Pith. sign in

REVIEW 3 major objections 4 minor 85 references

The paper's central claim is that LLM-generated chain-of-thought rationales, distilled into a compact mBERT student through dual-path representation and response alignment, consistently outperform strong baselines in cross-lingual stance de

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 14:37 UTC pith:CMSFJUPQ

load-bearing objection A competent, clearly-written extension of CoT distillation to cross-lingual stance detection with modest gains, but the ablation does not isolate the reasoning content and the statistics need tightening before publication. the 3 major comments →

arxiv 2607.18693 v1 pith:CMSFJUPQ submitted 2026-07-21 cs.CL

Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection

Qiuli Zhou , Jingyuan Yao , Shengeng Tang , Hongzhi Chen , Jun Tang , Richang Hong This is my paper
classification cs.CL
keywords cross-lingual stance detectionknowledge distillationchain-of-thought reasoningrationale generationin-context learningcontrastive learningmultilingual BERTlow-resource NLP
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.

Stance detection in low-resource languages like Catalan is held back by scarce training data. The paper proposes a framework in which a frozen large language model, prompted with chain-of-thought examples, writes short rationales explaining each stance, and those rationales are then distilled into a compact multilingual BERT model. During training, the student runs two branches that share the same weights—one that sees the rationale and one that does not—and the two branches are aligned on both hidden representations and prediction distributions. At inference only the rationale-free branch is used, so the model keeps the LLM's reasoning benefit without the LLM's cost. The authors report consistent gains over competitive baselines across three multilingual benchmarks, with the largest F1 improvement around 3.9 points.

Core claim

The central claim is that explicit chain-of-thought rationales generated by a frozen LLM can serve as transferable reasoning supervision for cross-lingual stance detection, and that a dual-branch distillation—aligning [CLS] representations via an InfoNCE contrastive objective and aligning prediction distributions via symmetric KL divergence between a rationale-enhanced branch and a rationale-free branch of the same mBERT student—lets the compact model internalize this reasoning while staying rationale-free at test time. The paper further claims that two contrastive strategies, a prototype-based contrastive loss in the source language and a dual-constraint contrastive alignment in the target

What carries the argument

Dual-path hierarchical distillation: the mBERT student is run twice on each training sample—once on (target, text, LLM-generated rationale) and once on (target, text) only—using the same shared encoder weights. InfoNCE pulls the two [CLS] embeddings of the same sample together while pushing apart different samples, and symmetric KL aligns the soft predictions of the two paths, so at deployment only the rationale-free path is needed. The rationales are produced by a frozen LLM using eight in-context chain-of-thought demonstrations, and a prototype-based contrastive loss on the source language provides additional stance discrimination.

Load-bearing premise

The load-bearing premise is that the dual-path alignment actually transfers the LLM's reasoning to the student: since both branches share the same mBERT weights, the InfoNCE and KL losses could simply teach the model to ignore the rationale input while gains come from contrastive regularization or extra tokens, and the paper provides no probing or control experiment showing the student acquired reasoning ability—it also concedes it still needs roughly 500 labeled target sampl

What would settle it

A control experiment that replaces the LLM rationale in the teacher branch with an equal-length, non-reasoning filler (such as shuffled rationale words or a generic template) while keeping everything else fixed: if accuracy and F1 stay at the same level, the reasoning content is not the active ingredient. Alternatively, probe the trained mBERT by asking it to produce rationales or by testing whether its rationale-free [CLS] embeddings encode reasoning structure; the absence of such signal would falsify the claim that reasoning was distilled.

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

If this is right

  • Low-resource languages can receive reasoning-aware stance detectors trained on roughly 500 labeled target samples, with no LLM inference required at deployment, keeping latency and cost low.
  • Because the student model is rationale-free at test time, the approach is practical for real-time applications where directly running a large language model is infeasible.
  • Higher-quality reasoning supervision matters: thinking-mode chain-of-thought rationales outperform non-thinking ones, so improving rationale quality should directly improve downstream accuracy.
  • The rationale path is load-bearing; removing it costs about 1.0–1.5 points in accuracy and F1, and moderate contrastive weights are optimal, indicating the two contrastive terms contribute beyond label supervision.
  • The method transfers across topics and domains (politics, Catalan independence, vaccines), suggesting the distilled reasoning knowledge is target-agnostic rather than topic-specific.

Where Pith is reading between the lines

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

  • The paper does not rule out that the gains come from the contrastive regularization or from seeing extra input tokens rather than from the reasoning content itself, since both branches share the same mBERT weights; a control that replaces rationales with non-reasoning filler text would settle this.
  • A direct test of whether reasoning was truly internalized would be to probe the trained student—for example, by prompting it to generate rationales or by examining whether its rationale-free embeddings encode reasoning structure such as clustering by rationale type.
  • The same dual-branch distillation recipe could transfer to other low-resource classification tasks where an LLM can produce evidence, such as fact-checking, hate-speech detection, or sentiment analysis.
  • The framework could be pushed toward fully zero-shot target languages by generating rationales for unlabeled target text and using them as self-supervision; the paper itself identifies zero-shot settings as future work.

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 a rationale-guided knowledge distillation framework for cross-lingual stance detection. Chain-of-thought prompting with in-context learning is used to generate rationales from a frozen LLM (Qwen3.5-Flash); these rationales are then used as additional input to a rationale-enhanced mBERT branch during target-language training, while a rationale-free mBERT branch is aligned to it through representation-level (InfoNCE) and response-level (KL) distillation. A prototype-based contrastive loss is used in the source-language stage, and a sample-wise contrastive alignment is used in the target stage. The method is evaluated on X-stance, CIC, and VaxxStance against several baselines, with ablations on rationale removal, number of ICL demonstrations, support/query construction, and contrastive loss weights. The central claim is that distilling LLM-generated reasoning into mBERT consistently outperforms competitive baselines, with gains of roughly 1–3% over CCSD.

Significance. If the claimed effect is real, the paper would make a useful contribution: it proposes a practical way to inject LLM-style reasoning into a compact encoder model for low-resource cross-lingual stance detection, and it addresses an architecture mismatch between decoder-only teachers and encoder-only students. The paper is clearly written, includes a reasonable set of baselines, and provides multiple ablation and sensitivity analyses. The idea of using rationales as auxiliary supervision is well-motivated and timely. However, the current experimental design does not isolate the effect of rationale content from the effect of simply adding extra input tokens and applying contrastive/consistency regularization, and the statistical evidence for the headline improvements is thin. These are fixable with additional experiments, so the manuscript is not fatally flawed, but the central claim needs stronger support before publication.

major comments (3)
  1. [§4.4, Eqs. (13)–(16), Fig. 5] The rationale ablation removes the entire reasoning path, which simultaneously removes (a) the second input branch, (b) the extra rationale tokens, and (c) the InfoNCE and KL alignment terms. Therefore the observed performance drop cannot be attributed specifically to the reasoning content of the rationales. Because both branches share the same trainable mBERT weights, the distillation objective in Eqs. (13)–(14) is essentially a consistency regularizer between two views of the same model; the teacher branch differs from the student branch only by the additional rationale tokens. A control that feeds length-matched non-reasoning filler tokens (e.g., shuffled or unrelated sentences) into the teacher branch would isolate the effect of rationale content. Without such a control, the paper's distinctive claim—that reasoning knowledge, not merely auxiliary regularization, is transferred—is not
  2. [Table 1] All results are reported as single-run accuracy and macro-F1 without standard deviations, confidence intervals, or significance tests. Several of the headline gains are small: on VaxxStance, the Qwen-thinking variant improves F1 over CCSD by only +0.12, and on X-Policy the accuracy gain is +1.09. Given this magnitude, the claim of 'consistent outperformance' across all datasets needs multiple seeds and a paired significance test (e.g., bootstrap or approximate randomization) to rule out variance. This is load-bearing because the central conclusion rests on the aggregate pattern of small improvements.
  3. [§4.5–§4.7, §4.10] The hyperparameters α=0.03, β=0.05, 8 ICL demonstrations, and the support/query configuration 8×2+4 are all tuned on X-Society and then applied unchanged to X-Economy, X-Education, X-Policy, CIC, and VaxxStance. The paper states that these settings give the 'best trade-off' on X-Society but provides no evidence that the same setting is near-optimal on other datasets and topics. If, for example, β or the ICL count were re-tuned per dataset, some of the reported gains might shrink or change sign. The authors should either report per-dataset selection or show sensitivity on at least one additional target language/dataset to support the generalization claim.
minor comments (4)
  1. [Table 1] Formatting issue: the CIC and VaxxStance columns run together (e.g., '70.1477.82' and '75.8377.86' are likely missing spaces). Please ensure all numbers are clearly separated.
  2. [Eqs. (6) and (14)] The symbol L_c is used for two different losses: the prototype-based contrastive loss in Eq. (6) and the InfoNCE loss in Eq. (14). Using distinct symbols (e.g., L_proto and L_infonce) would avoid confusion, especially because both appear in the overall objective.
  3. [§4.9] The claim that Qwen thinking mode outperforms non-thinking mode is based on a comparison visible in Table 1 and Fig. 5, but the manual qualitative inspection described in §4.9 is not reported in detail. A short table or example rationales from both modes would make the argument more concrete.
  4. [§4.8 and Fig. 7] The Integrated Gradients analysis would benefit from a brief description of how attributions were aggregated across tokens and whether the visualized examples are representative. Currently the two examples are illustrative but not quantified.

Circularity Check

0 steps flagged

No significant circularity; central result is a held-out empirical comparison, and the only self-citation is not load-bearing.

full rationale

The paper's central claim is empirical: it trains mBERT with a dual-path objective on labeled source/target data and evaluates on held-out test sets. LLM rationales are generated offline from training pairs (with the ground-truth stance included in the prompt) and are not used at test time, so there is no test-set leakage or fitted-parameter-then-prediction structure. Equations 13-18 define a consistency/contrastive objective between a rationale-conditioned branch and a rationale-free branch of the same mBERT; sharing weights means the 'teacher' is a self-distillation view rather than an independent LLM teacher, but this is a design choice and does not make the reported gain true by definition. The only self-citation, [14], appears in a broad citation list about LLM reasoning and is not used to justify the dual-branch design or to exclude alternatives. The §4.4 ablation removes the entire reasoning path, so the drop cannot be cleanly attributed to rationale content as opposed to the extra input/regularization; this is a confound in empirical attribution, not a circularity. Accordingly, no circular step is identified; score 2 reflects only the presence of a minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper makes no mathematical derivation; it builds a training objective from standard losses. The main burden is empirical: the framework's effectiveness rests on several domain assumptions about rationale quality and transfer, plus hyperparameters tuned on a test set.

free parameters (4)
  • α (source contrastive loss weight) = 0.03
    Tuned on X-Society test set (Fig. 6a); value chosen by maximizing accuracy on the same benchmark used for the main reported result.
  • β (target contrastive distillation weight) = 0.05
    Tuned on X-Society test set (Fig. 6b); best value selected on the test set.
  • Number of ICL demonstrations = 8
    Selected by comparing 2,4,6,8,10 in-context examples on X-Society/X-Economy/X-Education/X-Policy (Table 2); 8 chosen as best accuracy/compute trade-off.
  • Support/query split (support per class / query) = 8 support per class / 4 query
    Chosen empirically among 4/1, 4/2, 4/4, 8/4, 10/4 configurations (Table 3).
axioms (4)
  • domain assumption mBERT is a sufficient student architecture for cross-lingual stance detection.
    The framework is built on mBERT ([22]); if mBERT lacked multilingual capacity, the transfer would fail.
  • domain assumption LLM-generated rationales contain reasoning information that can be transferred to a compact encoder.
    This is the core premise of the paper (§3.3); no direct evidence that the student learns reasoning is provided.
  • domain assumption The hosted Qwen3.5-Flash API produces stable, reproducible rationales across calls.
    Rationales are generated via an API (§4.10); API versions and non-determinism are not controlled.
  • ad hoc to paper Hyperparameters tuned on X-Society generalize to other topics and datasets.
    α, β, ICL count, and support/query ratio are tuned on X-Society (§4.5–4.7) and then used for CIC and VaxxStance without re-tuning.

pith-pipeline@v1.3.0-alltime-deepseek · 20340 in / 16569 out tokens · 139945 ms · 2026-08-01T14:37:12.363047+00:00 · methodology

0 comments
read the original abstract

Stance detection aims to identify whether a text expresses a favorable or opposing attitude toward a given target, and serves as an important task for various downstream applications. Although existing studies have achieved strong performance in monolingual settings, especially in English, many low-resource languages such as Catalan still lack sufficient annotated data for training effective models. Cross-lingual stance detection alleviates this problem by transferring stance knowledge from resource-rich languages to low-resource languages. However, most existing methods mainly rely on semantic alignment between texts and targets, while ignoring the reasoning process required for reliable stance inference. Although Large Language Models provide strong reasoning ability, their high computational cost and inference latency limit practical deployment. To address these limitations, we propose a rationale-guided knowledge distillation framework for cross-lingual stance detection. Specifically, we use Chain-of-Thought prompting to guide Large Language Models in generating informative rationales, and distill the resulting reasoning knowledge into a compact student model. We further design a dual-path distillation mechanism to align rationale-enhanced and rationale-free representations, together with their prediction distributions. In addition, two contrastive learning strategies are introduced to improve stance discrimination. Experiments on multilingual benchmarks demonstrate that our method consistently outperforms competitive baselines.

Figures

Figures reproduced from arXiv: 2607.18693 by Hongzhi Chen, Jingyuan Yao, Jun Tang, Qiuli Zhou, Richang Hong, Shengeng Tang.

Figure 1
Figure 1. Figure 1: Illustrative example of Chain-of-Thought (CoT) reasoning for stance detection in English. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representation and structural gaps in knowledge distillation between decoder-only Large Language Models and encoder-only [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The pipeline of reasoning rationale preparation via In-Context Learning (ICL). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall architecture of the proposed framework. Step 1 trains the source-language mBERT model with cross-lingual templates [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effectiveness of the LLM reasoning on four X-stance topics. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of the contrastive loss weight on model performance. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison between the baseline model and our method using Integrated Gradients (IG). Red indicates positive [PITH_FULL_IMAGE:figures/full_fig_p017_7.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

85 extracted references · 9 linked inside Pith

  1. [1]

    Stance detection on social media: State of the art and trends.Information Processing & Management, 58(4):102597, 2021

    Abeer AlDayel and Walid Magdy. Stance detection on social media: State of the art and trends.Information Processing & Management, 58(4):102597, 2021

    2021

  2. [2]

    Towards analyzing the bias of news recommender systems using sentiment and stance detection

    Mehwish Alam, Andreea Iana, Alexander Grote, Katharina Ludwig, Philipp Müller, and Heiko Paulheim. Towards analyzing the bias of news recommender systems using sentiment and stance detection. InCompanion proceedings of the web conference 2022, pages 448–457, 2022

    2022

  3. [3]

    A survey on knowledge-aware news recommender systems.Semantic Web, 15(1):21–82, 2024

    Andreea Iana, Mehwish Alam, and Heiko Paulheim. A survey on knowledge-aware news recommender systems.Semantic Web, 15(1):21–82, 2024

    2024

  4. [4]

    Cross-domain label-adaptive stance detection

    Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. Cross-domain label-adaptive stance detection. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 9011–9028, 2021

    2021

  5. [5]

    Stance detection: A survey.ACM Computing Surveys (CSUR), 53(1):1–37, 2020

    Dilek Küçük and Fazli Can. Stance detection: A survey.ACM Computing Surveys (CSUR), 53(1):1–37, 2020

    2020

  6. [6]

    Zero-shot stance detection: A dataset and model using generalized topic representations

    Emily Allaway and Kathleen McKeown. Zero-shot stance detection: A dataset and model using generalized topic representations. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913–8931, 2020

    2020

  7. [7]

    Few-shot cross-lingual stance detection with sentiment-based pre-training

    Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. Few-shot cross-lingual stance detection with sentiment-based pre-training. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10729–10737, 2022

    2022

  8. [8]

    Cross-lingual cross-target stance detection with dual knowledge distillation framework

    Ruike Zhang, Hanxuan Yang, and Wenji Mao. Cross-lingual cross-target stance detection with dual knowledge distillation framework. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 10804–10819, 2023

    2023

  9. [9]

    Target-oriented relation alignment for cross-lingual stance detection

    Ruike Zhang, Nan Xu, Hanxuan Yang, Yuan Tian, and Wenji Mao. Target-oriented relation alignment for cross-lingual stance detection. InFindings of the Association for Computational Linguistics: ACL 2023, pages 6391–6404, 2023

    2023

  10. [10]

    An llm-enabled knowledge elicitation and retrieval framework for zero-shot cross-lingual stance identification

    Ruike Zhang, Yuan Tian, Penghui Wei, Daniel Dajun Zeng, and Wenji Mao. An llm-enabled knowledge elicitation and retrieval framework for zero-shot cross-lingual stance identification. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 12253–12266, 2024

    2024

  11. [11]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  12. [12]

    Teaching small language models to reason

    Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. Teaching small language models to reason. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1773–1781, 2023

    2023

  13. [13]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Hierarchical knowledge distillation for cross-lingual stance detection

    Qiuli Zhou, Jingyuan Yao, Shengeng Tang, Weidong Chen, Lechao Cheng, and Jun Tang. Hierarchical knowledge distillation for cross-lingual stance detection. In2025 4th International Conference on Artificial Intelligence, Human-Computer Interaction and Robotics (AIHCIR), pages 1–5. IEEE, 2025

    2025

  15. [15]

    Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. InFindings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017, 2023

    2023

  16. [16]

    Symbolic chain-of-thought distillation: Small models can also “think” step-by-step

    Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of-thought distillation: Small models can also “think” step-by-step. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2665–2679, 2023

    2023

  17. [17]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 1107–1128, 2024

    2024

  18. [18]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work? InProceedings of the 2022 conference on empirical methods in natural language processing, pages 11048–11064, 2022

    2022

  19. [19]

    From reasoning llms to bert: A two-stage distillation framework for search relevance.arXiv preprint arXiv:2510.11056, 2025

    Runze Xia, Yupeng Ji, Yuxi Zhou, Haodong Liu, Teng Zhang, and Piji Li. From reasoning llms to bert: A two-stage distillation framework for search relevance.arXiv preprint arXiv:2510.11056, 2025

    arXiv 2025

  20. [20]

    Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  21. [21]

    Knowledge distillation: A survey.International journal of computer vision, 129(6): 1789–1819, 2021

    Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey.International journal of computer vision, 129(6): 1789–1819, 2021

    2021

  22. [22]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019

    2019

  23. [23]

    How multilingual is multilingual bert? InProceedings of the 57th annual meeting of the association for computational linguistics, pages 4996–5001, 2019

    Telmo Pires, Eva Schlinger, and Dan Garrette. How multilingual is multilingual bert? InProceedings of the 57th annual meeting of the association for computational linguistics, pages 4996–5001, 2019

    2019

  24. [24]

    Improving efficient neural ranking models with cross-architecture knowledge distillation.arXiv preprint arXiv:2010.02666, 2020

    Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. Improving efficient neural ranking models with cross-architecture knowledge distillation.arXiv preprint arXiv:2010.02666, 2020

    Pith/arXiv arXiv 2010

  25. [25]

    Simcse: Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 6894–6910, 2021

    2021

  26. [26]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PmLR, 2020. Manuscript submitted to ACM Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection 21

    2020

  27. [27]

    Text style transfer with contrastive transfer pattern mining

    Jingxuan Han, Quan Wang, Licheng Zhang, Weidong Chen, Yan Song, and Zhendong Mao. Text style transfer with contrastive transfer pattern mining. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7914–7927, 2023

    2023

  28. [28]

    Contrastive representation distillation.arXiv preprint arXiv:1910.10699, 2019

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation.arXiv preprint arXiv:1910.10699, 2019

    Pith/arXiv arXiv 1910

  29. [29]

    A dataset for multi-target stance detection

    Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. A dataset for multi-target stance detection. InProceedings of the 15th conference of the European chapter of the association for computational linguistics: volume 2, short papers, pages 551–557, 2017

    2017

  30. [30]

    Semeval-2016 task 6: Detecting stance in tweets

    Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. Semeval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th international workshop on semantic evaluation (SemEval-2016), pages 31–41, 2016

    2016

  31. [31]

    Stance detection in covid-19 tweets

    Kyle Glandt, Sarthak Khanal, Yingjie Li, Doina Caragea, and Cornelia Caragea. Stance detection in covid-19 tweets. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1596–1611, 2021

    2021

  32. [32]

    Will-they-won’t-they: A very large dataset for stance detection on twitter

    Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. Will-they-won’t-they: A very large dataset for stance detection on twitter. InProceedings of the 58th annual meeting of the association for computational linguistics, pages 1715–1724, 2020

    2020

  33. [33]

    Stance classification with target-specific neural attention networks

    Jiachen Du, Ruifeng Xu, Yulan He, and Lin Gui. Stance classification with target-specific neural attention networks. In26th International Joint Conference on Artificial Intelligence, IJCAI 2017, pages 3988–3994. International Joint Conferences on Artificial Intelligence, 2017

    2017

  34. [34]

    Connecting targets to tweets: Semantic attention-based model for target-specific stance detection

    Yiwei Zhou, Alexandra I Cristea, and Lei Shi. Connecting targets to tweets: Semantic attention-based model for target-specific stance detection. In International Conference on Web Information Systems Engineering, pages 18–32. Springer, 2017

    2017

  35. [35]

    Fengyi Fu, Shancheng Fang, Weidong Chen, and Zhendong Mao. Sentiment-oriented transformer-based variational autoencoder network for live video commenting.ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1–24, 2024

    2024

  36. [36]

    Kaiwen Wei, Li Jin, Zequn Zhang, Zhi Guo, Xiaoyu Li, Qing Liu, and Weimiao Feng. More than syntaxes: Investigating semantics to zero-shot cross-lingual relation extraction and event argument role labelling.ACM Transactions on Asian and Low-Resource Language Information Processing, 23(5):1–21, 2024

    2024

  37. [37]

    Cascade cross-modal attention network for video actor and action segmentation from a sentence

    Weidong Chen, Guorong Li, Xinfeng Zhang, Hongyang Yu, Shuhui Wang, and Qingming Huang. Cascade cross-modal attention network for video actor and action segmentation from a sentence. InProceedings of the 29th ACM International Conference on Multimedia, pages 4053–4062, 2021

    2021

  38. [38]

    Improving radiology report generation with multi-grained abnormality prediction.Neurocomputing, 600:128122, 2024

    Yuda Jin, Weidong Chen, Yuanhe Tian, Yan Song, and Chenggang Yan. Improving radiology report generation with multi-grained abnormality prediction.Neurocomputing, 600:128122, 2024

    2024

  39. [39]

    End-to-end aspect-based sentiment analysis with combinatory categorial grammar

    Yuanhe Tian, Weidong Chen, Bo Hu, Yan Song, and Fei Xia. End-to-end aspect-based sentiment analysis with combinatory categorial grammar. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13597–13609, 2023

    2023

  40. [40]

    Improving video summarization by exploring the coherence between corresponding captions.IEEE Transactions on Image Processing, 2025

    Cheng Ye, Weidong Chen, Bo Hu, Lei Zhang, Yongdong Zhang, and Zhendong Mao. Improving video summarization by exploring the coherence between corresponding captions.IEEE Transactions on Image Processing, 2025

    2025

  41. [41]

    Multi-round mutual emotion-cause pair extraction for emotion-attributed video captioning

    Cheng Ye, Weidong Chen, Peipei Song, Xinyan Liu, Lei Zhang, and Zhendong Mao. Multi-round mutual emotion-cause pair extraction for emotion-attributed video captioning. InProceedings of the 33rd ACM International Conference on Multimedia, pages 3320–3329, 2025

    2025

  42. [42]

    Stance detection with a multi-target adversarial attention network.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(2):1–21, 2022

    Qingying Sun, Xuefeng Xi, Jiajun Sun, Zhongqing Wang, and Huiyan Xu. Stance detection with a multi-target adversarial attention network.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(2):1–21, 2022

    2022

  43. [43]

    Combatting data imbalance and noise in micro-action recognition

    Chuang Wang, Weidong Chen, Xu Cui, Yiming Zhao, Zhaobo Qi, Pengqi Huang, Xinyan Liu, and Weigang Zhang. Combatting data imbalance and noise in micro-action recognition. InProceedings of the 33rd ACM International Conference on Multimedia, pages 14229–14235, 2025

    2025

  44. [44]

    Matching street view and satellite images via drone imagery and semantic descriptions

    Xinyan Liu, Weidong Chen, Zhaobo Qi, Beichen Zhang, and Weigang Zhang. Matching street view and satellite images via drone imagery and semantic descriptions. InProceedings of the 3rd International Workshop on UA Vs in Multimedia: Capturing the World from a New Perspective, pages 4–9, 2025

    2025

  45. [45]

    Difference-aware iterative reasoning network for key relation detection

    Bowen Zhao, Weidong Chen, Bo Hu, Hongtao Xie, and Zhendong Mao. Difference-aware iterative reasoning network for key relation detection. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 276–281. IEEE Computer Society, 2023

    2023

  46. [46]

    Zero-shot cross-lingual stance detection via adversarial language adaptation.PeerJ Computer Science, 11:e2955, 2025

    A Bharathi and Arkaitz Zubiaga. Zero-shot cross-lingual stance detection via adversarial language adaptation.PeerJ Computer Science, 11:e2955, 2025

    2025

  47. [47]

    A reinforcement learning framework for cross-lingual stance detection using chain-of-thought alignment

    Binghui Li, Minghui Zou, Xiaowang Zhang, Shizhan Chen, and Zhiyong Feng. A reinforcement learning framework for cross-lingual stance detection using chain-of-thought alignment. InFindings of the Association for Computational Linguistics: ACL 2025, pages 21674–21688, 2025

    2025

  48. [48]

    Knowledge-enhanced prompt-tuning for stance detection.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(6):1–20, 2023

    Hu Huang, Bowen Zhang, Yangyang Li, Baoquan Zhang, Yuxi Sun, Chuyao Luo, and Cheng Peng. Knowledge-enhanced prompt-tuning for stance detection.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(6):1–20, 2023

    2023

  49. [49]

    Yuxin Huang, Yin Liang, Zhaoyuan Wu, Enchang Zhu, and Zhengtao Yu. Cross-lingual sentence embedding for low-resource chinese-vietnamese based on contrastive learning.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(6):1–18, 2023

    2023

  50. [50]

    Jiushun Ma, Yuxin Huang, Linqin Wang, Xiang Huang, Hao Peng, Zhengtao Yu, and Philip Yu. Augmenting low-resource cross-lingual summarization with progression-grounded training and prompting.ACM Transactions on Asian and Low-Resource Language Information Processing, 23(9):1–22, 2024

    2024

  51. [51]

    Towards efficient partially relevant video retrieval with active moment discovering.IEEE Transactions on Multimedia, 2025

    Peipei Song, Long Zhang, Long Lan, Weidong Chen, Dan Guo, Xun Yang, and Meng Wang. Towards efficient partially relevant video retrieval with active moment discovering.IEEE Transactions on Multimedia, 2025

    2025

  52. [52]

    Creatiparser: Generative image parsing of raster graphic designs into editable layers.arXiv preprint arXiv:2604.19632, 2026

    Weidong Chen, Dexiang Hong, Zhendong Mao, Yutao Cheng, Xinyan Liu, Lei Zhang, and Yongdong Zhang. Creatiparser: Generative image parsing of raster graphic designs into editable layers.arXiv preprint arXiv:2604.19632, 2026. Manuscript submitted to ACM 22 Zhou et al

    Pith/arXiv arXiv 2026

  53. [53]

    Face-net: Factual calibration and emotion augmentation for retrieval-enhanced emotional video captioning.arXiv preprint arXiv:2603.17455, 2026

    Weidong Chen, Cheng Ye, Zhendong Mao, Peipei Song, Xinyan Liu, Lei Zhang, Xiaojun Chang, and Yongdong Zhang. Face-net: Factual calibration and emotion augmentation for retrieval-enhanced emotional video captioning.arXiv preprint arXiv:2603.17455, 2026

    arXiv 2026

  54. [54]

    Improving image captioning via predicting structured concepts

    Ting Wang, Weidong Chen, Yuanhe Tian, Yan Song, and Zhendong Mao. Improving image captioning via predicting structured concepts. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 360–370, 2023

    2023

  55. [55]

    Weidong Chen, Guorong Li, Xinfeng Zhang, Shuhui Wang, Liang Li, and Qingming Huang. Weakly supervised text-based actor-action video segmentation by clip-level multi-instance learning.ACM Transactions on Multimedia Computing, Communications and Applications, 19(1):1–22, 2023

    2023

  56. [56]

    Rethinking pseudo word learning in zero-shot composed image retrieval: From an object-aware perspective

    Zhe Li, Lei Zhang, Kun Zhang, Weidong Chen, Yongdong Zhang, and Zhendong Mao. Rethinking pseudo word learning in zero-shot composed image retrieval: From an object-aware perspective. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 833–843, 2025

    2025

  57. [57]

    Dual-path collaborative generation network for emotional video captioning

    Cheng Ye, Weidong Chen, Jingyu Li, Lei Zhang, and Zhendong Mao. Dual-path collaborative generation network for emotional video captioning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 496–505, 2024

    2024

  58. [58]

    Heejin Do and Gary Geunbae Lee. Target-oriented knowledge distillation with language-family-based grouping for multilingual nmt.ACM Transactions on Asian and Low-Resource Language Information Processing, 22(2):1–18, 2023

    2023

  59. [59]

    Berttocnn: Similarity-preserving enhanced knowledge distillation for stance detection.Plos one, 16(9):e0257130, 2021

    Yang Li, Yuqing Sun, and Nana Zhu. Berttocnn: Similarity-preserving enhanced knowledge distillation for stance detection.Plos one, 16(9):e0257130, 2021

    2021

  60. [60]

    Contour-augmented concept prediction network for image captioning

    Ting Wang, Weidong Chen, Jingyu Li, Yixing Peng, and Zhendong Mao. Contour-augmented concept prediction network for image captioning. In International Conference on Artificial Neural Networks, pages 180–191. Springer, 2023

    2023

  61. [61]

    Subjective-objective emotion correlated generation network for subjective video captioning.IEEE Transactions on Image Processing, 2026

    Weidong Chen, Cheng Ye, Peipei Song, Lei Zhang, Yongdong Zhang, and Zhendong Mao. Subjective-objective emotion correlated generation network for subjective video captioning.IEEE Transactions on Image Processing, 2026

    2026

  62. [62]

    Query-based collaborative multimodal token pruning for audio-visual question answering

    Xilin Qin, Dexiang Hong, Weidong Chen, Cheng Ye, Xinyan Liu, Peipei Song, and Lei Zhang. Query-based collaborative multimodal token pruning for audio-visual question answering. In2025 4th International Conference on Artificial Intelligence, Human-Computer Interaction and Robotics (AIHCIR), pages 1–6. IEEE, 2025

    2025

  63. [63]

    Emoverse: A mllms-driven emotion representation dataset for interpretable visual emotion analysis.arXiv preprint arXiv:2511.12554, 2025

    Yijie Guo, Dexiang Hong, Weidong Chen, Zihan She, Cheng Ye, Xiaojun Chang, and Zhendong Mao. Emoverse: A mllms-driven emotion representation dataset for interpretable visual emotion analysis.arXiv preprint arXiv:2511.12554, 2025

    Pith/arXiv arXiv 2025

  64. [64]

    Zero-shot stance detection in practice: insights on training, prompting, and decoding with a capable lightweight llm.PeerJ Computer Science, 12:e3540, 2026

    Rachith Aiyappa, Shruthi Senthilmani, Jisun An, Haewoon Kwak, and Yong-Yeol Ahn. Zero-shot stance detection in practice: insights on training, prompting, and decoding with a capable lightweight llm.PeerJ Computer Science, 12:e3540, 2026

    2026

  65. [65]

    Stance reasoner: Zero-shot stance detection on social media with explicit reasoning

    Maksym Taranukhin, Vered Shwartz, and Evangelos Milios. Stance reasoner: Zero-shot stance detection on social media with explicit reasoning. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15257–15272, 2024

    2024

  66. [66]

    Stance detection with collaborative role-infused llm-based agents

    Xiaochong Lan, Chen Gao, Depeng Jin, and Yong Li. Stance detection with collaborative role-infused llm-based agents. InProceedings of the international AAAI conference on web and social media, volume 18, pages 891–903, 2024

    2024

  67. [67]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  68. [68]

    An explanation of in-context learning as implicit bayesian inference.arXiv preprint arXiv:2111.02080, 2021

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference.arXiv preprint arXiv:2111.02080, 2021

    Pith/arXiv arXiv 2021

  69. [69]

    Bootstrapping large language models for radiology report generation

    Chang Liu, Yuanhe Tian, Weidong Chen, Yan Song, and Yongdong Zhang. Bootstrapping large language models for radiology report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18635–18643, 2024

    2024

  70. [70]

    Multi-attention network for compressed video referring object segmentation

    Weidong Chen, Dexiang Hong, Yuankai Qi, Zhenjun Han, Shuhui Wang, Laiyun Qing, Qingming Huang, and Guorong Li. Multi-attention network for compressed video referring object segmentation. InProceedings of the 30th ACM International Conference on Multimedia, pages 4416–4425, 2022

    2022

  71. [71]

    Graph mixture of experts and memory-augmented routers for multivariate time series anomaly detection

    Xiaoyu Huang, Weidong Chen, Bo Hu, and Zhendong Mao. Graph mixture of experts and memory-augmented routers for multivariate time series anomaly detection. InProceedings of the AAAI conference on artificial intelligence, volume 39, pages 17476–17484, 2025

    2025

  72. [72]

    Prompting few-shot multi-hop question generation via comprehending type-aware semantics

    Zefeng Lin, Weidong Chen, Yan Song, and Yongdong Zhang. Prompting few-shot multi-hop question generation via comprehending type-aware semantics. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 3730–3740, 2024

    2024

  73. [73]

    Creatidesign: A unified multi-conditional diffusion transformer for creative graphic design.arXiv preprint arXiv:2505.19114, 2025

    Hui Zhang, Dexiang Hong, Maoke Yang, Yutao Cheng, Zhao Zhang, Weidong Chen, Jie Shao, Xinglong Wu, Zuxuan Wu, and Yu-Gang Jiang. Creatidesign: A unified multi-conditional diffusion transformer for creative graphic design.arXiv preprint arXiv:2505.19114, 2025

    arXiv 2025

  74. [74]

    Jingyu Li, Zhendong Mao, Hao Li, Weidong Chen, and Yongdong Zhang. Exploring visual relationships via transformer-based graphs for enhanced image captioning.ACM Transactions on Multimedia Computing, Communications and Applications, 20(5):1–23, 2024

    2024

  75. [75]

    Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.Advances in neural information processing systems, 35:22199–22213, 2022

    2022

  76. [76]

    Stimuli-aware emotion adaptor for enhancing llm in affective explanation captioning

    Zhiyan Zhang, Peipei Song, Jinpeng Hu, Weidong Chen, Lin Ni, and Xun Yang. Stimuli-aware emotion adaptor for enhancing llm in affective explanation captioning. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10662–10666. IEEE, 2026

    2026

  77. [77]

    A multi-agent framework with structured reasoning and reflective refinement for multimodal empathetic response generation.arXiv preprint arXiv:2604.18988, 2026

    Liping Wang, Cheng Ye, Weidong Chen, Peipei Song, Bo Hu, and Zhendong Mao. A multi-agent framework with structured reasoning and reflective refinement for multimodal empathetic response generation.arXiv preprint arXiv:2604.18988, 2026

    Pith/arXiv arXiv 2026

  78. [78]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5

    2026

  79. [79]

    X-stance: A multilingual multi-target dataset for stance detection.arXiv preprint arXiv:2003.08385, 2020

    Jannis Vamvas and Rico Sennrich. X-stance: A multilingual multi-target dataset for stance detection.arXiv preprint arXiv:2003.08385, 2020. Manuscript submitted to ACM Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection 23

    Pith/arXiv arXiv 2003

  80. [80]

    Multilingual stance detection in tweets: The catalonia independence corpus

    Elena Zotova, Rodrigo Agerri, Manuel Nuñez, and German Rigau. Multilingual stance detection in tweets: The catalonia independence corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1368–1375, 2020

    2020

Showing first 80 references.