction terms, a consistency loss minimizes the $|f_S-f_T|$ term, and a KL regularizer controls the prior-matching term.

What would settle it

Compute every term on the right-hand side of Theorem 1 from held-out source and target data for a trained model and compare with the measured target reconstruction error; any material violation of the claimed inequality would refute the theorem. A more targeted test is to check whether the decoder's negative log-likelihood loss is symmetric and satisfies the triangle inequality; if it is not, the proof step that invokes the classification-style domain-adaptation lemma for reconstruction is invalid.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is Theorem 1: for any hypothesis $h$ mapping a latent representation to a distribution over the vocabulary, with probability at least $1-\delta$ over random source and target samples, the expected target reconstruction error satisfies $\epsilon_T(h) \le p_T \hat\epsilon_T(h) + p_S \hat\epsilon_S(h) + \frac{1}{\lambda}KL(q\|p) + p_S \, d_{\mathcal H}(D_S(Z),D_T(Z)) + p_S \min\left\{\mathbb{E}_S|f_S-f_T|,\mathbb{E}_T|f_S-f_T|\right\} + O\!\left(\frac{K_\phi K_\theta \Delta + \frac{1}{\lambda}\log\frac{1}{\delta} + \lambda\Delta^2}{n_S+n_T}\right)$. Here $p_S$ and $p_T$ are sample proportions, $\hat\epsilon$ are empirical reconstruction errors, $d_{\mathcal H}$ is an H-divergence measuring how distinguishable the source and target latent distributions are, $f_S$ and $f_T$ are the optimal source and target reconstruction functions, and the last term collects sample-complexity and capacity terms. The paper reads this bound as a three-part recipe: reconstruct both domains well, align the latent spaces, and regularize, and it claims DALTA is the first topic model built to minimize these terms jointly.

Load-bearing premise

The proof's weakest point is that it applies a classical domain-adaptation bound originally proven for classification losses that are symmetric and obey the triangle inequality to a reconstruction loss (negative log-likelihood of decoded words), without showing the reconstruction setting satisfies those conditions. If that transfer of assumptions fails, the theoretical justification for the alignment and consistency losses collapses.

Editorial extensions

If this is right

  • Source quality is part of the target guarantee: the $p_S \hat\epsilon_S(h)$ term means a source model that reconstructs its own domain poorly caps how much transfer can help the target.
  • Latent-space alignment is not an optional bonus but a bound term: the $p_S d_{\mathcal H}$ term ties target error directly to how distinguishable source and target latent samples are.
  • The KL and sample-complexity terms make regularization part of the transfer guarantee, so the bound predicts that unregularized adaptation will pay a measurable penalty in target error.
  • Decoder consistency matters most when domains disagree: the $\min\{\mathbb{E}_S|f_S-f_T|,\mathbb{E}_T|f_S-f_T|\}$ term is exactly what the consistency loss minimizes, giving a precise reason for that design choice.
  • If the bound is tight, the four terms provide a diagnostic that can predict transfer success before full training, a principle the paper's appendix starts to exploit with an alignment score.

Reading between the lines

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

  • A rigorous version of the paper's early-training alignment score could follow directly: estimate H-divergence and target reconstruction loss after a few epochs and rank candidate source domains by the bound's leading terms, making source selection principled rather than heuristic.
  • If Theorem 1 holds generally, the same bound should transfer to other VAE-style generative models with paired domains, such as multilingual or cross-modal representation learning, where the recipe 'reconstruct both, align latents, regularize' would be the governing principle.
  • A controlled synthetic experiment with known topic overlap would test the bound's tightness: target performance should degrade smoothly as the H-divergence term grows, and a sharp threshold instead of a smooth curve would indicate the bound is loose.
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 / 4 minor

Summary. The paper formalizes domain adaptation for low-resource topic modeling, states a finite-sample generalization bound (Theorem 1) bounding target-domain reconstruction error by source and target empirical errors, a KL regularization term, latent H-divergence, and a discrepancy between optimal reconstruction functions, and introduces DALTA, a VAE-based method with a shared encoder, domain-specific decoders, adversarial alignment, a consistency loss, and a KL regularizer. Experiments on four low-resource target corpora with AG News as the source report improvements over eight baselines in topic coherence, diversity, classification, clustering, and qualitative topic interpretability, together with an ablation study and a heuristic source-selection case study.

Significance. The problem is timely and the proposed framework is coherent and empirically promising: the architecture follows the stated bound, and the evaluation covers multiple datasets, metrics, ablations, and a practical source-selection heuristic. The central theoretical result, however, is not established as stated. The proof chain applies a supervised classification bound to a reconstruction loss whose formal assumptions do not hold, and Proposition 1 overstates the relation between the neural adversarial objective and H-divergence. Because the generalization bound is the paper's main claimed contribution and the stated motivation for DALTA, the theoretical foundation needs substantial reworking before the paper's conclusions can be accepted.

major comments (4)
  1. [Appendix A, Lemma 3; §3.1, Theorem 1] The proof of Theorem 1 relies on Lemma 3, which invokes Lemma 1 (Ben-David et al., 2010) as if it applied to the reconstruction setting. Lemma 1 requires a symmetric loss satisfying the triangle inequality and deterministic labeling functions f_S and f_T; the VAE reconstruction loss in Eq. (2) is negative log-likelihood, which is neither symmetric nor triangle, and the reconstruction target is a random document drawn from p(x|z), not a deterministic function of z. Consequently the inequality ε_T(h) ≤ ε_S(h) + d_H + min{...} does not hold for the quantities used in Lemma 3, so the chain of inequalities leading to Theorem 1 is invalid.
  2. [Appendix A, Lemmas 2 and 3] The proof conflates two different objects: Lemma 1's ε_S(h) is a decoder-only error E_z[ℓ(h(z), f_S(z))] with a deterministic f_S(z), whereas Lemma 2 bounds the full VAE objective E_x E_{z∼q(z|x)}[-log p_θ(x|z)]. Lemma 3 substitutes the latter for the former, so the resulting bound does not apply to the model actually trained by Eq. (2).
  3. [§3.1, Theorem 1; Appendix A, proof of Theorem 1] The displayed bound in Theorem 1 contains an extra factor p_S multiplying the min term and an unmatched opening parenthesis, while the proof's final line has only one p_S factor for that term. The union-bound step also needs attention: combining Lemma 2 and Lemma 3, each with confidence 1−δ, gives at least 1−2δ unless δ/2 is used in each lemma; the statement claims 1−δ.
  4. [§3.2, Proposition 1] The proof of Proposition 1 constructs the Bayes-optimal classifier C*(z)=p_S(z)/(p_S(z)+p_T(z)) and identifies H-divergence with total variation. This equality holds only when the discriminator hypothesis class contains all measurable functions; for the neural discriminator class used in Eq. (1), H-divergence is generally strictly less than 2TV. Thus adversarial training does not exactly minimize the fourth term of Theorem 1, and the claim that DALTA is a bound-minimization method is not supported by Proposition 1 as stated.
minor comments (4)
  1. [§3.1, Theorem 1] The notation f_S and f_T ('optimal functions mapping latent representations to reconstructed outputs') is never formally defined; in particular it is unclear whether these are deterministic functions, conditional distributions, or decoder outputs, which makes the min term ambiguous.
  2. [§3.2, Eq. (3)] The consistency loss in Eq. (3) uses ∥p_θS(Z)−p_θT(Z)∥_2, but no argument connects this norm to the absolute difference |f_S−f_T| appearing in Theorem 1; the paper should state the formal relationship between these quantities.
  3. [§4, Tables and Algorithm 1] There are formatting errors that hinder readability: Algorithm 1 has 'end whilereturn' on one line, and Table 3 contains garbled entries such as '0.730.371'. These should be corrected.
  4. [§4, Experiments] No standard deviations, seeds, or multiple-run statistics are reported; given that the target sets are only 1,000 instances, the claim of consistent improvements would be strengthened by variance estimates.

Circularity Check

1 steps flagged · score 2.0 of 10

Localized proof-by-assumed-identity in Proposition 1; the central generalization bound is externally grounded and not circular.

  1. other [Section 3.2 / Appendix A, Proposition 1 and its proof]
    "The H-divergence is related to the total variation distance by: dH(qϕ(XS), qϕ(XT )) = 2TV(q ϕ(XS), qϕ(XT )). Thus, substituting the total variation distance: dH(qϕ(XS), qϕ(XT )) = 2 (1−2ϵ∗C)."

    Proposition 1 is meant to show that the adversarial discriminator error controls the H-divergence term in Theorem 1. The proof, however, does not derive dH = 2TV from the definition of H-divergence; it simply asserts this identity as a premise. All subsequent steps are algebraic substitution: with the Bayes classifier error giving TV = 1 − 2ϵ*C, substituting into the asserted identity yields exactly the proposition's claimed equality. Thus the conclusion is equivalent to the unproved premise, making the proposition self-justifying rather than derived. For a neural-domain-discriminator class, dH = 2TV is not an identity by construction, so the adversarial alignment loss is connected to the bound's fourth term only through this assumed equivalence.

full rationale

The paper's central contribution, Theorem 1, is a finite-sample generalization bound assembled from two named external results: Lemma 1 from Ben-David et al. (2010) and Lemma 2 from Mbacke et al. (2024). The bound is not fitted to data and its terms are not renamed versions of the method's losses; DALTA is then presented as a bound-minimization algorithm, which is the standard theory-to-algorithm flow and not circular. No fitted parameter is relabeled as a prediction, and the only self-citation (Akash and Chang, 2024) appears in related work and is not load-bearing. The one defensible circularity concern is Proposition 1, whose proof assumes the H-divergence/total-variation identity that is the substantive content of the proposition; this makes the adversarial-alignment justification rely on an unproved equivalence rather than on an independent derivation. That flaw is localized and does not make the central generalization bound construction-dependent, so the overall circularity score is low.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The theory rests on two imported lemmas and an unproven equality connecting neural discriminator error to H-divergence. The empirical method adds several hand-chosen loss weights and a sampling schedule. No new physical or mathematical entities are postulated.

free parameters (5)
  • omega_adv
    Weight of the adversarial loss L_adv in Eq. (5). Value is not reported; selected by hand or tuning.
  • omega_cons
    Weight of the consistency loss L_cons in Eq. (5). Value not reported.
  • omega_KL
    Weight of the KL regularizer L_KL in Eq. (5). Value not reported.
  • mu = 0.7 annealed to 0.3
    Domain-weight sampling probability, set to 0.7 initially and decreased to 0.3 during training (Appendix B).
  • lambda_alignment = 0.001
    Hand-chosen weight in the alignment score heuristic of Appendix D.
assumptions (5)
  • domain assumption Ben-David et al. (2010) domain adaptation bound (Lemma 1) applies to reconstruction loss in topic modeling.
    Lemma 3 in Appendix A substitutes reconstruction error into Lemma 1's classification error bound. The paper does not show that the negative log-likelihood reconstruction loss is symmetric and satisfies the triangle inequality required for the bound.
  • standard math Mbacke et al. (2024) PAC-Bayesian reconstruction bound (Lemma 2) holds for DALTA's encoder and decoder.
    Lemma 2 is imported from prior literature without re-derivation and is assumed to bound the reconstruction loss of the specific neural architectures used.
  • standard math Optimal reconstruction functions f_S and f_T exist for both domains and the expectation of |f_S - f_T| is finite.
    The bound and the consistency loss rely on these functions and on the integrability of their difference.
  • domain assumption A shared latent space Z and a standard Gaussian prior p(Z) can represent both domains well enough for transfer.
    The method assumes a single encoder can align the two domains in one latent space while separate decoders handle domain-specific output; this is a modeling assumption not guaranteed when topic spaces diverge.
  • ad hoc to paper The H-divergence between source and target latent distributions equals 2(1 - 2epsilon*_C) for the neural discriminator class.
    Proposition 1's proof uses the Bayes-optimal classifier and total variation distance, but the discriminator is a neural network with a restricted hypothesis class, so the equality is not established for the actual model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding Cross-Domain Adaptation in Low-Resource Topic Modeling." pith.science (2026). https://pith.science/paper/MNHI5O2T

@misc{pith2026250607453,
  author       = {Pith},
  title        = {Pith review of: Understanding Cross-Domain Adaptation in Low-Resource Topic Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MNHI5O2T}},
  note         = {Machine review of arXiv:2506.07453}
}
read the original abstract

Topic modeling plays a vital role in uncovering hidden semantic structures within text corpora, but existing models struggle in low-resource settings where limited target-domain data leads to unstable and incoherent topic inference. We address this challenge by formally introducing domain adaptation for low-resource topic modeling, where a high-resource source domain informs a low-resource target domain without overwhelming it with irrelevant content. We establish a finite-sample generalization bound showing that effective knowledge transfer depends on robust performance in both domains, minimizing latent-space discrepancy, and preventing overfitting to the data. Guided by these insights, we propose DALTA (Domain-Aligned Latent Topic Adaptation), a new framework that employs a shared encoder for domain-invariant features, specialized decoders for domain-specific nuances, and adversarial alignment to selectively transfer relevant information. Experiments on diverse low-resource datasets demonstrate that DALTA consistently outperforms state-of-the-art methods in terms of topic coherence, stability, and transferability.

Figures

Figures reproduced from arXiv: 2506.07453 by the authors.

Figure 1
Figure 1. An illustration of how a high-resource domain [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DALTA Framework ϵT (h) ≤ pT · ϵˆT (h) + pS · ϵˆS(h) + 1 λ KL(q∥p) + pS · (dH(DS(Z), DT (Z)) + pS · min  ES [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 24 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Pritom Saha Akash and Kevin Chang. 2024. Enhancing short-text topic modeling with llm-driven context expansion and prefix-tuned vaes. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 15635--15646

    work page 2024

  4. [4]

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. 2010. A theory of learning from different domains. Machine learning, 79:151--175

    work page 2010

  5. [5]

    Federico Bianchi, Silvia Terragni, and Dirk Hovy. 2020 a . Pre-training is a hot topic: Contextualized document embeddings improve topic coherence. arXiv preprint arXiv:2004.03974

    arXiv 2020

  6. [6]

    Federico Bianchi, Silvia Terragni, Dirk Hovy, Debora Nozza, and Elisabetta Fersini. 2020 b . Cross-lingual contextualized topic models with zero-shot learning. arXiv preprint arXiv:2004.07737

    arXiv 2020

  7. [7]

    David Blei and John Lafferty. 2006 a . Correlated topic models. Advances in neural information processing systems, 18:147

    work page 2006

  8. [8]

    David M Blei and John D Lafferty. 2006 b . Dynamic topic models. In Proceedings of the 23rd international conference on Machine learning, pages 113--120

    work page 2006

Show all 37 references
  1. [9]

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

    2003

  2. [10]

    Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning, 20(3):273--297

    1995

  3. [11]

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

    2020

  4. [12]

    Zhibin Duan, Dongsheng Wang, Bo Chen, Chaojie Wang, Wenchao Chen, Yewen Li, Jie Ren, and Mingyuan Zhou. 2021. Sawtooth factorial topic embeddings guided gamma belief network. In International Conference on Machine Learning, pages 2903--2913. PMLR

    2021

  5. [13]

    Zhibin Duan, Yishi Xu, Jianqiao Sun, Bo Chen, Wenchao Chen, Chaojie Wang, and Mingyuan Zhou. 2022. Bayesian deep embedding topic meta-learner. In International Conference on Machine Learning, pages 5659--5670. PMLR

    2022

  6. [14]

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Fran c ois Laviolette, Mario March, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1--35

    2016

  7. [15]

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

    2022 arXiv

  8. [16]

    Sungwon Han, Mingi Shin, Sungkyu Park, Changwook Jung, and Meeyoung Cha. 2023. Unified neural topic model via contrastive learning and term weighting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1802--1817

    2023

  9. [17]

    Tomoharu Iwata. 2021. Few-shot learning for topic modeling. arXiv preprint arXiv:2104.09011

    2021 arXiv

  10. [18]

    Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  11. [19]

    Bo Li, Yezhen Wang, Shanghang Zhang, Dongsheng Li, Kurt Keutzer, Trevor Darrell, and Han Zhao. 2021. Learning invariant representations and risks for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2021

  12. [20]

    Sokhna Diarra Mbacke, Florence Clerc, and Pascal Germain. 2024. Statistical guarantees for variational autoencoders using pac-bayesian theory. Advances in Neural Information Processing Systems, 36

    2024

  13. [21]

    Jon Mcauliffe and David Blei. 2007. Supervised topic models. Advances in neural information processing systems, 20

    2007

  14. [22]

    Yishu Miao, Lei Yu, and Phil Blunsom. 2016. Neural variational inference for text processing. In International conference on machine learning, pages 1727--1736. PMLR

    2016

  15. [23]

    Feng Nan, Ran Ding, Ramesh Nallapati, and Bing Xiang. 2019. Topic modeling with wasserstein autoencoders. arXiv preprint arXiv:1907.12374

    2019 arXiv

  16. [24]

    Thong Nguyen and Anh Tuan Luu. 2021. Contrastive learning for neural topic model. Advances in neural information processing systems, 34:11974--11986

    2021

  17. [25]

    Chau Minh Pham, Alexander Hoyle, Simeng Sun, Philip Resnik, and Mohit Iyyer. 2023. Topicgpt: A prompt-based topic modeling framework. arXiv preprint arXiv:2311.01449

    2023 arXiv

  18. [26]

    Duy-Tung Pham, Thien Trang Nguyen Vu, Tung Nguyen, Linh Ngo, Duc Nguyen, and Thien Nguyen. 2024. Neuromax: Enhancing neural topic modeling via maximizing mutual information and group topic regularization. In Findings of the Association for Computational Linguistics: EMNLP 2024...

    2024

  19. [27]

    Hinrich Sch \"u tze, Christopher D Manning, and Prabhakar Raghavan. 2008. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge

    2008

  20. [28]

    Suzanna Sia and Kevin Duh. 2021. Adaptive mixed component lda for low resource topic modeling. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2451--2469

    2021

  21. [29]

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

    2017 arXiv

  22. [30]

    Raymond E Wright. 1995. Logistic regression

    1995

  23. [31]

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

    2023

  24. [32]

    Xiaobao Wu, Chunping Li, Yan Zhu, and Yishu Miao. 2020. Short text topic modeling with topic distribution quantization and negative sampling decoder. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1772--1782

    2020

  25. [33]

    Xiaobao Wu, Thong Thanh Nguyen, Delvin Ce Zhang, William Yang Wang, and Anh Tuan Luu. 2024. Fastopic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  26. [34]

    Weijie Xu, Wenxiang Hu, Fanyou Wu, and Srinivasan Sengamedu. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.606 D e T i ME : Diffusion-enhanced topic modeling using encoder-decoder based LLM . In Findings of the Association for Computational Linguistics: EMNLP 2023, pag...

    2023 doi

  27. [35]

    Yishi Xu, Jianqiao Sun, Yudi Su, Xinyang Liu, Zhibin Duan, Bo Chen, and Mingyuan Zhou. 2024. Context-guided embedding adaptation for effective topic modeling in low-resource regimes. Advances in Neural Information Processing Systems, 36

    2024

  28. [36]

    Han Zhao, Remi Tachet Des Combes, Kun Zhang, and Geoffrey Gordon. 2019. On learning invariant representations for domain adaptation. In International conference on machine learning, pages 7523--7532. PMLR

    2019

  29. [37]

    He Zhao, Dinh Phung, Viet Huynh, Trung Le, and Wray Buntine. 2020. Neural topic model via optimal transport. arXiv preprint arXiv:2008.13537

    2020 arXiv

Pith tools

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