Back

JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection

Bin LiangQinglin ZhuXiang LiMin YangLin GuiYulan HeRuifeng Xu

article2022ACL132 citations

Presents a joint contrastive learning framework that uses stance contrastive learning and prototypical graph structures to transfer stance-reasoning capabilities from known targets to unseen ones, achieving state-of-the-art results in zero-shot stance detection.

Listen

Modern organizations and public monitoring systems need to automatically identify viewpoints or attitudes expressed in text towards specific topics, propositions, or entities. However, real-world debates continuously produce new subjects that have no historical training data. Existing stance detection methods struggle when faced with these unobserved topics because they cannot reliably generalize what they learned from familiar topics to entirely new ones during deployment.

The main objective of the article is to demonstrate a joint contrastive learning framework, named JointCL, designed to accurately detect stances on previously unseen targets. The article evaluates how combining context-based stance reasoning with target-based relational representations improves performance across zero-shot, few-shot, and cross-target settings.

The evaluated approach uses language representations generated by a pre-trained transformer model and enhances them through two complementary strategies. First, a stance contrastive learning mechanism groups instances sharing the same viewpoint while pushing away differing viewpoints. Second, the system clusters training examples into representative concept prototypes and builds graphs connecting each instance to these prototypes. Using a graph attention network and a novel edge-oriented contrastive objective, the model learns structural relationships between known topics and prototypes and transfers those patterns to unseen targets. The authors validated this framework across three benchmark datasets covering diverse general topics, political debates, and corporate financial mergers.

The experimental findings show that the proposed framework consistently establishes new state-of-the-art benchmarks in zero-shot stance detection. On the large, multi-topic benchmark, the framework achieved an overall Macro F1 score of 72.3%, significantly outperforming established neural and graph baselines. Ablation analyses showed that removing the stance contrastive module or the prototypical graph module caused noticeable performance drops across all datasets, confirming both components are necessary. The framework also generalized effectively to few-shot scenarios with a leading 71.5% F1 score and achieved superior performance across all cross-target evaluation pairs.

These findings imply that automated sentiment and stance analysis systems can reliably handle fast-emerging public topics without requiring expensive, time-consuming data re-labeling or target-specific retraining. The success of bridging known and unknown targets through shared prototype graphs demonstrates that relational structure is critical when contextual text alone is insufficient to deduce a viewpoint. Consequently, adopting this unified framework reduces deployment latency and ongoing annotation costs for monitoring systems.

For practical implementation, organizations deploying stance detection should adopt prototype-driven graph contrastive architectures rather than standard fine-tuning or adversarial methods. Teams should tune the number of prototype clusters to match the diversity of their target domain, using larger cluster counts for broad topic distributions and smaller counts for narrow, domain-specific settings. A sensible next step is to run domain-specific pilot deployments to confirm prototype stability under evolving real-time data streams.

Confidence in these findings is supported by rigorous multi-run evaluations and statistical significance testing across diverse benchmarks. However, practitioners should note that performance remains sensitive to hyper-parameter choices, such as cluster quantity and loss weighting, and zero-shot accuracy is inherently lower than cross-target accuracy where target relationships are partially known in advance.

Harbin Institute of TechnologyJoint Lab of HITSZ and China Merchants SecuritiesPeng Cheng LaboratoryShenzhen Institute of Advanced Technology, Chinese Academy of SciencesThe Alan Turing InstituteUniversity of Warwickcontrastive learningdetection algorithmgraph structureszero-shot stance detectiontarget-aware prototypical graph contrastive learningstance contrastive learning
HITSZ-HLT/JointCLLiang et al (2022).pdf
  • Paper: Prototypical Networks for Few-shot Learning, Jake Snell et al. (2017). Its class-prototype construction provides the foundation for understanding JointCL’s use of prototypes to transfer target-based representations to unseen targets.
  • Paper: Graph Contrastive Learning with Augmentations, Yuning You et al. (2020). GraphCL introduces contrastive representation learning on graph-structured data, preparing you to follow JointCL’s prototypical graph contrastive component.
  • Paper: Supervised Contrastive Learning, Prannay Khosla et al. (2020). Its supervised contrastive objective clarifies the class-aware contrastive learning principles JointCL adapts to improve stance-feature generalization.
  • Paper: SimCSE: Simple Contrastive Learning of Sentence Embeddings, Tianyu Gao et al. (2021). SimCSE shows how contrastive objectives can shape transferable sentence representations, a useful prerequisite for understanding JointCL’s stance representation learning.

No sufficiently relevant recommendations were found.

Cover for JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection

Abstract

Zero-shot stance detection (ZSSD) aims to detect the stance for an unseen target during the inference stage. In this paper, we propose a joint contrastive learning (JointCL) framework, which consists of stance contrastive learning and target-aware prototypical graph contrastive learning. Specifically, a stance contrastive learning strategy is employed to better generalize stance features for unseen targets. Further, we build a prototypical graph for each instance to learn the target-based representation, in which the prototypes are deployed as a bridge to share the graph structures between the known targets and the unseen ones. Then a novel target-aware prototypical graph contrastive learning strategy is devised to generalize the reasoning ability of target-based stance representations to the unseen targets. Extensive experiments on three benchmark datasets show that the proposed approach achieves state-of-the-art performance in the ZSSD task¹.

Topics

contrastive learningdetection algorithmgraph structuresinferencereasoningdirected graphs

Keywords

zero-shot stance detectiontarget-aware prototypical graph contrastive learningstance contrastive learningJointCL framework

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 2.1 Zero-Shot Stance Detection
  • 2.2 Contrastive Learning
  • 3 Methodology
  • 3.1 Task Description
  • 3.2 Encoder Module
  • 3.3 Stance Contrastive Learning
  • 3.4 Prototypes Generation
  • 3.5 Prototypical Graph
  • 3.6 Target-Aware Prototypical Graph Contrastive Learning
  • 3.7 Stance Detection
  • 3.8 Learning Objective
  • 4 Experimental Setup
  • 4.1 Datasets
  • 4.2 Implementation Detail
  • 4.3 Comparison Models
  • 5 Experimental Results
  • 5.1 Main Results
  • 5.2 Ablation Study
  • 5.3 Impact of the Values of k
  • 5.4 Generalizability Analysis
  • 5.5 Visualization
  • 6 Conclusion
  • Acknowledgments
  • References

Knowls

  1. Knowl 1 — JointCL combines stance-level and target-aware generalization

    model/method

    JointCL addresses zero-shot stance detection (ZSSD), where a model trains on text–target–stance examples for source targets and predicts stance for targets absent from training. It combines two complementary objectives: stance contrastive learning groups representations with the same stance and separates different stances, while target-aware prototypical graph contrastive learning transfers relationships between known targets and learned prototypes to unseen targets. For each text and its target, the model encodes an instance representation, constructs a graph representation using prototypes, concatenates the two representations, and predicts the stance. The framework is designed to generalize from known targets both through stance features in the text context and through learned target-related structure.

  2. Knowl 2 — Supervised stance contrastive loss

    equation

    For a minibatch B\mathcal{B} of NbN_b encoded training instances, JointCL treats every other instance with the same stance as a positive for anchor ii; all other instances in the batch remain in the normalization denominator. The loss encourages representations of same-stance instances to have high cosine similarity. Here, hih_i is the hidden vector for instance ii, yiy_i is its stance label, τs>0\tau_s>0 is a temperature, and f(u,v)=u⊤v/(∥u∥2∥v∥2)f(u,v)=u^\top v/(\|u\|_2\|v\|_2) is cosine similarity: Lstance=−1Nb∑i∈Blog⁡∑j∈B, j≠i1[yi=yj]exp⁡(f(hi,hj)/τs)∑j∈B, j≠iexp⁡(f(hi,hj)/τs)\mathcal{L}_{\mathrm{stance}}=-\frac{1}{N_b}\sum_{i\in\mathcal{B}}\log\frac{\sum_{j\in\mathcal{B},\,j\ne i}\mathbf{1}[y_i=y_j]\exp(f(h_i,h_j)/\tau_s)}{\sum_{j\in\mathcal{B},\,j\ne i}\exp(f(h_i,h_j)/\tau_s)}. The representations are formed by feeding the text rr and target tt as [CLS]r[SEP]t[SEP][\mathrm{CLS}]r[\mathrm{SEP}]t[\mathrm{SEP}] to BERT and using the [CLS][\mathrm{CLS}] vector as hih_i.

  3. Knowl 3 — Prototype-based graph construction

    model/method

    JointCL derives kk prototypes by applying kk-means to the hidden vectors of the training instances; each prototype represents a cluster of semantically similar examples, rather than a target class mean. The clusters are recalculated each training epoch. For a training instance with hidden vector hih_i, the graph nodes comprise all kk prototypes and hih_i. JointCL uses a fully connected adjacency matrix, with every entry equal to 1, and feeds the nodes and adjacency matrix to a graph attention network. The network produces attention weights αi\alpha_i over graph edges and a graph representation ziz_i for the instance. The prototypes provide shared graph structure through which target-based representations of known targets can be related to those of unseen targets.

  4. Knowl 4 — Target-aware edge-oriented graph contrastive loss

    equation

    JointCL contrasts graph edge-attention patterns so that instances with related target-based information have similar graph structures. For anchor ii, a positive instance jj must have the same stance and either concern the same target or be assigned to the same prototype; every other instance in the minibatch contributes to the denominator. Let αi\alpha_i be the graph attention matrix for instance ii, flattened into a vector for cosine similarity, and let qij=1q_{ij}=1 exactly when the positive-pair conditions hold and qij=0q_{ij}=0 otherwise. With minibatch B\mathcal{B} of size NbN_b and temperature τg>0\tau_g>0, the loss is Lgraph=−1Nb∑i∈Blog⁡∑j∈B, j≠iqijexp⁡(f(αi,αj)/τg)∑j∈B, j≠iexp⁡(f(αi,αj)/τg)\mathcal{L}_{\mathrm{graph}}=-\frac{1}{N_b}\sum_{i\in\mathcal{B}}\log\frac{\sum_{j\in\mathcal{B},\,j\ne i}q_{ij}\exp(f(\alpha_i,\alpha_j)/\tau_g)}{\sum_{j\in\mathcal{B},\,j\ne i}\exp(f(\alpha_i,\alpha_j)/\tau_g)}, where ff is cosine similarity. This objective learns relationships between instances through their prototype-linked graph edges, rather than contrasting only their hidden or graph-level instance vectors.

  5. Knowl 5 — Prediction and joint optimization

    equation

    For training instance ii, JointCL concatenates its BERT hidden vector hih_i and graph representation ziz_i into vi=hi⊕ziv_i=h_i\oplus z_i. A trainable linear layer with parameters W,bW,b and a softmax predicts the distribution over stance classes: y^i=softmax⁡(Wvi+b)\hat{y}_i=\operatorname{softmax}(Wv_i+b). The classification loss Lclass\mathcal{L}_{\mathrm{class}} is cross-entropy between predicted and one-hot gold stance distributions. The full objective is L=γcLclass+γsLstance+γgLgraph+λ∥Θ∥22\mathcal{L}=\gamma_c\mathcal{L}_{\mathrm{class}}+\gamma_s\mathcal{L}_{\mathrm{stance}}+\gamma_g\mathcal{L}_{\mathrm{graph}}+\lambda\|\Theta\|_2^2, where Θ\Theta denotes trainable model parameters, γc,γs,γg\gamma_c,\gamma_s,\gamma_g weight the three task losses, and λ\lambda weights L2 regularization.

  6. Knowl 6 — Zero-shot benchmark performance

    empirical result

    On the paper’s three zero-shot stance detection benchmarks, JointCL improved the aggregate VAST score and all four reported WT-WT target scores over the strongest listed baseline for each metric. On VAST, the reported Pro, Con, Neutral, and All scores (%) were 64.9, 63.2, 88.9, and 72.3 for JointCL, compared with 61.2, 61.2, 88.0, and 70.2 for CKE-Net. On SEM16, for targets DT, HC, FM, LA, A, and CC respectively, JointCL scored 50.5, 54.8, 53.8, 49.5, 54.5, and 39.7; the strongest baseline scores for those targets were 49.5 (TOAD), 51.2 (TOAD), 54.1 (TOAD), 46.5 (TPDG), 55.2 (BERT), and 37.3 (BERT). Thus JointCL was not highest on every SEM16 target. On WT-WT targets CA, CE, AC, and AH, JointCL scored 72.4, 70.2, 76.0, and 75.2, versus 66.8, 65.6, 74.2, and 73.1 for TPDG. Scores are percentages; VAST reports class-wise macro F1 and an All score, SEM16 reports its target scores, and WT-WT reports macro F1 by target. The paper reports significance at p<0.05p<0.05 on most evaluation metrics.

  7. Knowl 7 — Benchmark data and training protocol

    experimental setup

    JointCL was evaluated on VAST, SEM16, and WT-WT. VAST uses Pro, Con, and Neutral labels and has 13,477/2,062/3,006 train/dev/test examples, 1,845/682/786 unique comments, 4,003/383/600 zero-shot topics, and 638/114/159 few-shot topics. SEM16 contains six stance targets with Favor, Against, and Neutral labels; WT-WT consists of company-pair targets and uses Support, Refute, Comment, and Unrelated labels. SEM16 and WT-WT use leave-one-target-out evaluation. The encoder is uncased BERT-base with 768-dimensional token embeddings. Training uses Adam, learning rate 3×10−53\times10^{-5}, batch size 16, and L2 coefficient 10−510^{-5}. Both contrastive temperatures are 0.07. The number of clusters is 100 for VAST and 10 for SEM16 and WT-WT; γc=0.8\gamma_c=0.8 and γs=1\gamma_s=1, while γg=0.1\gamma_g=0.1 for VAST and 0.5 for SEM16 and WT-WT. Early stopping uses patience 5, and reported results average 10 runs. VAST is evaluated with macro-averaged F1 by label, SEM16 with the average F1 on Favor and Against, and WT-WT with macro F1 for each target.

  8. Knowl 8 — Ablations support the contributions of both contrastive objectives and graph structure

    empirical result

    Removing any of the tested JointCL components reduced performance across the reported benchmark metrics. On VAST, the All score (%) was 72.3 for the complete model, compared with 69.8 without stance contrastive learning, 70.7 without graph contrastive learning, 70.3 when target-aware contrastive learning was applied to hidden vectors instead of prototypical graphs, 69.5 without clustering (using a mean representation as prototype), and 71.4 without edge-oriented contrast on the attention matrices. The paper also reports reductions for these variants across SEM16 and WT-WT. These comparisons indicate that the stance-based objective, prototype generation, graph-based target representation, and explicit edge information each contributed to the tested system’s performance.

  9. Knowl 9 — Evaluation in few-shot and cross-target settings

    empirical result

    The authors additionally evaluated JointCL in a VAST few-shot condition and in cross-target stance detection on SEM16. In the VAST few-shot evaluation, JointCL’s Pro, Con, Neutral, and All scores (%) were 63.2, 66.7, 84.6, and 71.5. Its All score exceeded the strongest comparison score, 70.1 from CKE-Net, although CKE-Net scored higher on Pro (64.4 versus 63.2). For cross-target training-to-test pairs HC→DT, DT→HC, FM→LA, and LA→FM, JointCL scored 52.8, 54.3, 58.8, and 54.5, respectively; the strongest prior scores shown for those pairs were 50.4, 52.9, 58.3, and 54.1. These results show that JointCL’s reported gains extended to these settings, without implying that it led on every few-shot label score.

  10. Knowl 10 — Performance depends on the number of prototypes

    empirical result

    The paper varied the number of k-means clusters kk and found dataset-dependent optima. On VAST, which has a comparatively large target inventory, performance rose as kk increased and peaked at k=100k=100; larger values then reduced performance. On SEM16 and WT-WT, the better-performing region was k∈[10,20]k\in[10,20], with the peak at k=10k=10. The selected settings for the main experiments were therefore k=100k=100 on VAST and k=10k=10 on SEM16 and WT-WT.

Coverage note — The per-class SEM16 and WT-WT corpus-count breakdown and the t-SNE visualization are omitted because they characterize the datasets or qualitatively illustrate representation separation rather than adding independent quantitative findings.

References

  1. 1.Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913–8931, Online. Association for Computational Linguistics.
  2. 2.Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756–4767, Online. Association for Computational Linguistics.
  3. 3.Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876–885, Austin, Texas. Association for Computational Linguistics.
  4. 4.Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020a. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1597–1607. PMLR.
  5. 5.Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. 2020b. Big self-supervised models are strong semi-supervised learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
  6. 6.Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2020. Will-they-won’t-they: A very large dataset for stance detection on Twitter. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1715–1724, Online. Association for Computational Linguistics.
  7. 7.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), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics.
  8. 8.John Giorgi, Osvald Nitski, Bo Wang, and Gary Bader. 2021. DeCLUTR: Deep contrastive learning for unsupervised textual representations. In Proceedings 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 879–895, Online. Association for Computational Linguistics.
  9. 9.Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Representativeness of abortion legislation debate on twitter: A case study in argentina and chile. In Companion Proceedings of the Web Conference 2020, page 765–774.
  10. 10.Beliz Gunel, Jingfei Du, Alexis Conneau, and Veselin Stoyanov. 2021. Supervised contrastive learning for pre-trained language model fine-tuning. In International Conference on Learning Representations.
  11. 11.Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2, pages 1735–1742. IEEE.
  12. 12.Mohammad Kachuee, Hao Yuan, Young-Bum Kim, and Sungjin Lee. 2021. Self-supervised contrastive learning for efficient user satisfaction prediction in conversational agents. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4053–4064, Online. Association for Computational Linguistics.
  13. 13.Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
  14. 14.Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. 2021. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations.
  15. 15.Bin Liang, Yonghao Fu, Lin Gui, Min Yang, Jiachen Du, Yulan He, and Ruifeng Xu. 2021a. Target-adaptive graph for cross-target stance detection. In the Web Conference 2021 (WWW ’21).
  16. 16.Bin Liang, Wangda Luo, Xiang Li, Lin Gui, Min Yang, Xiaoqi Yu, and Ruifeng Xu. 2021b. Enhancing Aspect-Based Sentiment Analysis with Supervised Contrastive Learning, page 3242–3247. Association for Computing Machinery, New York, NY, USA.
  17. 17.Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3152–3157, Online. Association for Computational Linguistics.
  18. 18.Yixin Liu and Pengfei Liu. 2021. SimCLS: A simple framework for contrastive learning of abstractive summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 1065–1072, Online. Association for Computational Linguistics.
  19. 19.Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31–41, San Diego, California. Association for Computational Linguistics.
  20. 20.Mitra Mohtarami, Ramy Baly, James Glass, Preslav Nakov, Lluís Màrquez, and Alessandro Moschitti. 2018. Automatic stance detection using end-to-end memory networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 767–776, New Orleans, Louisiana. Association for Computational Linguistics.
  21. 21.Yujia Qin, Yankai Lin, Ryuichi Takanobu, Zhiyuan Liu, Peng Li, Heng Ji, Minlie Huang, Maosong Sun, and Jie Zhou. 2021. ERICA: Improving entity and relation understanding for pre-trained language models via contrastive learning. In Proceedings 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 3350–3363, Online. Association for Computational Linguistics.
  22. 22.Yao Qiu, Jinchao Zhang, and Jie Zhou. 2021. Improving gradient-based adversarial training for text classification by contrastive learning and auto-encoder. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1698–1707, Online. Association for Computational Linguistics.
  23. 23.T. Y.S.S. Santosh, Srijan Bansal, and Avirup Saha. 2019. Can siamese networks help in stance detection? In Proceedings of the ACM India Joint International Conference on Data Science and Management of Data, page 306–309.
  24. 24.Jake Snell, Kevin Swersky, and Richard S. Zemel. 2017. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4077–4087.
  25. 25.Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. 2017. A dataset for multi-target stance detection. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 551–557, Valencia, Spain. Association for Computational Linguistics.
  26. 26.Swapna Somasundaran and Janyce Wiebe. 2010. Recognizing stances in ideological on-line debates. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 116–124, Los Angeles, CA. Association for Computational Linguistics.
  27. 27.Yonglong Tian, Dilip Krishnan, and Phillip Isola. 2020. Contrastive multiview coding. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, pages 776–794. Springer.
  28. 28.Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579–2605.
  29. 29.Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.
  30. 30.Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021. CLEVE: Contrastive Pre-training for Event Extraction. In Proceedings 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 6283–6297, Online. Association for Computational Linguistics.
  31. 31.Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. 2018. Unsupervised feature learning via non-parametric instance discrimination. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 3733–3742. IEEE Computer Society.
  32. 32.Chang Xu, Cécile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778–783, Melbourne, Australia. Association for Computational Linguistics.
  33. 33.Nan Yang, Furu Wei, Binxing Jiao, Daxing Jiang, and Linjun Yang. 2021. xMoCo: Cross momentum contrastive learning for open-domain question answering. In Proceedings 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 6120–6129, Online. Association for Computational Linguistics.
  34. 34.Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. Advances in Neural Information Processing Systems, 33:5812–5823.
  35. 35.Bowen Zhang, Min Yang, Xutao Li, Yunming Ye, Xiaofei Xu, and Kuai Dai. 2020. Enhancing cross-target stance detection with transferable semantic-emotion knowledge. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3188–3197, Online. Association for Computational Linguistics.
  36. 36.Dejiao Zhang, Feng Nan, Xiaokai Wei, Shang-Wen Li, Henghui Zhu, Kathleen McKeown, Ramesh Nallapati, Andrew O. Arnold, and Bing Xiang. 2021. Supporting clustering with contrastive learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5419–5430, Online. Association for Computational Linguistics.

Citation

MLA
Liang, B., et al. “JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 81–91, https://doi.org/10.18653/V1/2022.ACL-LONG.7.
APA
Liang, B., Zhu, Q., Li, X., Yang, M., Gui, L., He, Y., & Xu, R. (2022). JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 81–91. https://doi.org/10.18653/V1/2022.ACL-LONG.7
Chicago
Liang, B., Q. Zhu, X. Li, et al. 2022. “JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 81–91. https://doi.org/10.18653/V1/2022.ACL-LONG.7.
Harvard
Liang, B. et al. (2022) “JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection”, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 81–91. Available at: https://doi.org/10.18653/V1/2022.ACL-LONG.7.
Vancouver
1. Liang B, Zhu Q, Li X, Yang M, Gui L, He Y, Xu R (2022) JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 81–91

BibTeX

@inproceedings{Liang_2022, title={JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection}, url={http://dx.doi.org/10.18653/V1/2022.ACL-LONG.7}, DOI={10.18653/v1/2022.acl-long.7}, booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, publisher={Association for Computational Linguistics}, author={Liang, Bin and Zhu, Qinglin and Li, Xiang and Yang, Min and Gui, Lin and He, Yulan and Xu, Ruifeng}, year={2022}, pages={81–91} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: https://creativecommons.org/licenses/by/4.0/