Pith. sign in

REVIEW 3 major objections 4 minor 52 references

Entriever: Energy-based Retriever for Knowledge-Grounded Dialog Systems

T0 review · 3 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 An energy-based retriever scores whole candidate knowledge sets jointly and outperforms independent-scoring baselines in knowledge-grounded dialog retrieval.

desk verdict Sensible energy-based joint retriever for task-oriented dialog, but the missing set-level cross-encoder baseline leaves the core mechanistic claim untested. read the letter →

arxiv 2506.00585 v1 pith:HR22ZBWH submitted 2025-05-31 cs.CL

Yucheng Cai , Ke Li , Yi Huang , Junlan Feng , Zhijian Ou This is my paper
classification cs.CL
keywords energy-basedretrieverknowledge-groundeddialogset-levelretrievalsemi-supervisedsystemslanguagemodelcross-encoderrerankingtask-orientedscoring
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

This work proposes Entriever, an energy-based retriever that scores an entire candidate retrieval result, a whole set of knowledge pieces, rather than scoring each piece independently. Traditional retrievers factor the retrieval probability as a product over knowledge pieces, so they miss dependencies such as two pieces that are individually likely but jointly inconsistent. Entriever assigns the set a single score via an energy function, trained by maximum likelihood with importance or Metropolis sampling, and reports improved retrieval on four task-oriented dialog datasets. It also uses the unnormalized energy score to weight pseudo-labeled knowledge in semi-supervised dialog training without accessing the full knowledge base. If the results hold, retrieval in knowledge-grounded dialog should be treated as a collective decision, not a ranking of isolated facts.

What carries the argument

The load-bearing mechanism is an energy function $U_{\theta}$ over the concatenated dialog context and candidate knowledge-set sequence, converted into a retrieval probability through $\exp(-U_{\theta})/Z_{\theta}$. In the residual variant, the energy function exponentiates the independent cross-encoder distribution, so the model learns only the difference between the independent baseline and the target set-level distribution. During inference, the top-$K$ candidate sets are produced by the independent retriever with the Viterbi algorithm and then rescored by Entriever as wholes; in semi-supervised training, the unnormalized energy score supplies the retrieval probability inside the importance weight for pseudo labels, making the full knowledge base unnecessary.

What would settle it

Compute, on each dataset, the rank of every oracle knowledge set under the independent cross-encoder; if a non-negligible fraction of oracle sets fall outside the top 16, those turns are impossible for Entriever to retrieve. A direct oracle-coverage measurement at $K=16$ on all four datasets would separate reordering the proposer's candidates from recovering sets the proposer missed.

Watch

Extended reading notes

Core claim

The central claim is that modeling the probability of a candidate set jointly, instead of as a product of independent piece probabilities, improves both retrieval and downstream dialog. Entriever defines $p(\xi_t \mid c_t,u_t) \propto \exp(-U_{\theta}(c_t,u_t,\xi_t))$, where $\xi_t$ is a full collection of knowledge pieces and $U_{\theta}$ is a bidirectional-transformer energy function over the concatenated sequence. Because the energy is unnormalized, the score can be compared across candidate sets and used in importance weights without enumerating the knowledge base. The paper finds that a residual form, which starts from the independent cross-encoder distribution and learns only an exponential correction, trains more stably and gives the best retrieval results, and that in semi-supervised experiments Entriever improves the combined success-and-fluency score of the dialog system on the MobileCS dataset.

Load-bearing premise

Entriever can only choose among the top-$K$ candidate sets that the independent retriever proposes, so its success depends on $K$ being large enough that every oracle knowledge set appears in that list.

Editorial extensions

If this is right

  • Retrieval becomes a set-level decision: a candidate set can be rejected for internal inconsistency even when every member scores well individually.
  • Semi-supervised systems can compute an importance weight for pseudo-labeled knowledge without enumerating the knowledge base, because the unnormalized energy score cancels in the acceptance ratio.
  • The residual form lets an existing independent retriever serve as reference distribution, so Entriever only needs to learn the correction on top of it.
  • Using Entriever for inference costs more than a cross-encoder, but the overhead is acceptable when the knowledge base is small, which is the vertical-domain setting the paper targets.

Reading between the lines

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

  • A natural extension is to replace the independent proposer with a diversity-aware proposer; since Entriever's ceiling is the proposer's top-K list, the comparison would reveal how much joint scoring adds beyond candidate coverage.
  • The same set-wise energy scoring could transfer to passage-level retrieval for grounded generation, where retrieved passages should be mutually consistent as a block rather than independently relevant.
  • Entriever could also be trained on unlabeled data alone once the energy score supplies its own retrieval signal; the paper leaves retriever training over unlabeled data as future work.
  • A lighter test would be to measure how often Entriever's winning set is a set that the independent retriever would not rank first, isolating the contribution of dependency modeling from base relevance ranking.
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

3 major / 4 minor

Summary. The paper introduces Entriever, an energy-based retriever for knowledge-grounded dialog systems. Instead of scoring each knowledge piece independently as in dual-encoder or cross-encoder retrievers, Entriever assigns a score to the whole candidate set of retrieved knowledge pieces via an energy function U(c_t, u_t, ξ_t), with a residual form that exponentiates a cross-encoder reference distribution. The model is trained by maximum likelihood estimation using importance sampling (IS) or Metropolis independence sampling (MIS), and at test time Entriever reranks the top-K candidate sets proposed by an independent cross-encoder via Viterbi search. Experiments on MobileCS, CamRest, In-Car, and Woz2.1 report retrieval improvements in Joint-acc, Inform, and F1, and experiments on MobileCS report end-to-end improvements for semi-supervised dialog systems when Entriever is used to score pseudo knowledge labels in JSA training. The paper also includes ablations on residual versus non-residual forms, on the number of proposed candidates K, and on computational overhead.

Significance. If the claims hold, this is a useful contribution: it is the first energy-based retriever for dialog knowledge retrieval, it demonstrates that modelling candidate retrieval results jointly can improve retrieval accuracy, and it shows a practical benefit in semi-supervised dialog systems where the full KB is unavailable. The paper is generally clearly written, the code is open-sourced, and the experiments cover four datasets with both retrieval and end-to-end evaluations. However, the central mechanistic claim — that the gains come from the energy-based formulation capturing inter-knowledge dependencies — is not yet established, because no non-energy jointly-scoring baseline is compared, and the oracle coverage of the top-K proposal step is not reported. In addition, there is a sign inconsistency in the training-gradient and importance-weight equations that must be resolved for the method to be reproducible as written.

major comments (3)
  1. [§4.3.1, Table 1] The main claim that Entriever captures inter-knowledge dependencies through its energy function is underdetermined because the only baselines are independent-scoring dual-encoder and cross-encoder retrievers. A set-level neural reranker that scores the same concatenated input (c_t, u_t, ξ_t) with ordinary binary cross-entropy, trained on the same top-K Viterbi proposals and the same labels, is never run. If such a control matches Entriever's Joint-acc, Inform, and F1 numbers, then the reported improvements are attributable to the joint-set representation and the reranking stage, not to the energy-based formulation or MLE training. Please add this baseline, or explicitly soften the mechanistic interpretation of the results.
  2. [§4.3.1, Algorithm 1, Table 5] Entriever scores only the K candidate subsets proposed by the independent cross-encoder, so any knowledge piece outside the Viterbi top-K is never considered. The paper does not report the oracle recall of the top-K proposals, i.e., the fraction of turns where the ground-truth ξ_t is contained in the K candidates. Without this number, the observed gains over the cross-encoder are upper-bounded by the proposal quality, and the paper cannot show that Entriever corrects inter-knowledge dependencies that the independent retriever systematically misses. Table 5 only reports final retrieval F1 as K varies; please report the oracle coverage of the top-K proposals for each dataset and discuss how the chosen K=16 relates to that ceiling.
  3. [§4.2, Eqs. (8), (12), (13)] There is a sign inconsistency in the training equations. With p_ret ∝ exp(−U), the negative log-likelihood J = −log p_ret has gradient ∇U − E[∇U], but Eq. (12) reports −∇U + E[∇U]. Similarly, since the residual model is p_ret ∝ p_ref exp(−U) and the proposal is q = p_ref, the importance weight p_ret/q should be proportional to exp(−U), not exp(U) as stated in Eq. (13). Because the IS and MIS training procedures use these weights to estimate the gradient, the authors should correct the sign convention and confirm that the released code implements the intended objective. If the equations are meant to use a different sign convention for U, that convention must be stated explicitly.
minor comments (4)
  1. [Table 1] The retrieval results are reported without variance or significance tests; given that some differences are modest (e.g., F1 0.824 vs 0.816 on CamRest), please report multiple-run means with standard deviations or paired significance tests for the retrieval metrics.
  2. [Table 2] The caption contains a typo: 'colomun' should be 'column'. The two p-value columns are also ambiguous; please label explicitly which comparison each column tests.
  3. [§5.2] The text says the sample size in IS and the number of Markov steps in MIS are both set to 12, but it does not specify the number of gradient-estimation samples per turn or the burn-in policy for the Markov chain; please add these training details for reproducibility.
  4. [§4.3.2] There is a missing space in 'forξ t generated from the inference model'; the sentence should read 'for ξ_t generated from the inference model'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Entriever's gains are empirical, and its use of the cross-encoder as proposal/reference does not make the comparison self-fulfilling.

full rationale

The paper's central claims are empirical rather than definitional. Entriever's energy function is trained by maximum likelihood on human-annotated knowledge labels (Eq. 11), not on outputs of the cross-encoder baseline. The cross-encoder enters the method in two auxiliary roles: as the proposal distribution for importance sampling/MCMC in training (Section 4.2) and as the reference distribution for the residual form (Eq. 10). In both roles it is fixed and does not absorb Entriever's parameters; the residual energy is learned to fit the difference between the target and the reference. Thus the final retrieval score is not equal to the baseline score by construction, and outperformance (Tables 1, 2, 4) is an empirical result. The top-K proposal restriction (Section 4.3.1, Table 5) is a genuine limitation on oracle coverage, but it is not a circular reduction: Entriever can rerank within the proposed set and is not guaranteed to match the baseline. Self-citations (e.g., Cai et al. 2023 as baseline and setup, Ou et al. 2024 for the gradient identity) are either baseline implementations or standard mathematical identities, not load-bearing premises imported to force a conclusion. The absence of a non-energy jointly-scoring reranker weakens the mechanistic attribution of gains to the energy formulation, but that is an experimental underdetermination, not circularity. No quoted equation reduces to another by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; it is a new scoring function over already-known knowledge items.

free parameters (3)
  • sample_size = 12
    Number of samples in IS and Markov steps in MIS for estimating the gradient in Eq. (12), set to 12.
  • beam_width_K = 16
    Number of top candidate retrieval sets proposed by the cross-encoder for Entriever to score at test time; chosen based on ablation in Table 5.
  • learning_rate = not reported
    Training hyperparameters for BERT/GPT2 are not specified in the paper.
assumptions (2)
  • domain assumption The gradient estimator using IS or MIS with the fixed proposal q (the independent cross-encoder distribution) is unbiased and has acceptable variance for the sample size used.
    Section 4.2: the expectation in Eq. (12) is approximated with 12 samples from q. Standard theory requires q to be close to the target for low variance, but no diagnostic is given.
  • domain assumption The top-K candidate set from the independent cross-encoder (Eq. 2) contains the oracle joint retrieval result with high probability.
    Section 4.3.1 and Table 5: Viterbi top-K from independent probabilities is the only source of candidates for Entriever; if the oracle set contains an individually low-scored knowledge piece, it cannot be recovered. The paper's K=16 choice is heuristic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entriever: Energy-based Retriever for Knowledge-Grounded Dialog Systems." pith.science (2026). https://pith.science/paper/HR22ZBWH

@misc{pith2026250600585,
  author       = {Pith},
  title        = {Pith review of: Entriever: Energy-based Retriever for Knowledge-Grounded Dialog Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HR22ZBWH}},
  note         = {Machine review of arXiv:2506.00585}
}
read the original abstract

A retriever, which retrieves relevant knowledge pieces from a knowledge base given a context, is an important component in many natural language processing (NLP) tasks. Retrievers have been introduced in knowledge-grounded dialog systems to improve knowledge acquisition. In knowledge-grounded dialog systems, when conditioning on a given context, there may be multiple relevant and correlated knowledge pieces. However, knowledge pieces are usually assumed to be conditionally independent in current retriever models. To address this issue, we propose Entriever, an energy-based retriever. Entriever directly models the candidate retrieval results as a whole instead of modeling the knowledge pieces separately, with the relevance score defined by an energy function. We explore various architectures of energy functions and different training methods for Entriever, and show that Entriever substantially outperforms the strong cross-encoder baseline in knowledge retrieval tasks. Furthermore, we show that in semi-supervised training of knowledge-grounded dialog systems, Entriever enables effective scoring of retrieved knowledge pieces and significantly improves end-to-end performance of dialog systems.

Figures

Figures reproduced from arXiv: 2506.00585 by the authors.

Figure 1
Figure 1. An illustration of the difference between (a) traditional retriever and (b) Entriever in the retrieval task for [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The models in semi-supervised training proce [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The architecture of retrieval models: (a) the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages

  1. [1]

    URL: " 'urlintro :=

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

  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]

    Anton Bakhtin, Yuntian Deng, Sam Gross, Myle Ott, Marc'Aurelio Ranzato, and Arthur Szlam. 2021. Residual energy-based models for text. The Journal of Machine Learning Research, 22(1):1840--1880

    work page 2021

  4. [4]

    Yucheng Cai, Si Chen, Yuxuan Wu, Yi Huang, Junlan Feng, and Zhijian Ou. 2024. The 2nd futuredial challenge: Dialog systems with retrieval augmented generation (futuredial-rag). In 2024 IEEE Spoken Language Technology Workshop (SLT), pages 1091--1098. IEEE

    work page 2024

  5. [5]

    Yucheng Cai, Hong Liu, Zhijian Ou, Yi Huang, and Junlan Feng. 2022. Advancing semi-supervised task oriented dialog systems by JSA learning of discrete latent variable models. In Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 456--467

    work page 2022

  6. [6]

    Yucheng Cai, Hong Liu, Zhijian Ou, Yi Huang, and Junlan Feng. 2023. Knowledge-retrieval task-oriented dialog systems with semi-supervision. In INTERSPEECH

    work page 2023

  7. [7]

    Kevin Clark, Minh-Thang Luong, Quoc Le, and Christopher D. Manning. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.20 Pre-training transformers as energy-based cloze models . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 285--294

    work page doi:10.18653/v1/2020.emnlp-main.20 2020

  8. [8]

    Jingcheng Deng, Liang Pang, Huawei Shen, and Xueqi Cheng. 2023. Regavae: A retrieval-augmented gaussian mixture variational auto-encoder for language modeling. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2500--2510

    work page 2023

Show all 52 references
  1. [9]

    Yuntian Deng, Anton Bakhtin, Myle Ott, Arthur Szlam, and Marc'Aurelio Ranzato. 2020. Residual energy-based models for text generation. In International Conference on Learning Representations

    2020

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language T...

    2019

  3. [11]

    Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations

    2018

  4. [12]

    Zeyuan Ding, Zhihao Yang, Ling Luo, Yuanyuan Sun, and Hongfei Lin. 2024. From retrieval to generation: A simple and unified generative model for end-to-end task-oriented dialogue. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17907--17914

    2024

  5. [13]

    Mihail Eric, Rahul Goel, Shachi Paul, Abhishek Sethi, Sanchit Agarwal, Shuyang Gao, Adarsh Kumar, Anuj Kumar Goyal, Peter Ku, and Dilek Hakkani-T \"u r. 2020. Multiwoz 2.1: A consolidated multi-domain dialogue dataset with state corrections and state tracking baselines. In LREC

    2020

  6. [14]

    Mihail Eric, Lakshmi Krishnan, Francois Charette, and Christopher D Manning. 2017. Key-value retrieval networks for task-oriented dialogue. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 37--49

    2017

  7. [15]

    G David Forney. 1973. The viterbi algorithm. Proceedings of the IEEE, 61(3):268--278

    1973

  8. [16]

    Silin Gao, Zhijian Ou, Wei Yang, and Huifang Xu. 2020. Integrating discrete and neural features via mixed-feature trans-dimensional random field language models. In International Conference on Acoustics, Speech, and Signal Processing (ICASSP)

    2020

  9. [17]

    Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2022

  10. [18]

    Michael Gutmann and Aapo Hyv \"a rinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297--304

    2010

  11. [19]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, pages 3929--3938

    2020

  12. [20]

    Tianxing He, Bryan McCann, Caiming Xiong, and Ehsan Hosseini-Asl. 2021. https://doi.org/10.18653/v1/2021.eacl-main.151 Joint energy-based model training for better calibrated natural language understanding models . In Proceedings of the 16th Conference of the European Chapter ...

    2021 doi

  13. [21]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022 a . Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research

    2022

  14. [22]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022 b . Few-shot learning with retrieval augmented language models. arXiv e-prints, pages arXiv--2208

    2022

  15. [23]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020 a . Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMN...

    2020

  16. [24]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020 b . Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMN...

    2020

  17. [25]

    Ekaterina Khramtsova, Shengyao Zhuang, Mahsa Baktashmotlagh, and Guido Zuccon. 2024. Leveraging llms for unsupervised dense retriever ranking. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1307--1317

    2024

  18. [26]

    Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In International Conference on Learning Representations ( ICLR )

    2020

  19. [27]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  20. [28]

    Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2020. Zero-resource knowledge-grounded dialogue generation. Advances in Neural Information Processing Systems, 33:8475--8485

    2020

  21. [29]

    Yu Li, Baolin Peng, Yelong Shen, Yi Mao, Lars Liden, Zhou Yu, and Jianfeng Gao. 2022. Knowledge-grounded dialogue generation with a unified knowledge representation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Lingui...

    2022

  22. [30]

    Hong Liu, Zhaobiao Lv, Zhijian Ou, Wenbo Zhao, and Qing Xiao. 2023. Exploring energy-based language models with different architectures and training methods for speech recognition. arXiv preprint arXiv:2305.12676

    2023 arXiv

  23. [31]

    Hong Liu, Hao Peng, Zhijian Ou, Juanzi Li, Yi Huang, and Junlan Feng. 2022. Information extraction and human-robot dialogue towards real-life tasks: A baseline study with the mobilecs dataset. In EMNLP 2022 SereTOD Workshop

    2022

  24. [32]

    Jun S Liu. 2001. Monte Carlo strategies in scientific computing. Springer

    2001

  25. [33]

    Weichen Lu, Yifei Wang, Weizhen Zhao, Buxian Chen, Xiaojie Chen, Jinsong Pan, Wentao Liang, and Yongquan Lai. 2022. http://seretod.org/Challenge.html Team passion at seretod-emnlp 2022: End-to-end task-oriented dialog system with improved prompting scheme

    2022

  26. [34]

    Zhijian Ou, Junlan Feng, Juanzi Li, Yakun Li, Hong Liu, Hao Peng, Yi Huang, and Jiangjiang Zhao. 2022. A challenge on semi-supervised and reinforced task-oriented dialog systems. arXiv preprint arXiv:2207.02657

    2022 arXiv

  27. [35]

    Zhijian Ou and Yunfu Song. 2020. Joint stochastic approximation and its application to learning discrete latent variable models. In Conference on Uncertainty in Artificial Intelligence, pages 929--938. PMLR

    2020

  28. [36]

    Zhijian Ou et al. 2024. Energy-based models with applications to speech and language processing. Foundations and Trends in Signal Processing , 18(1-2):1--199

    2024

  29. [37]

    Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2021. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In International Conference on Learning Representations

    2021

  30. [38]

    Tetiana Parshakova, Jean-Marc Andreoli, and Marc Dymetman. 2019. https://doi.org/10.18653/v1/K19-1084 Global autoregressive models for data-efficient sequence learning . In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 900--909

    2019 doi

  31. [39]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9

    2019

  32. [40]

    Stephen Robertson, Hugo Zaragoza, and Michael Taylor. 2004. Simple bm25 extension to multiple weighted fields. In Proceedings of the thirteenth ACM international conference on Information and knowledge management, pages 42--49

    2004

  33. [41]

    Tao Shen, Guodong Long, Xiubo Geng, Chongyang Tao, Yibin Lei, Tianyi Zhou, Michael Blumenstein, and Daxin Jiang. 2024. Retrieval-augmented retrieval: Large language models are strong zero-shot retriever. In Findings of the Association for Computational Linguistics ACL 2024, pa...

    2024

  34. [42]

    Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya Chen, Kushal Arora, Joshua Lane, et al. 2022. Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage. arXiv preprint arXiv:2208.03188

    2022 arXiv

  35. [43]

    Bin Wang and Zhijian Ou. 2017. Language modeling with neural trans-dimensional random fields. In IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 294--300. IEEE

    2017

  36. [44]

    Bin Wang and Zhijian Ou. 2018 a . https://doi.org/10.1109/SLT.2018.8639591 Improved training of neural trans-dimensional random field language models with dynamic noise-contrastive estimation . In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 70--76

    2018

  37. [45]

    Bin Wang and Zhijian Ou. 2018 b . Learning neural trans-dimensional random field language models with noise-contrastive estimation. In International Conference on Acoustics, Speech, and Signal Processing (ICASSP)

    2018

  38. [46]

    Bin Wang, Zhijian Ou, and Zhiqiang Tan. 2015. https://doi.org/10.3115/v1/P15-1076 Trans-dimensional random fields for language modeling . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Na...

    2015 doi

  39. [47]

    Bin Wang, Zhijian Ou, and Zhiqiang Tan. 2017. Learning trans-dimensional random fields with applications to language modeling. IEEE transactions on pattern analysis and machine intelligence, 40(4):876--890

    2017

  40. [48]

    Tsung-Hsien Wen, David Vandyke, Nikola Mrk s i \'c , Milica Gasic, Lina M Rojas Barahona, Pei-Hao Su, Stefan Ultes, and Steve Young. 2017. A network-based end-to-end trainable task-oriented dialogue system. In Proceedings of the 15th Conference of the European Chapter of the A...

    2017

  41. [49]

    Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. Retromae: Pre-training retrieval-oriented language models via masked auto-encoder. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 538--548

    2022

  42. [50]

    Zhitong Yang, Xing Ma, Anqi Liu, and Zheyu Zhang. 2022. Discovering customer-service dialog system with semi-supervised learning and coarse-to-fine intent detection. In EMNLP 2022 SereTOD Workshop

    2022

  43. [51]

    Weihao Zeng, Keqing He, Zechen Wang, Dayuan Fu, Guanting Dong, Ruotong Geng, Pei Wang, Jingang Wang, Chaobo Sun, Wei Wu, et al. 2022. Semi-supervised knowledge-grounded pre-training for task-oriented dialog systems. In EMNLP 2022 SereTOD Workshop

    2022

  44. [52]

    Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledge-grounded dialogue generation with pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3377--3390

    2020

Pith tools

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