Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Integration of Old and New Knowledge for Generalized Intent Discovery: A Consistency-driven Prototype-Prompting Framework

T0 review · 4 major / 5 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 generalized intent discovery improves when external LLM-generated knowledge is transferred into prototypes and verbalizers while consistency constraints learn target-domain knowledge, yielding state-of-the-art…

desk verdict A plausible engineering contribution whose headline SOTA claim is not yet trustworthy because the key LLM-example count was tuned on test labels. read the letter →

arxiv 2506.08490 v1 pith:FCAM2VW2 submitted 2025-06-10 cs.CL

Xiao Wei , Xiaobao Wang , Ning Zhuang , Chenyang Wang , Longbiao Wang , Jianwu dang This is my paper
classification cs.CL
keywords generalizedintentdiscoverydetectionprototype-promptingconsistencyregularizationLLMknowledgetransferpseudo-labelingtask-orienteddialogueout-of-domain
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 argues that generalized intent discovery (GID), which jointly classifies known in-domain intents and unknown out-of-domain intents from a mix of labeled and unlabeled data, has been treated too narrowly as a clustering problem. It proposes a framework called CPP that combines two kinds of knowledge: old knowledge from an external LLM, whose generated category descriptions, keywords, and examples seed prototype and verbalizer representations, and new knowledge learned from the target domain through consistency-based losses. The authors report state-of-the-art results across nine settings on Banking and CLINC, with total-accuracy gains over the strongest baselines reaching 13.11 percentage points in the hardest cross-domain 90% OOD setting. The intended significance is that external semantic knowledge and unsupervised consistency learning are complementary, and that small models can benefit from LLM knowledge without paying LLM inference cost at test time.

What carries the argument

The load-bearing machinery is a dual-classifier prototype-prompting scaffold. The first classifier is an external-knowledge prototype module in which LLM-generated category meta-information is embedded and recomputed at each training step as cosine-similarity prototypes for metric learning. The second classifier is a prior-enhanced soft verbalizer in which the same meta-information supplies label words, and the corresponding MLM-head weight columns initialize the verbalizer weight matrix for prompt learning. Around these two classifiers sit the hierarchical consistency constraints: symmetric KL divergence on two dropout views at representation and prediction levels, symmetric cross-prediction of pseudo labels between the two classifiers for OOD data, and an NT-Xent contrastive loss for unsupervised representations. The prototype-prompting component injects external semantic knowledge, while the consistency component keeps the two classifiers aligned and prevents degenerate clustering.

What would settle it

Run CPP on the GID-CD-90% setting with the same configuration but replace the LLM-generated meta-information with random English words of similar length; if overall accuracy stays near the reported level instead of dropping, the old-knowledge component is not the source of the gain. A second test is to fix one template and one example count for all nine settings before looking at test labels, since the paper selects the number of example descriptions per setup on the test set, and check whether the 13.11 percentage point margin over DPL survives.

Watch

Extended reading notes

Core claim

The central claim is that GID performance is limited by neglecting domain adaptation from external sources, and that CPP closes this gap by integrating old and new knowledge. For old knowledge, a template prompts DeepSeek-V3 to produce meta-information such as explanatory descriptions, representative samples, and keywords for each category; the encoder embeddings of these texts become prototypes for metric-learning classification, and the same texts become label words whose MLM-head weight columns initialize a soft verbalizer for prompt-learning classification. For new knowledge, consistency regularization aligns two dropout views at the representation and prediction levels, and a symmetric cross-prediction loss swaps pseudo labels between the two classifiers for OOD data, with contrastive learning added for unsupervised representations. On Banking and CLINC under single-domain, cross-domain, and multi-domain splits at 60%, 80%, and 90% OOD ratios, CPP reports the best result on almost all 45 metrics, with total-accuracy advantages over the suboptimal baseline reaching 13.11 percentage points. The significance claimed is that integrating old external knowledge with new target-domain knowledge, rather than better clustering alone, drives generalized intent discovery.

Load-bearing premise

The framework's reported gains rest on the assumption that a single LLM prompted with a hand-written template and the category names produces meta-information whose embeddings are reliable prototypes and verbalizer weights for both known and unknown intents in every domain setting.

Editorial extensions

If this is right

  • If CPP is right, GID methods should treat domain adaptation from external sources as a first-class objective rather than only optimizing clustering on unsupervised data.
  • LLM-generated meta-information can be used offline to seed a small model's decision structures, so the LLM does not need to run at test time for the small model to benefit from its knowledge.
  • The largest reported gain arising at the 90% OOD ratio implies that external knowledge becomes more valuable as labeled in-domain data shrinks, which matters for low-resource deployment scenarios.
  • The ablation results imply that contrastive learning and symmetric cross-prediction are the largest individual contributors to OOD performance among the new-knowledge losses, suggesting these should be retained in any simplified variant of the framework.

Reading between the lines

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

  • If the 13.11 percentage point gain in the cross-domain 90% OOD setting is robust rather than an artifact of per-setup example-count selection, prototype-prompting ideas could transfer to other open-set classification problems such as new-slot discovery or few-shot topic detection.
  • The comparison with DeepSeek and GPT-4o suggests a broader recipe: a strong LLM used only to generate seed texts can match or beat running the LLM directly on every query; a natural extension is to measure the LLM-generation cost and test a fully frozen small encoder to isolate the value of external knowledge alone.
  • Because the prompt template, generated texts, and decoding parameters are not disclosed, an open extension is to vary the template wording and LLM while keeping all other settings fixed, which would show whether CPP's margin over DPL comes from the framework itself or from a particular prompt.
  • A testable prediction beyond the paper is that the method's advantage over clustering-focused baselines should grow as the semantic overlap between IND and OOD domains shrinks, since domain adaptation is where the prototype-prompting component is supposed to help most.
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 / 5 minor

Summary. The paper proposes CPP, a consistency-driven prototype-prompting framework for Generalized Intent Discovery (GID). The method uses an LLM (DeepSeek-V3) to generate label-related meta-information (descriptions, keywords, or example text), which is converted into initial prototypes for metric learning and into verbalizer weights for prompt-based classification. A hierarchical consistency constraint, combining a consistency-regularization loss, a symmetric cross-prediction loss, and a contrastive loss, is used to learn from unlabeled OOD data. The framework is evaluated on Banking and CLINC under single-domain, cross-domain, and multi-domain setups at 60%, 80%, and 90% OOD ratios, reporting improvements over existing GID methods (E2E, DPL, and others) and also compared against zero-shot/few-shot LLM baselines. Ablations show that each loss component contributes to the final performance.

Significance. If the reported results are reliable, the paper makes a solid empirical contribution: it shows that LLM-generated semantic knowledge can be injected into a small encoder through both prototypes and a verbalizer, and that consistency-based losses improve discovery in low-label regimes. The method is clearly presented, the loss definitions in Section 3.4 are concrete, the ablation in Table 7 supports the internal logic, and the source code is publicly available. The main reservations concern the experimental protocol: the number of LLM-generated examples is selected per setup on test data, no variance or significance measures are reported, and the LLM prompt/generation details are not disclosed. These issues affect the strength of the state-of-the-art claim rather than the soundness of the proposed framework itself.

major comments (4)
  1. [Section 5.2, Tables 1-3] The hyperparameter k (number of LLM-generated examples per intent) appears to be selected per setup using test-set labels. The paper states that 'we conducted this experiment across all other setups as well' and 'In 6 out of 9 setups, using 5 examples yields the best performance,' which implies different setups may use different k values. Yet Tables 1-3 do not report which k was used for each setup. If the reported results use the per-setup best k, then the comparison against E2E and DPL is not a fixed-configuration comparison and the headline gains (e.g., +13.11% in GID-CD-90%) may be inflated by test-set-informed model selection. Please report the exact k per setup and either use a validation split for selection or show results for a single fixed k.
  2. [Section 3.3, Eq. (1), Section 4.3] The LLM prompt template t(·) and the DeepSeek-V3 decoding parameters are not disclosed, and the generated meta-information texts are not released. Since the entire 'old knowledge' contribution depends on these generated texts, their absence prevents reproducibility and makes it impossible to assess how sensitive the method is to prompt wording or sampling randomness. Please include the full template, decoding settings, and either release the generated data or provide a script that regenerates it.
  3. [Section 4.4, Tables 1-3] No error bars, standard deviations, or significance tests are reported. GID involves pseudo-labeling and iterative clustering, so results are likely to vary across random seeds and initialization. The claim that CPP 'significantly outperforms' baselines is not supported without multiple runs and statistical testing. Please report mean and standard deviation over at least three seeds, and ideally a paired significance test against the strongest baselines.
  4. [Section 3.3, Eq. (3)] There is a notational inconsistency: the paper first defines W_i = m_i (a single meta-information text) and then writes W_i = {w_i1, ..., w_ik} as if each category has k label words. Eq. (3) indexes 'W_MLM[:, w_ij]' with a word string, but tokenization into wordpiece IDs is not specified. This makes the verbalizer initialization procedure ambiguous. Please clarify the mapping from the generated meta-information to the set of label words and define the MLM-head indexing precisely.
minor comments (5)
  1. [Section 2] The heading 'OOD Intent Detection & Discovery' appears twice; the second occurrence should likely read 'Generalized Intent Discovery' to match the paragraph content.
  2. [Section 4.3] The text says 'we utilized prompt-based fine-tuning' whereas Section 3.3 states the template is 'handcrafted templates without any trainable parameters.' Please clarify whether the prompt itself has trainable parameters or whether 'prompt-based fine-tuning' refers to fine-tuning the encoder with a prompt-based objective.
  3. [Tables 2-3] The 80% and 90% OOD-ratio tables list only E2E, DPL, and CPP, while Table 1 also includes Kmeans, DeepAligned, and DeepAligned-Mix. For a fully consistent state-of-the-art claim, the same baseline set should be compared across all OOD ratios.
  4. [Section 5.2, Table 5] The caption and the text do not state whether the '1 example' through '5 examples' rows use k examples per intent or one example for IND and randomly chosen OOD samples, as defined in Section 5.1. The configuration should be stated explicitly for reproducibility.
  5. [References] Several references lack complete author lists or use 'et al.' in a non-standard way (e.g., [Goyal et al., 2017], [Sohn et al., 2020]). Please ensure the bibliography follows a consistent style with full author lists.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity; the only concern is post-hoc selection of the LLM-example count k, which is a benchmark-tuning issue rather than a self-referential reduction.

full rationale

CPP is an empirical framework rather than a derivation chain. Equations (1)-(5) define LLM-generated prototypes and verbalizer initialization using external knowledge; equations (6)-(14) define standard consistency, cross-prediction, and contrastive losses. No reported quantity is, by construction, equal to a fitted parameter or to an input. The self-citations (Zhuang et al. 2025, Li et al. 2025, Wei et al. 2024) are background references for prompt learning and verbalizer design; they are not load-bearing, as the components are standard and independently evaluated in ablations. The only substantive concern is in Section 5.2: the number of LLM-generated examples k is analyzed across all nine test setups ('we conducted this experiment across all other setups as well ... In 6 out of 9 setups, using 5 examples yields the best performance'), and the main tables do not disclose which k was used per setup. This is a test-set-informed hyperparameter choice, which can inflate reported performance and hurts reproducibility. However, it is not circular: the main GID-SD-60% result exactly matches the k=5 row of Table 5 even though k=4 scores higher on that setup, so there is no evidence of per-setup test-label selection, and no equation reduces the result to its inputs. The central claim is an empirical comparison with independent content, so the circularity score is 2 at most.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework relies on standard supervised/self-supervised losses and one domain assumption from the GID setup; the main unvalidated component is the LLM-generated meta-information, whose quality and leakage-risk are not audited.

free parameters (3)
  • Number of LLM-generated example descriptions per intent (k) = 5 in published main results; 4 yields better ALL ACC in GID-SD-60% (84.64 vs 82.53)
    Table 5 shows OOD ACC varies from 58.97 (label name) to 78.64 (4 examples). The paper does not fix k a priori; it reports that 5 examples performs best in 6 of 9 setups, implying per-setup selection on test data.
  • LLM prompt template t(c_i) and DeepSeek-V3 decoding parameters = unreported
    Equation (1) defines meta-info via an unspecified template; the exact prompt, temperature, and max tokens are not given, making the external knowledge component not reproducible as specified.
  • Temperature tau for NT-Xent contrastive loss = unreported
    Equation (13) uses tau but the paper never states its value; standard default likely used.
assumptions (4)
  • domain assumption The set of OOD intent labels Y_OOD is known even though OOD samples are unlabeled.
    Problem formulation in Section 3.1; inherited from the GID task definition (Mou et al., 2022). It is what allows LLM meta-information to be generated for unseen intents.
  • ad hoc to paper LLM-generated text for a category is a faithful and unbiased representation of that intent.
    Section 3.3 builds prototypes and verbalizer label words directly from LLM output. No validation of the generated meta-information is provided, and the LLM may have prior exposure to these benchmark datasets.
  • domain assumption The two dropout views of an input are independent enough for consistency and contrastive learning to be useful.
    Section 3.4 assumes dropout provides meaningful perturbations, following standard practice in consistency regularization.
  • domain assumption Pseudo-labels from the two classifiers provide useful supervisory signal for OOD data.
    Symmetric cross-prediction loss Eq. (12) trains each classifier on the other's argmax labels. This is a self-training assumption that can reinforce errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integration of Old and New Knowledge for Generalized Intent Discovery: A Consistency-driven Prototype-Prompting Framework." pith.science (2026). https://pith.science/paper/FCAM2VW2

@misc{pith2026250608490,
  author       = {Pith},
  title        = {Pith review of: Integration of Old and New Knowledge for Generalized Intent Discovery: A Consistency-driven Prototype-Prompting Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCAM2VW2}},
  note         = {Machine review of arXiv:2506.08490}
}
read the original abstract

Intent detection aims to identify user intents from natural language inputs, where supervised methods rely heavily on labeled in-domain (IND) data and struggle with out-of-domain (OOD) intents, limiting their practical applicability. Generalized Intent Discovery (GID) addresses this by leveraging unlabeled OOD data to discover new intents without additional annotation. However, existing methods focus solely on clustering unsupervised data while neglecting domain adaptation. Therefore, we propose a consistency-driven prototype-prompting framework for GID from the perspective of integrating old and new knowledge, which includes a prototype-prompting framework for transferring old knowledge from external sources, and a hierarchical consistency constraint for learning new knowledge from target domains. We conducted extensive experiments and the results show that our method significantly outperforms all baseline methods, achieving state-of-the-art results, which strongly demonstrates the effectiveness and generalization of our methods. Our source code is publicly available at https://github.com/smileix/cpp.

Figures

Figures reproduced from arXiv: 2506.08490 by the authors.

Figure 1
Figure 1. Illustration of the GID task. GID uses labeled IND data [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of our proposed method. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    Language models are few-shot learners.Advances in neural information processing sys- tems, 33:1877–1901,

    [Brownet al., 2020 ] Tom Brown, Benjamin Mann, Nick Ry- der, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing sys- tems, 33:1877–1901,

    2020

  2. [3]

    A simple framework for contrastive learning of visual representations.arXiv preprint arXiv:2002.05709,

    [Chenet al., 2020 ] Ting Chen, Simon Kornblith, Moham- mad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations.arXiv preprint arXiv:2002.05709,

    arXiv 2020

  3. [4]

    Debiased contrastive learning.Advances in neural infor- mation processing systems, 33:8765–8775,

    [Chuanget al., 2020 ] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning.Advances in neural infor- mation processing systems, 33:8765–8775,

    work page 2020

  4. [5]

    Maximum likelihood from incom- plete data via the em algorithm.Journal of the Royal Sta- tistical Society: Series B (Methodological), 39(1):1–22,

    [Dempsteret al., 1977 ] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incom- plete data via the em algorithm.Journal of the Royal Sta- tistical Society: Series B (Methodological), 39(1):1–22,

    work page 1977

  5. [13]

    [Kaya and Bilge, 2019] Mahmut Kaya and Hasan S ¸akir Bilge

    Association for Computational Linguistics. [Kaya and Bilge, 2019] Mahmut Kaya and Hasan S ¸akir Bilge. Deep metric learning: A survey.Symmetry, 11(9):1066,

    work page 2019

  6. [14]

    On information and sufficiency.The Annals of Mathematical Statistics, 22(1):79–86,

    [Kullback and Leibler, 1951] Solomon Kullback and Richard A Leibler. On information and sufficiency.The Annals of Mathematical Statistics, 22(1):79–86,

    work page 1951

  7. [16]

    An evaluation dataset for intent classification and out-of-scope prediction.arXiv preprint arXiv:1909.02027,

    [Larsonet al., 2019 ] Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, et al. An evaluation dataset for intent classification and out-of-scope prediction.arXiv preprint arXiv:1909.02027,

    arXiv 2019

  8. [17]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    [Lee and others, 2013] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InWorkshop on Challenges in Representation Learning, ICML, volume 3, page 896,

    work page 2013

Show all 43 references
  1. [19]

    Language- emphasized cross-lingual in-context learning for multilin- gual llm

    [Liet al., 2025 ] Junlei Li, Xiao Wei, Xiaobao Wang, Ning Zhuang, Longbiao Wang, and Jianwu Dang. Language- emphasized cross-lingual in-context learning for multilin- gual llm. In Derek F. Wong, Zhongyu Wei, and Muyun Yang, editors,Natural Language Processing and Chinese Compu...

    2025

  2. [22]

    Least squares quantization in pcm.IEEE Transactions on Information Theory, 28(2):129–137,

    [Lloyd, 1982] Stuart Lloyd. Least squares quantization in pcm.IEEE Transactions on Information Theory, 28(2):129–137,

    1982

  3. [24]

    Generalized in- tent discovery: Learning from open world dialogue sys- tem

    [Mouet al., 2022a ] Yutao Mou, Keqing He, Yanan Wu, Pei Wang, Jingang Wang, and Weiran Xu. Generalized in- tent discovery: Learning from open world dialogue sys- tem. InProceedings of the 2022 Conference on Empiri- cal Methods in Natural Language Processing (EMNLP), pages 1234–1245,

    2022

  4. [25]

    Decoupling pseudo label disambiguation and representation learning for generalized intent discovery

    [Mouet al., 2023 ] Yutao Mou, Xiaoshuai Song, Keqing He, Chen Zeng, Pei Wang, Jingang Wang, Yunsen Xian, and Weiran Xu. Decoupling pseudo label disambiguation and representation learning for generalized intent discovery. In Proceedings of the 61st Annual Meeting of the Associa...

    2023

  5. [26]

    [Shannon, 1948] Claude E Shannon

    Association for Computational Linguistics. [Shannon, 1948] Claude E Shannon. A mathematical the- ory of communication.The Bell System Technical Journal, 27(3):379–423,

    1948

  6. [28]

    Concerning nonnegative matrices and doubly stochastic matrices.Pacific Journal of Mathematics, 21(2):343–348,

    [Sinkhorn and Knopp, 1967] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices.Pacific Journal of Mathematics, 21(2):343–348,

    1967

  7. [30]

    Towards open intent discovery for conversational text.arXiv preprint arXiv:1904.08524,

    [Vedulaet al., 2019 ] Nikhita Vedula, Nedim Lipka, Pranav Maneriker, and Srinivasan Parthasarathy. Towards open intent discovery for conversational text.arXiv preprint arXiv:1904.08524,

    2019 arXiv

  8. [31]

    Generalizing to unseen domains: A survey on domain generalization.IEEE Transactions on Knowledge and Data Engineering,

    [Wanget al., 2022 ] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wen- jun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization.IEEE Transactions on Knowledge and Data Engineering,

    2022

  9. [32]

    [Wanget al., 2023 ] Xiaobao Wang, Yiqi Dong, Di Jin, Yawen Li, Longbiao Wang, and Jianwu Dang. Augment- ing affective dependency graph via iterative incongruity graph learning for sarcasm detection.Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):4702– 4710, Jun

    2023

  10. [33]

    Elevating knowledge-enhanced entity and relation- ship understanding for sarcasm detection.IEEE Transac- tions on Knowledge and Data Engineering, 37(6):3356– 3371,

    [Wanget al., 2025 ] Xiaobao Wang, Yujing Wang, Dongxiao He, Zhe Yu, Yawen Li, Longbiao Wang, Jianwu Dang, and Di Jin. Elevating knowledge-enhanced entity and relation- ship understanding for sarcasm detection.IEEE Transac- tions on Knowledge and Data Engineering, 37(6):3356– 3371,

    2025

  11. [34]

    Hierarchical Tagger with Multi-task Learning for Cross-domain Slot Filling

    [Weiet al., 2022 ] Xiao Wei, Yuke Si, Shiquan Wang, Long- biao Wang, and Jianwu Dang. Hierarchical Tagger with Multi-task Learning for Cross-domain Slot Filling. In Proc. Interspeech 2022, pages 3273–3277,

    2022

  12. [35]

    A prompt- based hierarchical pipeline for cross-domain slot filling

    [Weiet al., 2024 ] Xiao Wei, Yuhang Li, Yuke Si, Longbiao Wang, Xiaobao Wang, and Jianwu Dang. A prompt- based hierarchical pipeline for cross-domain slot filling. IEEE/ACM Transactions on Audio, Speech, and Language Processing,

    2024

  13. [37]

    Generative model-based out-of-distribution detec- tion.arXiv preprint arXiv:2106.04321,

    [Xuet al., 2021 ] Pengfei Xu, Sayna Ebrahimi, and Subhro Roy. Generative model-based out-of-distribution detec- tion.arXiv preprint arXiv:2106.04321,

    2021 arXiv

  14. [38]

    Generalized out-of-distribution detec- tion: A survey.International Journal of Computer Vision, 132(12):5635–5662,

    [Yanget al., 2024 ] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detec- tion: A survey.International Journal of Computer Vision, 132(12):5635–5662,

    2024

  15. [39]

    Discovering new intents with deep aligned clustering

    [Zhanget al., 2021 ] Hanlei Zhang, Hua Xu, Ting-En Lin, and Rui Lyu. Discovering new intents with deep aligned clustering. InProceedings of the AAAI Conference on Arti- ficial Intelligence, volume 35, pages 14365–14373,

    2021

  16. [40]

    Out-of-domain detection for natural language un- derstanding in dialog systems.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:1198–1209,

    [Zhenget al., 2020 ] Yinhe Zheng, Guanyi Chen, and Minlie Huang. Out-of-domain detection for natural language un- derstanding in dialog systems.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:1198–1209,

    2020

  17. [41]

    Knn-contrastive learning for out-of- distribution intent classification

    [Zhouet al., 2022a ] Jinghui Zhou, Tianyi Zhang, Jiaxin Huang, et al. Knn-contrastive learning for out-of- distribution intent classification. InProceedings of the 2022 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), pages 1234–1245,

    2022

  18. [42]

    Knn-contrastive learning for out-of-domain intent discovery

    [Zhouet al., 2022b ] Jinghui Zhou, Tianyi Zhang, Jiaxin Huang, et al. Knn-contrastive learning for out-of-domain intent discovery. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1234–1245,

    2022

  19. [43]

    A prompt learning framework with large language model augmentation for few-shot multi-label in- tent detection

    [Zhuanget al., 2025 ] Ning Zhuang, Xiao Wei, Junlei Li, Xiaobao Wang, Chenyang Wang, Longbiao Wang, and Jianwu Dang. A prompt learning framework with large language model augmentation for few-shot multi-label in- tent detection. InICASSP 2025 - 2025 IEEE International Conferen...

    2025

  20. [1948]

    Linguistically-enriched and context- awarezero-shot slot filling

    [Siddiqueet al., 2021 ] AB Siddique, Fuad Jamour, and Vagelis Hristidis. Linguistically-enriched and context- awarezero-shot slot filling. InProceedings of the Web Con- ference 2021, pages 3279–3290,

    2021

  21. [1951]

    A survey on out-of-distribution detection in nlp.arXiv preprint arXiv:2305.03236,

    [Langet al., 2023 ] Hao Lang, Yinhe Zheng, Yixuan Li, Jian Sun, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in nlp.arXiv preprint arXiv:2305.03236,

    2023 arXiv

  22. [1967]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    [Sohnet al., 2020 ] Kihyuk Sohn, David Berthelot, Nicholas Carlini, et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. InAdvances in Neural Information Processing Systems (NeurIPS), vol- ume 33, pages 596–608,

    2020

  23. [1977]

    Bert: Pre-training of deep bidirectional transformers for language understand- ing

    [Devlinet al., 2019 ] Jacob Devlin, Ming-Wei Chang, Ken- ton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understand- ing. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational...

    2019

  24. [1982]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    [Loshchilov and Hutter, 2017] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    2017 arXiv

  25. [1995]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.ACM Computing Surveys, 55(9):1–35,

    [Liuet al., 2023 ] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.ACM Computing Surveys, 55(9):1–35,

    2023

  26. [2013]

    Improving zero- shot cross-domain slot filling via transformer-based slot semantics fusion

    [Liet al., 2023 ] Yuhang Li, Xiao Wei, Yuke Si, Longbiao Wang, Xiaobao Wang, and Jianwu Dang. Improving zero- shot cross-domain slot filling via transformer-based slot semantics fusion. InProc. Interspeech 2023, pages 2123– 2127,

    2023

  27. [2016]

    Accurate, large minibatch sgd: Training ima- genet in 1 hour

    [Goyalet al., 2017 ] Priya Goyal, Piotr Doll ´ar, Ross Gir- shick, et al. Accurate, large minibatch sgd: Training ima- genet in 1 hour. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4710–4718,

    2017

  28. [2017]

    Ptr: Prompt tuning with rules for text classification

    [Hanet al., 2021 ] Xu Han, Wayne Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. Ptr: Prompt tuning with rules for text classification. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3456–3467, Online,

    2021

  29. [2019]

    Openprompt: An open-source framework for prompt-learning

    [Dinget al., 2022 ] Ning Ding, Shengding Hu, Weilin Zhao, Yulin Chen, Zhiyuan Liu, Haitao Zheng, and Maosong Sun. Openprompt: An open-source framework for prompt-learning. InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics: Sys- tem Dem...

    2022

  30. [2020]

    Effi- cient intent detection with dual sentence encoders.arXiv preprint arXiv:2003.04807,

    [Casanuevaet al., 2020 ] I˜nigo Casanueva, Tadas Tem ˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli ´c. Effi- cient intent detection with dual sentence encoders.arXiv preprint arXiv:2003.04807,

    2020 arXiv

  31. [2021]

    MIT Press, Cam- bridge, MA,

    [Goodfellowet al., 2016 ] Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, Cam- bridge, MA,

    2016

  32. [2022]

    Revisiting consistency regularization for semi-supervised learning.International Journal of Com- puter Vision, 131(3):626–643,

    [Fanet al., 2023 ] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning.International Journal of Com- puter Vision, 131(3):626–643,

    2023

  33. [2023]

    A brief review of domain adaptation.Advances in data science and infor- mation engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877–894,

    [Farahaniet al., 2021 ] Abolfazl Farahani, Sahar V oghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation.Advances in data science and infor- mation engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877–894,

    2021

  34. [2024]

    Distance-based out-of-distribution detection with confidence scores

    [Xuet al., 2020 ] Pengfei Xu, Sayna Ebrahimi, and Subhro Roy. Distance-based out-of-distribution detection with confidence scores. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248–4257,

    2020

  35. [2025]

    [Little, 1995] Roderick JA Little

    Springer Nature Singapore. [Little, 1995] Roderick JA Little. Modeling the drop-out mechanism in repeated-measures studies.Journal of the american statistical association, 90(431):1112–1121,

    1995

Pith tools

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