Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Minimally Supervised Hierarchical Domain Intent Learning for CRS

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper shows that a stable hierarchical intent structure for a food-domain conversational recommender needs only about 20,400 utterances, roughly 46% of the available 44,112 questions.

desk verdict A useful algorithmic combination undercut by a self-contradictory evaluation: the paper's own NMI/ARI numbers fail its stated 0.85 stability criterion. read the letter →

arxiv 2505.02209 v1 pith:P3EC42AB submitted 2025-05-04 cs.IR cs.HC

Safikureshi Mondal , Subhasis Dasgupta , Amarnath Gupta This is my paper · ORCID
classification cs.IRcs.HC
keywords hierarchicalintentclusteringconversationalrecommendationsystemneuralattentionstratifiedsamplingdomaindiscoveryclusterstabilityminimaldatacoverage
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 is trying to establish that a domain intent hierarchy for a conversational recommendation system can be built from a surprisingly small fraction of the available user questions. Using a neural attention-based hierarchical clustering algorithm on 44,112 food-business questions, the authors report that the cluster structure stabilizes after approximately 20,400 utterances, meaning additional data no longer changes leaf clusters, hierarchy size, or centroid positions. If true, this lowers the data and labeling cost of building and maintaining dynamic intent structures, because a stratified sample across business categories can stand in for the full corpus. The intended payoff is a self-learning domain model that can absorb new questions without frequent retraining.

What carries the argument

The load-bearing mechanism is the hierarchical clustering procedure the paper calls Efficient Hierarchical Clustering via Neural Attention and Adaptive Entropy Merging. It refines transformer embeddings with a neural attention layer, initializes clusters with ANN-accelerated k-means, iteratively merges clusters using attention-weighted entropy with adaptive threshold annealing, and selects the top-3 centroid-nearest utterances as prototypes for each leaf cluster. This procedure, combined with stratified sampling across the business categories, is what lets the paper detect the point at which additional utterances no longer change the intent structure.

What would settle it

Re-run the hierarchy on all 44,112 questions with the same embeddings, attention settings, and merging parameters, then compare leaf intents and validation-set assignments against the 20,400-utterance result; if validation questions assign to substantially different leaves, for example below 90% agreement in leaf membership, the claim that additional questions are redundant is refuted.

Watch

Extended reading notes

Core claim

The central claim is that hierarchical domain intent structures can be effectively constructed with a minimal number of utterances, approximately 20,400, which is about 46% of the 44,112-question dataset, when training samples are drawn by stratified sampling across 255 food-business categories. The paper reports that clustering stabilizes at that size: the leaf cluster count derivative falls below 0.001, the total hierarchy node count stabilizes at 10,200 utterances, centroid movement drops below 5%, and NMI and ARI stability scores level off near 0.80 and 0.65. Expert review of 500 clusters and their prototypes reports 92.5% accuracy, and the paper concludes that the hierarchy from the 20,400-utterance sample matches the structure built from the full dataset, leaving the remaining 54% of questions redundant for intent coverage.

Load-bearing premise

The load-bearing premise is that the algorithm's own stability signals, the tree stopping its growth, centroids moving less than 5%, and cluster-comparison scores flattening, genuinely mean the intent space is covered, even though no external labels or downstream task verify that the discovered intents are the ones users care about.

Editorial extensions

If this is right

  • A system can reach adequate domain coverage with roughly 46% of the question set, cutting data collection and manual labeling by more than half.
  • The optimal dataset size can be detected during incremental sampling from internal signals, so training can stop when the cluster count derivative, centroid movement, and NMI or ARI scores stabilize.
  • Stratified sampling across the 255 business categories is what preserves coverage while shrinking the set.
  • Because the hierarchy is stable under additional data, new questions can be assigned to existing intents without full retraining, and only genuinely novel clusters require structure updates.
  • The reported 92.5% expert accuracy on sampled clusters and prototypes indicates the discovered clusters are usable as intents, not merely mathematical partitions.

Reading between the lines

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

  • The 20,400-utterance figure is not a universal minimum; it is a property of this dataset, this embedding model, and the chosen merging thresholds, so the method's general value is the protocol for finding the stabilization point rather than the specific number.
  • A testable extension is to apply the identical protocol to a different domain, such as travel or healthcare questions, and check whether internal stability again predicts a sharp data-reduction ratio.
  • Because coverage is measured internally, a stronger validation would compare the 20,400-utterance hierarchy's prototypes against human-labeled intent taxonomies or downstream retrieval quality; the paper's own 500-cluster expert review is a first step but not an independent coverage measure.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 a neural attention-driven hierarchical clustering algorithm for conversational recommendation systems, combining ideas from flat clustering methods DEC and NAM. The method is evaluated on 44,112 food-business-domain questions with 255 business categories, using stratified sampling. The central claim is that the hierarchical intent structure stabilizes after approximately 20,400 utterances, so that this much data is sufficient for adequate domain intent coverage. The evaluation consists of internal stability metrics (cluster count derivative, centroid movement, NMI/ARI between consecutive runs), cluster quality metrics, and a human evaluation of 500 sampled clusters.

Significance. If the 20,400-utterance claim were supported by external ground truth, the result would be practically valuable for reducing annotation cost in evolving conversational recommendation domains. The paper gives a fairly detailed algorithmic description, and the idea of merging DEC-style embedding optimization with NAM-style attention in a hierarchical merging framework is reasonable. The stratified sampling procedure is clearly described. However, the significance is currently conditional: the central claim rests on internal stability metrics rather than external validation, no baseline comparisons are provided, and the paper's own reported numbers conflict with its stated stabilization criteria. With additional experiments against the 255 available category labels and repaired consistency of the decision rule, the contribution could be meaningfully established.

major comments (4)
  1. [Section 4 and Section 4.2] The decision rule in Section 4 states that the optimal dataset size is identified when 'NMI and ARI exceed 0.85,' but Section 4.2 reports that at the claimed stabilization point the NMI stability score is 'closure to .80' and the ARI stability score is 'closure to 0.65.' The conclusion that the structure stabilized at approximately 20,400 utterances is therefore not supported by the paper's own criterion. Because this stabilization point is the sole empirical basis for the Section 6 claim that approximately 20,400 utterances are sufficient for adequate domain coverage, this internal inconsistency is load-bearing.
  2. [Section 4, Section 4.2, and Section 6] All the stability criteria used to identify the 20,400-utterance point (cluster count derivative, centroid movement, and NMI/ARI between consecutive runs) are computed from the algorithm's own clusterings on the same embedding space. Satisfying these thresholds shows that the algorithm's partitions converge as data are added; it does not show that the resulting hierarchy covers the true domain intent space. The human evaluation in Section 4.2 rates coherence and practical usefulness of 500 clusters, not coverage of the 255 business categories. An external evaluation against the 255 category labels is necessary before the paper can claim 'adequate domain knowledge coverage.'
  3. [Section 4.2 and Section 6] The paper claims its hierarchical algorithm is 'more suitable' than flat DEC and NAM, but no experimental comparison against these or any other baselines is reported. Without baseline runs at the same dataset sizes, the paper cannot support the causal claim that the proposed method, rather than the stratified sampling or the embedding setup, is responsible for the observed stabilization behavior.
  4. [Section 3.1 and Algorithm 1] Several free parameters and implementation details needed to reproduce the experiments are not specified: kmax, tau_min, the annealing factor alpha, the large-cluster split threshold m, the contrastive threshold tau_contrast, the number of ANN neighbors M, and the sentence-embedding model. Additionally, Equations (2) and (3) appear dimensionally inconsistent: W2 has shape 1 x h, so the attention vector A is not shaped as an n x d mask, making the element-wise multiplication in Equation (3) ill-defined as written. These details are necessary for verification and for interpreting the quantitative claims.
minor comments (5)
  1. [Figure 5] The caption contains a typo: 'Qulaity metrics' should read 'Quality metrics.'
  2. [Section 4.2] The sentence 'the percentage of validation samples assigned consistently the same with low confidence' is unclear; please define the low-confidence assignment rate precisely and state the value observed at the stabilization point.
  3. [Section 3.2] The complexity analysis mixes worst-case and practical assumptions: after deriving O(n^2.5 d) worst-case complexity from the contrastive refinement term, the text states a practical complexity closer to O(n log n + n d) without specifying the conditions under which the lower bound applies. This transition should be justified.
  4. [Conclusion] The sentence 'Our algorithm does not support learning in one domain, and after it can be transferred or adapted to new domains' appears to be missing words and is difficult to parse; please revise.
  5. [References] Reference [7] is cited as the source of Deep Embedded Clustering (DEC), but the original DEC method is usually attributed to Xie et al. (Reference [8]); please verify the attribution or clarify that a modified variant is used.

Circularity Check

1 steps flagged · score 6.0 of 10

The 20,400-utterance sufficiency result is the point where the paper's own internal stability criteria are asserted to be met; the reported NMI/ARI values are below the stated 0.85 threshold.

  1. self definitional [Section 4 (optimal-size criteria); Section 4.2 (Result Analysis); Section 6 (Conclusion)]
    "We determine the optimal dataset size by identifying: 1. When the cluster count stabilizes (derivative approaches zero) 2. When cluster movement falls below 5 percent. 3. When NMI and ARI exceed 0.85. 4. When quality metrics reach their optimal values. The analysis of various stability metrics revealed that the clustering structure stabilized after approximately 20400 utterances. But the NMI stability score closure to .80 at this dataset size ... and similarly, the ARI stability score closure to 0.65 ..."

    The claimed 'adequate domain knowledge coverage' (Section 6) is operationalized entirely by stability metrics computed from the algorithm's own clusterings: cluster-count derivative, centroid movement, and NMI/ARI between consecutive runs. There is no external labeled intent hierarchy or independent coverage measure. The 20,400-utterance number is therefore, by construction, the dataset size at which the author-chosen internal thresholds are asserted to be met; the 'prediction' restates the criterion rather than testing it. In addition, the values actually reported at that point (≈0.80 NMI, ≈0.65 ARI) are below the paper's own 0.85 criterion, so the stabilization conclusion is not even forced by the stated decision rule; it is an asserted fit.

full rationale

The central circularity is in the evaluation design, not in citation practice. The paper cites DEC [7] and NAM [9] as external prior work, and there is no reliance on the authors' own prior publications or a uniqueness theorem, so patterns 3-5 are not present. The load-bearing issue is that the 'optimal/minimal dataset size' (≈20,400) is identified by four internal stability criteria applied to the algorithm's own clusterings, and the same stabilization is then presented as evidence of 'adequate domain knowledge coverage.' No external ground-truth intent structure, held-out category coverage test, or downstream retrieval evaluation is used to connect stability to coverage; the human evaluation rates coherence and usefulness of 500 clusters, not coverage of the 255 business categories. Thus the 20,400 number is a fitted cutoff: the point where the algorithm's self-defined curves are asserted to plateau. This makes the central 'prediction' reduce by construction to the criterion. The internal inconsistency that NMI (≈0.80) and ARI (≈0.65) are below the stated 0.85 threshold compounds the issue: the conclusion is not even supported by the paper's own decision rule. Score 6 rather than 8 because the clustering algorithm and sampling methodology are independent contributions and the human coherence evaluation provides some internal quality signal; the circularity is confined to the coverage/minimal-data claim.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several author-defined quantities, most with unreported values, and on domain assumptions about the curated categories and embedding quality. The merging similarity H is never defined, and the contrastive effect is asserted rather than derived. No invented entities are introduced; the hierarchy is a data structure, not a new physical or conceptual entity.

free parameters (8)
  • Adaptive cluster count k = min(sqrt(n), kmax) = The value of kmax is not reported.
    Initial cluster count is a heuristic formula with an unreported kmax; it controls the granularity of the hierarchy and the merging cost.
  • Initial merging threshold tau (75th percentile of H(Ci,Cj)) = The value is data-dependent and not reported.
    The merging threshold is set as a quantile of the undefined similarity H; changing it changes when merging stops.
  • Minimum threshold tau_min = The value is not reported.
    Minimum merging threshold in Algorithm 1; controls depth of hierarchy.
  • Annealing factor alpha = The value is not reported.
    Threshold decay in Eq. 7; no value or schedule is given.
  • Large-cluster split threshold m = The value is not reported.
    Triggers contrastive refinement for clusters with more than m members.
  • Contrastive threshold tau_contrast = The value is not reported.
    Used in Eq. 11; not defined anywhere in the paper.
  • Stability criteria (derivative < 0.001, centroid movement < 5%, NMI/ARI > 0.85) = Author-chosen thresholds; results do not meet them.
    These criteria determine the claimed minimal size of 20,400; Section 4.2 reports NMI around 0.80 and ARI around 0.65, below the stated 0.85.
  • Prototype count (top-3) = 3
    Number of prototypes per leaf is chosen without justification; changing it changes coverage claims.
assumptions (5)
  • domain assumption The curated 44,112 questions and 255 business domain categories represent the full business food intent space.
    Stratified sampling is based on these categories, but how they were collected or validated is not described in Sections 1 and 4.1.
  • domain assumption Pretrained sentence embeddings, refined by a two-layer attention network, capture intent-relevant semantic similarity.
    The method's success depends on the embedding space, but no specific embedding model or training details are given in Eqs. 1 to 3.
  • ad hoc to paper Centroid-based reassignment 'implicitly optimizes the same objective' as a contrastive loss without an explicit loss.
    Stated in Section 3.1; this is asserted, not derived, and the equivalence is not formalized.
  • domain assumption Internal self-consistency metrics (cluster count derivative, centroid movement, NMI/ARI stability) are valid measures of domain coverage.
    The central 20,400-utterance claim rests on these author-defined thresholds in Section 4; no external ground truth is used.
  • ad hoc to paper Entropy-based cluster similarity H(Ci,Cj) is well-defined and can be computed.
    H is used throughout Algorithm 1 and Eqs. 6 to 8 but never defined in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Minimally Supervised Hierarchical Domain Intent Learning for CRS." pith.science (2026). https://pith.science/paper/P3EC42AB

@misc{pith2026250502209,
  author       = {Pith},
  title        = {Pith review of: Minimally Supervised Hierarchical Domain Intent Learning for CRS},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3EC42AB}},
  note         = {Machine review of arXiv:2505.02209}
}
read the original abstract

Modeling domain intent within an evolving domain structure presents a significant challenge for domain-specific conversational recommendation systems (CRS). The conventional approach involves training an intent model using utterance-intent pairs. However, as new intents and patterns emerge, the model must be continuously updated while preserving existing relationships and maintaining efficient retrieval. This process leads to substantial growth in utterance-intent pairs, making manual labeling increasingly costly and impractical. In this paper, we propose an efficient solution for constructing a dynamic hierarchical structure that minimizes the number of user utterances required to achieve adequate domain knowledge coverage. To this end, we introduce a neural network-based attention-driven hierarchical clustering algorithm designed to optimize intent grouping using minimal data. The proposed method builds upon and integrates concepts from two existing flat clustering algorithms DEC and NAM, both of which utilize neural attention mechanisms. We apply our approach to a curated subset of 44,000 questions from the business food domain. Experimental results demonstrate that constructing the hierarchy using a stratified sampling strategy significantly reduces the number of questions needed to represent the evolving intent structure. Our findings indicate that this approach enables efficient coverage of dynamic domain knowledge without frequent retraining, thereby enhancing scalability and adaptability in domain-specific CSRs.

Figures

Figures reproduced from arXiv: 2505.02209 by the authors.

Figure 1
Figure 1. Minimal Data Requirement for Optimal Hierarchical Intent Clustering a crucial role in organizing user utterances into coherent groups that represent underlying intents. Existing clustering-based methods for intent recognition of￾ten suffer from inefficiencies in data utilization. Existing methods, such as flat clustering techniques like Deep Embedded Clustering (DEC) [7] and Neural At￾tention Models (NAM) [9], strug… view at source ↗
Figure 2
Figure 2. Dataflow of Hierarchical clustering algorithm Our hierarchical clustering algorithm strategically integrates key aspects from both DEC and NAM to create a more effective intent clustering system. From DEC, we adopt the concept of learning optimized embedding representations during the clustering process, but modify it to work within a hierarchical frame￾work. Specifically, while traditional DEC jointly optimizes emb… view at source ↗
Figure 3
Figure 3. Stability and Performance of the Hierarchical Clustering [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Clustering statistics [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Qulaity metrics [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

  1. [1]

    An Iterative Self- Learning Framework for Medical Domain Generalization

    Zhenbang Wu, Huaxiu Yao, David Liebovitz, and Jimeng Sun. An Iterative Self- Learning Framework for Medical Domain Generalization. In Advances in Neural Information Processing Systems, volume 36, pages 54833–54854. Curran Associates, Inc., 2023

    work page 2023

  2. [2]

    Discovering New Intents with Deep Aligned Clustering

    Hanlei Zhang, Hua Xu, Ting-En Lin, and Rui Lv. Discovering New Intents with Deep Aligned Clustering. CoRR, abs/2012.08987, 2020

    work page Pith review arXiv 2012

  3. [3]

    Open Intent Discovery through Unsupervised Semantic Clustering and Dependency Parsing

    Pengfei Liu, Youzhang Ning, King Keung Wu, Kun Li, and Helen Meng. Open In- tent Discovery through Unsupervised Semantic Clustering and Dependency Parsing. arXiv preprint arXiv:2104.12114 , 2021

    work page Pith review arXiv 2021

  4. [4]

    Supervised Clustering of Questions into Intents for Dialog System Applica- tions

    Iryna Haponchyk, Antonio Uva, Seunghak Yu, Olga Uryupina, and Alessandro Mos- chitti. Supervised Clustering of Questions into Intents for Dialog System Applica- tions. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2310–2321, Brussels, Belgium, October-November 2018. Association for Computational Linguistics

    work page 2018

  5. [5]

    Achtermann; Indrajit Bhattacharya; Kevin W

    Jeffrey M. Achtermann; Indrajit Bhattacharya; Kevin W. Fairfield; Shantanu R. Godbole,; Sachindra Joshi; Ashwin Srinivasan; Ashish Verma. Cross-domain clus- terability evaluation for cross-guided data clustering based on alignment between data domains. U.S. Patent No. US 8,229,929 B2, Jan. 6, 2010, Year, Jul. 24, 2012

    work page 2010

  6. [6]

    Supervised Clustering Loss for Clustering-Friendly Sentence Embeddings: Minimally Supervised Hierarchical Domain Intent Learning for CRS 15 an Application to Intent Clustering

    Giorgio Barnabò, Antonio Uva, Sandro Pollastrini, Chiara Rubagotti, and Davide Bernardi. Supervised Clustering Loss for Clustering-Friendly Sentence Embeddings: Minimally Supervised Hierarchical Domain Intent Learning for CRS 15 an Application to Intent Clustering. InFindings of the Association for Computa- tional Linguistics: IJCNLP-AACL 2023 (Findings) ...

    work page 2023

  7. [7]

    Deep Embedded Clustering with Data Augmentation

    Xifeng Guo, En Zhu, Xinwang Liu, and Jianping Yin. Deep Embedded Clustering with Data Augmentation. InAsian Conference on Machine Learning , 2018

    work page 2018

  8. [8]

    Girshick, and Ali Farhadi

    Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Unsupervised Deep Embedding for Clustering Analysis. CoRR, abs/1511.06335, 2015

    arXiv 2015

Show all 23 references
  1. [9]

    An Unsupervised Neural Attention Model for Aspect Extraction

    Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Dahlmeier. An Unsupervised Neural Attention Model for Aspect Extraction. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 388–397, Vancouver, Canada, July 2...

    2017

  2. [10]

    Bing Liu and Ian R. Lane. Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling.CoRR, abs/1609.01454, 2016

    2016 arXiv

  3. [11]

    BERT for Joint Intent Classification and Slot Filling

    Qian Chen, Zhu Zhuo, and Wen Wang. BERT for Joint Intent Classification and Slot Filling. CoRR, abs/1902.10909, 2019

    1902 arXiv

  4. [12]

    Discovering New Intents via Constrained Deep Adaptive Clustering with Cluster Refinement.CoRR, abs/1911.08891, 2019

    Ting-En Lin, Hua Xu, and Hanlei Zhang. Discovering New Intents via Constrained Deep Adaptive Clustering with Cluster Refinement.CoRR, abs/1911.08891, 2019

    1911 arXiv

  5. [13]

    Intent Mining from past conversations for Conversational Agent.CoRR, abs/2005.11014, 2020

    Ajay Chatterjee and Shubhashis Sengupta. Intent Mining from past conversations for Conversational Agent.CoRR, abs/2005.11014, 2020

    2005 arXiv

  6. [14]

    Crook, Bing Liu, Zhou Yu, Eunjoon Cho, and Zhiguang Wang

    Andrea Madotto, Zhaojiang Lin, Zhenpeng Zhou, Seungwhan Moon, Paul A. Crook, Bing Liu, Zhou Yu, Eunjoon Cho, and Zhiguang Wang. Continual Learning in Task-Oriented Dialogue Systems.CoRR, abs/2012.15504, 2020

    2012 arXiv

  7. [15]

    Incremental Domain Adaptation for Neural Machine Translation in Low-Resource Settings

    Marimuthu Kalimuthu, Michael Barz, and Daniel Sonntag. Incremental Domain Adaptation for Neural Machine Translation in Low-Resource Settings. InProceed- ings of the Fourth Arabic Natural Language Processing Workshop , pages 1–10, Flo- rence, Italy, August 2019. Association for...

    2019

  8. [16]

    Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,”J. Comput. Appl. Math. , vol. 20, pp. 53–65, 1987

    1987

  9. [17]

    A dendrite method for cluster analysis,

    T. Calinski and J. Harabasz, “A dendrite method for cluster analysis,”Commun. Stat., vol. 3, no. 1, pp. 1–27, 1974

    1974

  10. [18]

    A cluster separation measure,

    D. L. Davies and D. W. Bouldin, “A cluster separation measure,”IEEE Trans. Pattern Anal. Mach. Intell. , vol. 1, no. 2, pp. 224–227, 1979

    1979

  11. [19]

    Cluster ensembles—a knowledge reuse framework for combining multiple partitions,

    A. Strehl and J. Ghosh, “Cluster ensembles—a knowledge reuse framework for combining multiple partitions,”J. Mach. Learn. Res. , vol. 3, pp. 583–617, 2002

    2002

  12. [20]

    Comparing partitions,

    L. Hubert and P. Arabie, “Comparing partitions,”J. Classif., vol. 2, pp. 193–218, 1985

    1985

  13. [21]

    Algorithmic bias: on the implicit biases of social technology,

    G. M. Johnson, “Algorithmic bias: on the implicit biases of social technology,” Synthese, vol. 198, pp. 9941–9961, 2021, doi: 10.1007/s11229-020-02696-y

    2021 doi

  14. [22]

    Dial-In LLM: Human-Aligned LLM-in-the-loop Intent Clustering for Customer Service Dialogues,

    M. Hong et al., “Dial-In LLM: Human-Aligned LLM-in-the-loop Intent Clustering for Customer Service Dialogues,”arXiv preprint arXiv:2412.09049 , 2025

    2025

  15. [23]

    ClusterPrompt: Cluster Semantic Enhanced Prompt Learn- ing for New Intent Discovery,

    J. Liang and L. Liao, “ClusterPrompt: Cluster Semantic Enhanced Prompt Learn- ing for New Intent Discovery,” inFindings of the Association for Computational Linguistics: EMNLP 2023 , Singapore, Dec. 2023, pp. 10468–10481

    2023

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.