Pith. sign in

REVIEW 3 major objections 4 minor 54 references

A fully unsupervised multimodal intent discovery method uses MLLM-generated concepts to guide semantic propagation, beating existing methods on three benchmarks while producing interpretable concept labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

MCSP improves unsupervised multimodal intent discovery by using MLLM-generated semantic concepts as anchors and propagating them over a concept-weighted graph before contrastively refining representations.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Useful pipeline with real ablations, but the headline empirical claim is not yet supported: the evaluation mixes two split protocols and tunes key temperatures on test labels. the 3 major comments →

arxiv 2607.21908 v1 pith:JKO5OMSN submitted 2026-07-24 cs.MM

Unsupervised Multimodal Intent Discovery via MLLM-Guided Concept Generation and Semantic Propagation

Yunjin Gu , Qianrui Zhou , Hua Xu This is my paper
classification cs.MM
keywords unsupervised intent discoverymultimodal dialogue understandingMLLM-guided concept generationsemantic propagationrandom walk label propagationcontrastive clusteringinterpretable clusteringmultimodal intent
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 paper proposes MCSP, a fully unsupervised pipeline for discovering intents in unlabeled multimodal dialogues (text, audio, and video). Instead of refining clusters by geometric proximity alone, it asks a multimodal large language model to read a few confident examples from each cluster and name the intent in a short phrase, producing an interpretable concept. These concepts are then diffused over a neighborhood graph through a random walk, and the resulting pseudo-labels drive contrastive learning that reshapes the embedding space. The paper shows the method outperforms prior clustering and intent-discovery baselines on MIntRec, MIntRec2.0, and MELD-DA, and that the discovered clusters come with readable concept labels. A sympathetic reader would take away that adding high-level semantic guidance, via a handful of MLLM calls, can both improve accuracy and make the clusters explainable.

Core claim

MCSP claims that semantic concepts, not geometry, should anchor unsupervised multimodal intent discovery. Starting from k-means++ clusters, it picks high-margin representative samples per cluster (Eq. 2-3), uses an MLLM to contrast the target cluster against its nearest neighbor and produce a 2-6 word gerund phrase naming the cluster's communicative function (Eq. 4), and globally normalizes the resulting concept set (Eq. 5). The concepts are encoded as anchors, a k-NN graph is built with edge weights modulated by semantic consistency between concept distributions (Eq. 6-8), and a random walk initialized at representatives propagates conceptual beliefs to all samples (Eq. 9). The propagated p

What carries the argument

The central mechanism is the combination of MLLM-guided contrastive concept generation with semantic propagation over a concept-aware graph: high-margin representatives seed the MLLM with clean evidence, the MLLM outputs a short concept phrase per cluster, and those phrases are encoded into anchors that modulate graph edge weights and initialize a random-walk label diffusion; a concept-supervised contrastive loss then closes the loop. In one phrase: 'random walk on a semantically weighted graph seeded by MLLM-generated concepts.' This carries the argument because it converts abstract cluster geometry into explicit semantics that can be propagated locally and used to re-shape the embedding ma

Load-bearing premise

The load-bearing assumption is that the initial k-means clusters are coherent enough that the top margin-ranked representatives are semantically pure, and that the number of intent categories is known and matches the ground truth; if either fails, the MLLM generates a wrong or noisy concept and the random-walk propagation reinforces that error.

What would settle it

Run MCSP on a benchmark where the ground-truth category count is deliberately mis-estimated by a small margin (e.g., K true 20 versus K set to 18 or 22) and compare against UMC with the correct K: the paper's F.1 experiment already reports a narrowing of the advantage to 0.72 average points when K is estimated, so a setting where the estimate is off by more than a couple of categories would show whether the method still beats the baseline.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If MCSP is correct, fully unsupervised intent discovery no longer has to choose between accuracy and interpretability: clusters come with human-readable concept labels (e.g., 'announcing departure', 'confronting or accusing').
  • The method shows that a small number of MLLM calls (one per cluster plus one global refinement) can inject semantic guidance into clustering, suggesting a practicable route to semantic discovery at scale without exhaustive labeling.
  • Since the pseudo-labels arise from propagated concepts, the resulting representation is structured by semantics, which could make downstream tasks like intent classification and out-of-scope detection start from better organization of unlabeled data.
  • The ablation study (w/o G etc.) indicates that the graph propagation step, not just the concept names, is responsible for the gains; thus concept-driven diffusion is the load-bearing component.
  • The reported stability across seeds and concepts suggests the discovered intents are reproducible, which is a prerequisite for treating them as reliable labels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper assumes K is known; when K is estimated, the advantage over UMC shrinks to under one average point (F.1), so the practical upside may be smaller in genuinely open-world settings where the number of intents is unknown.
  • If the MLLM's reasoning fails on a dataset whose intents are more visually/graphically nuanced or where the top-3 representatives are not semantically pure, the propagation will reinforce wrong boundaries; a natural test is to swap the MLLM for a text-only LLM and see whether the multimodal grounding contributes.
  • The concept-stability metric (0.91 average cosine similarity across seeds) suggests the method could generate reusable intent taxonomies for a domain; these taxonomies might bootstrap few-shot intent classifiers, but that extrapolation goes beyond the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes MCSP, an unsupervised multimodal intent discovery method. It first uses UMC-style contrastive pre-training, then selects high-margin representatives from k-means clusters, prompts an MLLM to generate discriminative semantic concepts for each cluster (Eqs. 2–5), and finally propagates these concepts over a semantically reweighted k-NN graph to produce pseudo-labels that supervise a concept-contrastive loss (Eqs. 6–12). The authors claim MCSP consistently outperforms existing methods on MIntRec, MIntRec2.0, and MELD-DA while yielding interpretable concept labels.

Significance. If the empirical claims are supported, the paper would make a useful contribution: it moves multimodal intent clustering from purely geometric refinement toward semantically interpretable, MLLM-grounded concepts, and it backs this with a coherent three-stage pipeline. The ablation study (Table 2) is informative, the stability analyses (Tables 7–8) and cost analysis (Appendix G) are thoughtful, and the case study (Section 5.4) illustrates interpretability. The main positive result is not verified by released code, but the appendices contain enough detail for replication attempts. The contribution is significant for the multimedia clustering community, provided the evaluation-protocol and hyperparameter-selection concerns below are resolved.

major comments (3)
  1. [§4.1 vs. Appendix C.2] The evaluation protocol is internally contradictory and this is load-bearing for every number in Table 1. Section 4.1 states that 'we merge the original train, validation, and test splits and repartition the data 4:1 for training and testing.' Appendix C.2 states, for the same datasets, that 'we merge the original training and validation splits for unlabeled training and retain the original test split for evaluation.' These protocols produce different test sets, so the reported gains over UMC and SPILL cannot be interpreted unless all methods were re-run under exactly one protocol. Please specify the single protocol used for all methods, confirm that every baseline was re-evaluated under that protocol, and report the results in a way that removes split incomparability.
  2. [§4.3, §5.2, and §D.3] The temperature triplets (tau, tau_c, tau_g) are tuned with ground-truth labels, which contradicts the claim of a fully unsupervised evaluation. Section 4.3 says the triplets are configured 'as discussed in Section 5.2,' and Section 5.2 reports accuracy peaks on MIntRec2.0—necessarily using ground-truth labels to compute accuracy. The label-free stability criterion in Appendix D.3, run on MIntRec, selects a different configuration (0.3, 1.5, 1.0) rather than the used (0.5, 1.0, 0.7). Thus the main results are not label-free. Report the main table under the label-free stability criterion of D.3, or clearly separate oracle-tuned upper-bound results from label-free results.
  3. [§F.1 and §4.4] The paper's central claim is stated without the caveat that the main experiments assume K is known and equal to the ground-truth category count. Section F.1 shows that with estimated K, MCSP's advantage over UMC shrinks to 0.72 average points (24.47 vs. 23.75), much smaller than the 1.5-point margin in the main table. This is not a fatal flaw—known K is a standard assumption in unsupervised clustering—but the abstract and conclusion should qualify 'fully unsupervised' and the consistency claim accordingly, since the discovery setting in the title implies unknown K.
minor comments (4)
  1. [References] References [2] and [3] are the same Ben-Hur et al. paper duplicated with different citation years (2002 in both). Please merge them and fix the citations in §D.3 and §F.3.
  2. [Figures 10–11] The captions for Figures 10 and 11 say 'MIntRec2.0' and 'MELD-DA', respectively, but the panel labels inside the figures appear to be carried over from Figure 5 (e.g., 'MIntRec / MIntRec2.0'). Please make the panel labels consistent with the dataset actually shown.
  3. [§5.2] The sensitivity analysis reports only accuracy ('peaks at 29.19%') for each temperature. Since the main evaluation uses four metrics, specify whether other metrics behave similarly or provide the curves for ARI/NMI/FMI in Figure 4.
  4. [§4.3] The statement 'we keep most hyperparameters the same across datasets and adjust only a few parameters' is followed by per-dataset temperatures, learning rates, and batch size. Clarify which are considered 'training stability' parameters and which are method hyperparameters, since the latter should not be tuned per dataset in a fully unsupervised claim.

Circularity Check

1 steps flagged

Reported SOTA gains are partly fitted: temperature hyperparameters are chosen by test-set accuracy, not by the paper's label-free criterion.

specific steps
  1. fitted input called prediction [Section 4.3; Section 5.2; Appendix D.3]
    ""The temperature triplets (τ,τ_c,τ_g) are configured as (0.5, 1.0, 0.7), (0.3, 0.3, 7.5), and (2.5, 1.0, 3.5), respectively, as discussed in Section 5.2." In Section 5.2: "For the τ in concept-supervised contrastive learning, the accuracy reaches a peak of 29.19% when τ is set to 0.5.""

    The paper's central empirical claim is that MCSP is a fully unsupervised method that 'consistently outperforms' baselines. But the per-dataset temperature triplets used in the final tables are selected through Section 5.2, which maximizes ground-truth ACC on MIntRec2.0. The reported 29.19% accuracy on MIntRec2.0 is the very peak used to choose the temperature, so that result (and the correlated gains) are not independent predictions; they are the result of fitting control parameters to the evaluation labels. Appendix D.3's label-free stability criterion selects different configurations (e.g., highest stability 0.4000 at (0.3,1.5,1.0)), so the paper's own unsupervised selection procedure does not justify the configurations used to claim SOTA performance.

full rationale

The clearest circular element is in the evaluation, not the algorithm: Section 4.3 fixes the temperature triplets 'as discussed in Section 5.2', and Section 5.2 explicitly reports peaks in accuracy (ground-truth) on MIntRec2.0. Thus the Table 1 performance numbers are selected on the test labels, undermining the 'fully unsupervised ... consistently outperforms' claim. I do not count the Eq. (3)->Eq. (9)->Eq. (12) pseudo-label loop as a definitional circularity: representatives are chosen from an initial k-means partition and their labels seed the random walk, but propagation can reassign non-anchor samples, so the final clustering is not identical to the input by construction; it is a self-training loop with external MLLM naming. This loop does mean the 'semantic concepts' are summaries of the same clusters they later supervise, so they are not independent evidence, but that is a limitation of the method, not a formal reduction. The UMC architecture citation is legitimate prior work, not load-bearing self-citation. The paper also contains a non-circular evaluation inconsistency: Section 4.1 says all splits are merged and repartitioned 4:1, while Appendix C.2 says the original train+val are used and the original test is retained; this should be resolved but is not itself circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central claim rests on a self-training loop over initial k-means clusters and on several human-chosen hyperparameters. The free parameters are dominated by final temperatures and confidence cutoffs selected with label-dependent metrics; the axioms show the key domain assumptions (K known, margin samples semantic, MLLM concepts correct) that are imperfect by the paper's own case study.

free parameters (6)
  • cluster number K = 20 (MIntRec), 30 (MIntRec2.0), 12 (MELD-DA)
    Main experiments set K to the ground-truth category count, which is oracle information in an unsupervised setting; estimated-K results in Table 6 show smaller gains.
  • temperature triplets (tau, tau_c, tau_g) = (0.5,1.0,0.7), (0.3,0.3,7.5), (2.5,1.0,3.5) per dataset
    Chosen per dataset; Section 5.2 motivates them using clustering accuracy on MIntRec2.0, a label-dependent metric.
  • confidence ratio rho = 0.6
    Selected as the peak of Figure 9 using ARI, a ground-truth-dependent metric.
  • graph size k and semantic weight lambda = k=10, lambda=3.0
    Fixed by hand; no label-free selection procedure is provided.
  • diffusion factor alpha and propagation steps = alpha=0.95, steps=40
    Fixed uniformly across datasets; no sensitivity analysis is reported for these choices.
  • representative/candidate counts (m, N_r) = m=40, N_r=3
    Chosen to limit MLLM cost rather than optimized for clustering performance.
axioms (6)
  • domain assumption k-means++ clusters in the fused embedding space approximate the true latent intents
    Section 3.2: all representative selection and concept generation builds on these initial clusters.
  • domain assumption Euclidean margin score identifies semantically unambiguous, prototypical samples
    Equations (2)-(3): margin is a geometric heuristic, not a semantic guarantee.
  • domain assumption MLLM contrastive reasoning over three representatives yields correct, generalizable concepts
    Equations (4)-(5); Section 5.4 admits failures such as sarcasm and coarse boundaries, so this assumption is only partially met.
  • standard math Random-walk label propagation on a k-NN graph with fixed alpha converges to meaningful soft labels
    Equation (9) is standard harmonic label propagation (Zhu et al. 2003); convergence is known but quality depends on the graph construction.
  • domain assumption Multi-view masking preserves the semantic identity of an utterance
    Section 3.3: masking audio/video assumes the textual modality alone carries the intent for the positive pairs.
  • ad hoc to paper True category count K is available in the main evaluation
    Used throughout Section 4; the paper treats unknown K as a separate robustness setting in Appendix F.1.
invented entities (1)
  • MLLM-generated concept anchors (t'_k encoded as c_k) no independent evidence
    purpose: Semantic pseudo-labels that seed graph propagation and supervise contrastive refinement
    The concepts are introduced by the paper and are not checked against any external concept taxonomy; their validity is inferred only from final clustering accuracy.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Multimodal Intent Discovery via MLLM-Guided Concept Generation and Semantic Propagation." pith.science (2026). https://pith.science/paper/JKO5OMSN

@misc{pith2026260721908,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Multimodal Intent Discovery via MLLM-Guided Concept Generation and Semantic Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKO5OMSN}},
  note         = {Machine review of arXiv:2607.21908}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unsupervised multimodal intent discovery aims to uncover latent intents from unlabeled multimodal dialogues, but remains challenging due to the lack of explicit semantic supervision. Existing methods often provide limited interpretability, as their refinement mainly relies on geometric similarity rather than high-level semantic guidance. To address these limitations, we propose MCSP, a fully unsupervised method that introduces semantic refinement based on concepts into multimodal intent discovery. To obtain reliable semantic evidence for intent discovery, we identify high-quality representative samples for each cluster and use them to support MLLM-guided contrastive reasoning against neighboring clusters, which produces interpretable high-level semantic concepts. Building on these concepts, we perform semantic propagation over a semantically weighted graph to align conceptual information with local structural consistency and generate reliable pseudo-labels for representation refinement. Extensive experiments on three challenging multimodal intent datasets show that MCSP consistently outperforms state-of-the-art methods while producing interpretable clusters grounded in semantic concepts.

Figures

Figures reproduced from arXiv: 2607.21908 by Hua Xu, Qianrui Zhou, Yunjin Gu.

Figure 1
Figure 1. Figure 1: Comparison of intent discovery paradigms. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed MCSP method. The MCSP proceeds through three synergistic stages to achieve precise [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multimodal Contrastive Reasoning. strategy, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Temperature hyper-parameter sensitivity analysis on MIntRec2.0 measured by accuracy (%). [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of representations on the MIntRec dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Case study on MIntRec. Representative samples display video, text, and GT labels. Purple text indicates generated [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Promptcont for multimodal contrastive reasoning, focusing on cluster A against nearest neighbor B. an alignment protocol that binds video segments to transcripts. Crucially, a contrastive mechanism tasks the model with isolating unique functions and subtracting shared traits from nearest neigh￾bors, thereby avoiding generic outputs in favor of discriminative labels. The output is constrained to a 2 to 6 wo… view at source ↗
Figure 8
Figure 8. Figure 8: Promptglobal used to ensure cross-cluster label con￾sistency and syntactical alignment. unlabeled training and retain the original test split for evaluation. No ground-truth labels are used during training or clustering. C.3 Evaluation Details We evaluate clustering performance using four standard metrics: Accuracy (ACC), Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), and Fowlkes-Mallows I… view at source ↗
Figure 9
Figure 9. Figure 9: Effect of confidence ratio 𝜌 on the ARI metric. its stability is computed as the average pairwise NMI between the resulting pseudo-label assignments. This criterion requires no ground-truth labels and favors configurations that consistently re￾cover similar cluster structures across runs [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of representations on the MELD-DA dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 11 canonical work pages

  1. [1]

    David Arthur and Sergei Vassilvitskii. 2007. k-means++: The Advantages of Careful Seeding. InProceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA ’07), Vol. 8. 1027–1035. doi:10.1145/1283383.1283494

    arXiv 2007

  2. [3]

    Asa Ben-Hur, Andre Elisseeff, and Isabelle Guyon. 2002. A Stability Based Method for Discovering Structure in Clustered Data.Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing2002 (02 2002), 6–17. doi:10.1142/9789812799623_0002

    work page doi:10.1142/9789812799623_0002 2002

  3. [4]

    Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. 2018. Deep Clustering for Unsupervised Learning of Visual Features. InProceedings of the European Conference on Computer Vision (ECCV)

    2018

  4. [6]

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Xiangzhan Yu, and Furu Wei. 2022. WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing.IEEE Journal of Selected Topics in Sig...

    arXiv 2022

  5. [7]

    Zhanpeng Chen, Zhihong Zhu, Xianwei Zhuang, Zhiqi Huang, and Yuexian Zou

  6. [8]

    Jackie Chi Kit Cheung and Xiao Li. 2012. Sequence clustering and labeling for unsupervised query intent discovery. InProceedings of the Fifth ACM International Conference on Web Search and Data Mining(Seattle, Washington, USA)(WSDM ’12). Association for Computing Machinery, New York, NY, USA, 383–392. doi:10. 1145/2124295.2124342

    arXiv 2012

  7. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Jill Burstein, Christy...

    2019

  8. [10]

    Jairo Diaz-Rodriguez. 2026. Summaries as Centroids for Interpretable and Scal- able Text Clustering. InThe Fourteenth International Conference on Learning Representations. https://openreview.net/forum?id=Uzku7RZXvI

    2026

  9. [11]

    Zomaya, Sebti Foufou, and Abdelaziz Bouras

    Adil Fahad, Najlaa Alshatri, Zahir Tari, Abdullah Alamri, Ibrahim Khalil, Albert Y. Zomaya, Sebti Foufou, and Abdelaziz Bouras. 2014. A Survey of Clustering Algorithms for Big Data: Taxonomy and Empirical Analysis.IEEE Transactions on Emerging Topics in Computing2, 3 (2014), 267–279. doi:10.1109/TETC.2014. 2330519

    work page doi:10.1109/tetc.2014 2014

  10. [12]

    Zijin Feng, Luyang Lin, Lingzhi Wang, Hong Cheng, and Kam-Fai Wong. 2024. LLMEdgeRefine: Enhancing Text Clustering with LLM-Based Boundary Point Refinement. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Fl...

    doi:10.18653/v1/2024.emnlp-main.1025 2024

  11. [13]

    Godfrey, Edward C

    John J. Godfrey, Edward C. Holliman, and Jane McDaniel. 1992. SWITCHBOARD: Telephone speech corpus for research and development. InProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, I/517–I/520. doi:10.1109/ICASSP.1992.225957

    arXiv 1992

  12. [14]

    Chidananda Gowda and G

    K. Chidananda Gowda and G. Krishna. 1978. Agglomerative clustering using the concept of mutual nearest neighbourhood.Pattern Recognition10, 2 (1978), 105–112. doi:10.1016/0031-3203(78)90018-3

    work page doi:10.1016/0031-3203(78)90018-3 1978

  13. [15]

    Derek Greene, Derek O’Callaghan, and Padraig Cunningham. 2014. How Many Topics? Stability Analysis for Topic Models. doi:10.1007/978-3-662-44848-9_32

    work page doi:10.1007/978-3-662-44848-9_32 2014

  14. [16]

    Dilek Hakkani-Tür, Yun-Cheng Ju, Geoffrey Zweig, and Gokhan Tur. 2015. Clus- tering novel intents in a conversational interaction system with semantic parsing. InInterspeech 2015. 1854–1858. doi:10.21437/Interspeech.2015-70

    work page doi:10.21437/interspeech.2015-70 2015

  15. [17]

    Iryna Haponchyk, Antonio Uva, Seunghak Yu, Olga Uryupina, and Alessandro Moschitti. 2018. Supervised Clustering of Questions into Intents for Dialog System Applications. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii (Eds.). Association for Comput...

    work page doi:10.18653/v1/d18-1254 2018

  16. [18]

    Sua Kim, Taeyoung Jeong, Seokyoung Hong, Seongjun Kim, Jeongpil Lee, Du- Seong Chang, and Myoung-Wan Koo. 2025. KSTC: Keyphrase-driven Sentence embedding and Task independent prompting for filling slot in the Generation of theme label. InProceedings of the Twelfth Dialog System Technology Challenge, Behnam Hedayatnia, Vivian Chen, Zhang Chen, Raghav Gupta...

    2025

  17. [19]

    Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K

    Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars. 2019. An Evaluation Dataset for Intent Clas- sification and Out-of-Scope Prediction. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing...

    2019

  18. [20]

    Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng

  19. [21]

    Jinggui Liang and Lizi Liao. 2023. ClusterPrompt: Cluster Semantic Enhanced Prompt Learning for New Intent Discovery. InFindings of the Association for Computational Linguistics: EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 10468–10481. doi:10.18653/v1/2023.findings-emnlp.702

    work page doi:10.18653/v1/2023.findings-emnlp.702 2023

  20. [22]

    I-Fan Lin, Faegheh Hasibi, and Suzan Verberne. 2025. SPILL: Domain-Adaptive Intent Clustering based on Selection and Pooling with Large Language Models. InFindings of the Association for Computational Linguistics: ACL 2025, Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (Eds.). Association for Computational Linguistics, Vienna...

    2025

  21. [23]

    Jianghan Liu, Ziyu Shang, Wenjun Ke, Peng Wang, Zhizhao Luo, Jiajun Liu, Guozheng Li, and Yining Li. 2025. LLM-Guided Semantic-Aware Clustering for Topic Modeling. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (E...

    doi:10.18653/v1/ 2025

  22. [24]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows . In2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, Los Alamitos, CA, USA, 9992–10002. doi:10.1109/ICCV48922.2021.00986

    arXiv 2021

  23. [25]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. InProceedings of the International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  24. [26]

    J. B. MacQueen. 1967. Some Methods for Classification and Analysis of Multi- Variate Observations. InProc. of the fifth Berkeley Symposium on Mathematical Statistics and Probability, L. M. Le Cam and J. Neyman (Eds.), Vol. 1. University of California Press, 281–297

    1967

  25. [27]

    Yutao Mou, Keqing He, Pei Wang, Yanan Wu, Jingang Wang, Wei Wu, and Weiran Xu. 2022. Watch the Neighbors: A Unified K-Nearest Neighbor Contrastive Learn- ing Framework for OOD Intent Discovery. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (Eds.). Association for C...

    work page doi:10.18653/v1/2022.emnlp-main.98 2022

  26. [28]

    Yutao Mou, Keqing He, Yanan Wu, Pei Wang, Jingang Wang, Wei Wu, Yi Huang, Junlan Feng, and Weiran Xu. 2022. Generalized Intent Discovery: Learning from Open World Dialogue System. InProceedings of the 29th International Conference on Computational Linguistics, Nicoletta Calzolari, Chu-Ren Huang, Hansaem Kim, James Pustejovsky, Leo Wanner, Key-Sun Choi, Pu...

    2022

  27. [29]

    A Padmasundari and Srinivas Bangalore. 2018. Intent Discovery Through Unsu- pervised Semantic Text Clustering. InInterspeech 2018. 606–610. doi:10.21437/ Interspeech.2018-2436

    2018

  28. [30]

    Anup Pattnaik, Cijo George, Rishabh Kumar Tripathi, Sasanka Vutla, and Jithen- dra Vepa. 2024. Improving Hierarchical Text Clustering with LLM-guided Multi- view Cluster Representation. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, Franck Dernoncourt, Daniel Preoţiuc-Pietro, and Anastasia Shimori...

    doi:10.18653/v1/2024.emnlp- 2024

  29. [31]

    Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. 2020. Towards Emotion-aided Multi-modal Dialogue Act Classification. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.). Association for MM ’26, November 10–14, 2026, Rio de Jane...

    doi:10.18653/v1/2020.acl-main.402 2020

  30. [32]

    Amit Saxena, Mukesh Prasad, Akshansh Gupta, Neha Bharill, Om Prakash Patel, Aruna Tiwari, Meng Joo Er, Weiping Ding, and Chin-Teng Lin. 2017. A Review of Clustering Techniques and Developments.Neurocomputing267 (2017), 664–681. doi:10.1016/j.neucom.2017.06.053

    doi:10.1016/j.neucom.2017.06.053 2017

  31. [33]

    Wenkai Shi, Wenbin An, Feng Tian, Qinghua Zheng, QianYing Wang, and Ping Chen. 2023. A Diffusion Weighted Graph Framework for New Intent Discovery. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 8033–8042. doi:...

    doi:10.18653/v1/2023.emnlp- 2023

  32. [34]

    Xiaoshuai Song, Yutao Mou, Keqing He, Yueyan Qiu, Jinxu Zhao, Pei Wang, and Weiran Xu. 2023. Continual Generalized Intent Discovery: Marching Towards Dynamic and Open-world Intent Recognition. InFindings of the Association for Computational Linguistics: EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics...

    doi:10.18653/v1/2023.findings-emnlp.289 2023

  33. [35]

    Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov

    Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2019. Multimodal Transformer for Unaligned Multimodal Language Sequences. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Anna Korhonen, David Traum, and Lluís Màrquez (Eds.). Association for Comput...

    doi:10.18653/v1/p19-1656 2019

  34. [36]

    Laurens J. P. van der Maaten and Geoffrey E. Hinton. 2008. Visualizing Data Using t-SNE.Journal of Machine Learning Research9 (2008), 2579–2605

    2008

  35. [37]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in neural information processing systems. 5998–6008. http://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2017

  36. [38]

    Vijay Viswanathan, Kiril Gashteovski, Carolin Lawrence, Tongshuang Wu, and Graham Neubig. 2024. Large Language Models Enable Few-Shot Clustering. Transactions of the Association for Computational Linguistics12 (2024), 321–333. doi:10.1162/tacl_a_00648

    doi:10.1162/tacl_a_00648 2024

  37. [39]

    Junyuan Xie, Ross Girshick, and Ali Farhadi. 2016. Unsupervised deep embedding for clustering analysis. InProceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48(New York, NY, USA) (ICML’16). JMLR.org, 478–487

    2016

  38. [40]

    Sidiropoulos, and Mingyi Hong

    Bo Yang, Xiao Fu, Nicholas D. Sidiropoulos, and Mingyi Hong. 2017. Towards K- means-friendly spaces: simultaneous deep learning and clustering. InProceedings of the 34th International Conference on Machine Learning - Volume 70(Sydney, NSW, Australia)(ICML’17). JMLR.org, 3861–3870

    2017

  39. [41]

    Arnold, and Bing Xiang

    Dejiao Zhang, Feng Nan, Xiaokai Wei, Shang-Wen Li, Henghui Zhu, Kathleen McKeown, Ramesh Nallapati, Andrew O. Arnold, and Bing Xiang. 2021. Support- ing Clustering with Contrastive Learning. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Kristina Toutanova, ...

    2021

  40. [42]

    Hanlei Zhang, Xiaoteng Li, Hua Xu, Panpan Zhang, Kang Zhao, and Kai Gao

  41. [43]

    Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, jinyue Zhao, Wenrui Li, and Yanting Chen. 2024. MIntRec2.0: A Large-scale Benchmark Dataset for Multimodal Intent Recognition and Out-of-scope Detection in Conversations. InThe Twelfth International Conference on Learning Representations. https: //openreview.net/forum?id=nY9nITZQjc

    2024

  42. [44]

    Hanlei Zhang, Hua Xu, Fei Long, Xin Wang, and Kai Gao. 2024. Unsupervised Multimodal Clustering for Semantics Discovery in Multimodal Utterances. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics...

    work page doi:10.18653/v1/2024.acl-long.2 2024

  43. [45]

    TEXTOIR: An Integrated and Visualized Platform for Text Open Intent Recognition. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, Heng Ji, Jong C. Park, and Rui Xia (Eds.). Association for Computational Linguistics, O...

    2021

  44. [46]

    Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng

  45. [47]

    Yuwei Zhang, Zihan Wang, and Jingbo Shang. 2023. ClusterLLM: Large Language Models as a Guide for Text Clustering. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 13903–13920. doi:10.18653/v1/2023.emnlp-main.858

    doi:10.18653/v1/2023.emnlp-main.858 2023

  46. [48]

    Hanlei Zhang, Hua Xu, Xin Wang, Fei Long, and Kai Gao. 2024. A Clustering Framework for Unsupervised and Semi-Supervised New Intent Discovery.IEEE Transactions on Knowledge and Data Engineering36, 11 (2024), 5468–5481. doi:10. 1109/TKDE.2023.3340732

    arXiv 2024

  47. [49]

    Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. 2024. Token-Level Contrastive Learning with Modality-Aware Prompting for Multimodal Intent Recognition. InProceedings of the 37th AAAI Conference on Artificial Intelligence. AAAI Press, 17114–17122. doi:10.1609/aaai. v38i15.29656

    doi:10.1609/aaai 2024

  48. [50]

    Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. 2003. Semi-supervised learning using Gaussian fields and harmonic functions. InProceedings of the Twentieth International Conference on International Conference on Machine Learn- ing(Washington, DC, USA)(ICML’03). AAAI Press, 912–919

    2003

  49. [51]

    requesting help

    Zhihong Zhu, Xuxin Cheng, Zhaorun Chen, Yuyan Chen, Yunyan Zhang, Xian Wu, Yefeng Zheng, and Bowen Xing. 2024. InMu-Net: Advancing Multi-modal Intent Detection via Information Bottleneck and Multi-sensory Processing. In Proceedings of the 32nd ACM International Conference on Multimedia(Melbourne VIC, Australia)(MM ’24). Association for Computing Machinery...

    arXiv 2024

  50. [52]

    Yuwei Zhang, Haode Zhang, Li-Ming Zhan, Xiao-Ming Wu, and Albert Y.S. Lam

  51. [53]

    https://api.semanticscholar.org/CorpusID:248780207

    New Intent Discovery with Pre-training and Contrastive Learning.ArXiv abs/2205.12914 (2022). https://api.semanticscholar.org/CorpusID:248780207

    Pith/arXiv arXiv 2022

  52. [2021]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Contrastive Clustering. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. AAAI Press, 8547–8555. doi:10.1609/aaai.v35i10.17037

    work page doi:10.1609/aaai.v35i10.17037

  53. [2022]

    InProceedings of the 30th ACM International Conference on Multimedia(Lisboa, Portugal)(MM ’22)

    MIntRec: A New Dataset for Multimodal Intent Recognition. InProceedings of the 30th ACM International Conference on Multimedia(Lisboa, Portugal)(MM ’22). Association for Computing Machinery, New York, NY, USA, 1688–1697. doi:10.1145/3503161.3547906

    arXiv

  54. [2024]

    InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.)

    Dual-oriented Disentangled Network with Counterfactual Intervention for Multimodal Intent Detection. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 17554–17567. doi:10.18653/v1/2024.emnlp-main.972

    work page doi:10.18653/v1/2024.emnlp-main.972 2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.