Back

Stance Detection on Social Media with Background Knowledge

Ang LiBin LiangJingqian ZhaoBowen ZhangMin YangRuifeng Xu

article2023EMNLP70 citations

Proposes a knowledge-augmented stance detection framework that integrates Wikipedia-derived episodic context and language-model-interpreted social discourse to achieve state-of-the-art accuracy in both in-target and zero-shot settings.

Listen

Social media stance detection aims to automatically determine whether an author supports, opposes, or remains neutral toward a specific target or topic. However, social media posts are typically brief and heavily laden with slang, acronyms, and hashtags, while omitting the essential factual background required to comprehend implicit viewpoints. Existing automated methods struggle because they rely primarily on surface-level text or lack domain-general background information, making accurate public opinion analysis difficult.

The article evaluates whether augmenting stance detection models with external background knowledge improves classification performance. Specifically, it demonstrates the Knowledge-Augmented Stance Detection framework, which categorizes background context into factual understanding of the topic (episodic knowledge) and interpretations of social media informal language and references (discourse knowledge).

To implement this approach, the authors developed a heuristic retrieval pipeline that crawls Wikipedia pages relevant to specific topics and uses topic modeling alongside text-matching algorithms to identify relevant sections. A large language model filters this retrieved content to remove non-essential text and paraphrase informal social media posts by resolving abbreviations and hashtags. The framework was evaluated across four standard benchmark datasets covering political figures, COVID-19 policies, and diverse debate topics under both known-target (in-target) and unseen-target (zero-shot) conditions.

The analysis yields four critical findings. First, the framework consistently achieves top performance across standard benchmarks, significantly outperforming existing baseline models and pre-trained architectures. Second, knowledge augmentation substantially improves zero-shot stance detection on unseen topics, raising average performance on challenging datasets from roughly 40–53% up to 58%. Third, human evaluations confirmed high data extraction fidelity, with independent reviewers rating the generated and filtered knowledge as accurate and relevant in over 95% of sampled cases. Fourth, distilling background and contextual knowledge into smaller, specialized language models enables them to outperform large general-purpose foundation models while operating with 500 to 1,000 times fewer parameters.

These findings indicate that smaller, fine-tuned models augmented with targeted external knowledge provide a superior, cost-effective alternative to relying entirely on massive, computationally expensive language models for public opinion monitoring. Organizations can deploy compact, highly accurate stance detection systems that lower computational overhead, reduce deployment latency, and mitigate the risk of classification errors caused by ambiguous online slang.

Decision-makers building opinion monitoring or social media analytics capabilities should consider adopting knowledge-retrieval frameworks before investing heavily in larger proprietary models. Implementing a hybrid approach—using larger language models offline to retrieve and filter knowledge bases while running compact fine-tuned models in production—offers the best balance of accuracy and operating cost. Further work should explore incorporating real-time news sources to address emerging topics beyond static reference materials.

The primary limitation of this study is its reliance on pre-crawled Wikipedia articles, which may limit effectiveness for rapidly evolving, breaking events where online encyclopedias are not immediately updated. Nevertheless, given the consistent empirical improvements across diverse datasets and high human-evaluation scores, there is high confidence in the framework's effectiveness for general and zero-shot stance detection tasks.

Guangdong Provincial Key Laboratory of Novel Security Intelligence TechnologiesHarbin Institute of TechnologyPeng Cheng LaboratoryShenzhen Institute of Advanced Technology, Chinese Academy of SciencesShenzhen Technology UniversityThe Chinese University of Hong Kongbackground knowledgedetection algorithmretrieval algorithmsstance detectionbackground knowledgeknowledge-augmented stance detection
HITSZ-HLT/KA-Stance-DetectionLi et al (2023).pdf

No sufficiently relevant recommendations were found.

Cover for Stance Detection on Social Media with Background Knowledge

Abstract

Identifying users’ stances regarding specific targets/topics is a significant route to learning public opinion from social media platforms. Most existing studies of stance detection strive to learn stance information about specific targets from the context, in order to determine the user’s stance on the target. However, in real-world scenarios, we usually have a certain understanding of a target when we express our stance on it. In this paper, we investigate stance detection from a novel perspective, where the background knowledge of the targets is taken into account for better stance detection. To be specific, we categorize background knowledge into two categories: episodic knowledge and discourse knowledge, and propose a novel Knowledge-Augmented Stance Detection (KASD) framework. For episodic knowledge, we devise a heuristic retrieval algorithm based on the topic to retrieve the Wikipedia documents relevant to the sample. Further, we construct a prompt for ChatGPT to filter the Wikipedia documents to derive episodic knowledge. For discourse knowledge, we construct a prompt for ChatGPT to paraphrase the hashtags, references, etc., in the sample, thereby injecting discourse knowledge into the sample. Experimental results on four benchmark datasets demonstrate that our KASD achieves state-of-the-art performance in in-target and zero-shot stance detection.

Topics

background knowledgedetection algorithmretrieval algorithmsdomain knowledgemicro-blogtext documentexplicit knowledgetacit knowledge

Keywords

stance detectionbackground knowledgeknowledge-augmented stance detectionzero-shot stance detectionjoint contrastive learning frameworksocial mediaknowledge injectioncommonsense knowledgeknowledge retrievalLatent Dirichlet Allocation

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • Incorporating Episodic Knowledge
  • Incorporating Discourse Knowledge
  • 3 Methodology
  • 3.1 Episodic Knowledge Acquisition
  • 3.2 Retrieval and Filtering
  • 3.2.1 Topic Modeling
  • 3.2.2 Heuristic Retrieval
  • 3.2.3 Large Language Model Filtering
  • 3.3 Discourse Knowledge Injection
  • 3.4 Knowledge-Augmented Stance Detection
  • 3.4.1 Fine-tuned Model Stance Detection
  • 3.4.2 Large Language Model Stance Detection
  • 4 Experimental Setup
  • 4.1 Datasets
  • 4.2 Implementation Details
  • 4.3 Evaluation Metric
  • 4.4 Comparison Models
  • 5 Experimental Results
  • 5.1 In-Target Stance Detection
  • 5.2 Zero-Shot Stance Detection
  • 5.3 Ablation Study
  • 5.4 Analysis of Topic Model
  • 5.5 Human Evaluation
  • 5.6 Fine-tuned Model vs Large Language Model
  • 6 Conclusion
  • Acknowledgments
  • Limitations
  • Ethics Statement
  • References
  • A Case Study

Knowls

  1. Knowl 1 — KASD separates target knowledge from social-media discourse knowledge

    model/method

    Knowledge-Augmented Stance Detection (KASD) augments a post and its target with two different kinds of context. Episodic knowledge is relevant factual background about the target that may be needed to infer the post’s stance but is not stated in the post. Discourse knowledge clarifies social-media expressions such as acronyms, hashtags, slang, and references. KASD obtains episodic knowledge from Wikipedia documents retrieved for the sample and filtered by ChatGPT, and obtains discourse knowledge by having ChatGPT restate the post with its expressions clarified.

    For a fine-tuned classifier, KASD combines the target, the restated post, and the filtered episodic knowledge in one input sequence, separated with standard special tokens. The encoded sequence is classified into favor, against, or neutral. In the experiments, the encoder was RoBERTa and the classifier was a fully connected layer with batch normalization, LeakyReLU, and a softmax output, trained with cross-entropy and L2 regularization. Training used AdamW, batch size 16, at most 30 epochs, learning rate 10−510^{-5}, weight decay 10−310^{-3}, and warm-up ratio 0.2; reported fine-tuned results average five runs. KASD was also applied to ChatGPT by adding the retrieved and generated knowledge to the stance-detection prompt; the experiments used gpt-3.5-turbo-0301 with temperature zero.

  2. Knowl 2 — Topic-informed TF-IDF retrieval selects episodic knowledge

    algorithm

    Given a social-media text xx, its target, and a Wikipedia knowledge base, KASD retrieves candidate documents by combining topic relevance with the text’s own expressions. The knowledge base is made by retrieving the top 10 Wikipedia pages for each target and splitting each page into section-level documents averaging about 400 words. A corpus-wide TF-IDF model is built over these documents.

    For a target, an LDA topic model estimates the topic distribution P(z∣x)P(z\mid x) for a sample and the word distribution P(v∣z)P(v\mid z) for each topic zz. The paper fits LDA with online variational Bayes. It selects the number of topics per target from 5 through 19 using the minimum first-order difference in perplexity; for VAST, whose targets have too few examples for effective target-specific topic modeling, retrieval uses the text terms without LDA topics. For each candidate document dd, the score combines its TF-IDF weights for topic words, weighted by their topic and sample probabilities, with its TF-IDF weights for the sample’s own words:

    Stopic(d,x)=∑zP(z∣x)∑v∈V50(z)tfidf⁡d(v)P(v∣z)S_{\text{topic}}(d,x)=\sum_z P(z\mid x)\sum_{v\in V_{50}(z)}\operatorname{tfidf}_d(v)P(v\mid z) and Stext(d,x)=1M∑m=1Mtfidf⁡d(wm)S_{\text{text}}(d,x)=\frac{1}{M}\sum_{m=1}^{M}\operatorname{tfidf}_d(w_m); the final score is Stopic(d,x)+Stext(d,x)S_{\text{topic}}(d,x)+S_{\text{text}}(d,x).

    Here, zz ranges over the target’s LDA topics; V50(z)V_{50}(z) is the set of the 50 highest-probability words for topic zz; vv is a vocabulary word; wmw_m is the word at position mm in the input text; MM is the number of word positions in that text; and tfidf⁡d(q)\operatorname{tfidf}_d(q) is the corpus-model TF-IDF weight of term qq in candidate document dd. The top 50 words cover more than 0.99 of each topic’s word probability mass in the authors’ experiments. KASD retains candidate documents scoring above 0.02. Topic terms capture the sample’s broader themes, while the text-term component accounts for sample-specific wording.

  3. Knowl 3 — ChatGPT filters and summarizes retrieved Wikipedia documents

    model/method

    After heuristic retrieval, KASD processes each candidate Wikipedia document separately with a prompt containing the social-media sentence, its target, and the document. ChatGPT is instructed to return “None” if the document is not relevant to both the sentence and the target; otherwise, it summarizes the document’s knowledge relevant to that sentence-target pair. KASD discards documents returning “None” and concatenates the relevant summaries into the sample’s episodic knowledge.

    The prompt constrains extraction to the supplied Wikipedia document rather than allowing ChatGPT to supply background from its internal knowledge. This filtering is intended to remove irrelevant or redundant material while retaining document context that might be lost by further splitting or retrieval.

  4. Knowl 4 — ChatGPT restates posts to inject discourse knowledge

    model/method

    To produce discourse-augmented text, KASD prompts ChatGPT with the original sentence and asks it to expand abbreviations, slang, and hashtags into complete phrases and sentences, thereby restating the text. The intended output clarifies the expressions in context, including references that help identify what a user is talking about. This restatement is combined with the target and filtered episodic knowledge for stance classification; it is not a replacement for the original target or the episodic-knowledge retrieval step.

  5. Knowl 5 — KASD improves in-target and zero-shot stance scores across benchmarks

    empirical result

    The paper evaluates in-target stance detection, where training and testing use the same target, and zero-shot detection, where the test target is unseen during training. Scores below are average F1 percentages. For SemEval-2016 and P-Stance, the reported average is the mean of favor and against F1; for COVID-19-Stance and VAST, it is the mean of favor, against, and neutral F1.

    In in-target testing, KASD-BERT achieved 73.42 on SemEval-2016, compared with 71.90 for the strongest reported fine-tuned baseline (KEprompt); on P-Stance it achieved 83.80, compared with 82.77 for WS-BERT-Dual; and on COVID-19-Stance it achieved 86.17, compared with 84.35 for WS-BERT-Dual. KASD-ChatGPT scored 71.52 on SemEval-2016 versus 69.80 for ChatGPT, 83.20 on P-Stance versus 83.10, and 71.64 on COVID-19-Stance versus 69.26.

    In zero-shot testing, KASD-BERT scored 57.85 on SemEval-2016, compared with 52.70 for JointCL; 74.99 on P-Stance, compared with 72.93 for WS-BERT-Dual; and 76.82 on VAST, compared with 75.30 for WS-BERT-Dual. KASD-ChatGPT scored 71.15 on SemEval-2016 versus 68.70 for ChatGPT, 82.52 on P-Stance versus 81.50, and 67.03 on VAST versus 62.30. The paper reports statistical-significance markers for selected fine-tuned target-level comparisons, not for every aggregate score. For a further zero-shot P-Stance comparison, KASD-RoBERTa-large scored 84.36 for Biden, 79.69 for Sanders, and 85.25 for Trump (83.10 average), versus ChatGPT’s 82.30, 79.40, and 82.80 (81.50 average). The authors report that the fine-tuned model used about 500–1,000 times fewer parameters than ChatGPT.

  6. Knowl 6 — Ablations show episodic knowledge is useful and discourse gains depend on dataset

    data/table

    Removing either knowledge component reduced scores in the reported P-Stance in-target and VAST zero-shot tests. Each value below is average F1 percentage; the P-Stance columns give target-specific scores followed by their average.

    P-Stance, KASD-BERT: full system 85.66 (Biden), 80.39 (Sanders), 85.35 (Trump), 83.80 average; without episodic knowledge 83.41, 78.54, 85.03, 82.33; without discourse knowledge 83.69, 79.01, 84.19, 82.29. VAST: 76.82 for the full system, 74.53 without episodic knowledge, and 76.44 without discourse knowledge.

    P-Stance, KASD-ChatGPT: full system 84.59 (Biden), 79.96 (Sanders), 85.06 (Trump), 83.20 average; without episodic knowledge 82.59, 78.10, 81.69, 80.69; without discourse knowledge 82.87, 77.79, 83.09, 81.25. VAST: 67.03 for the full system, 65.22 without episodic knowledge, and 66.79 without discourse knowledge.

    The reported results show larger VAST drops when episodic knowledge is removed than when discourse knowledge is removed. The authors characterize discourse augmentation as not significantly improving VAST performance, consistent with VAST’s more standardized debate text; they identify episodic knowledge as especially useful for VAST’s many different targets.

  7. Knowl 7 — Retrieval followed by filtering is tested against alternative knowledge inputs

    empirical result

    The paper compares four ways of providing episodic knowledge: using only a Wikipedia page’s summary section; using the heuristic retrieval method to select the most similar Wikipedia document; asking ChatGPT to extract knowledge directly from the knowledge base without retrieval; and combining heuristic retrieval with ChatGPT filtering. These comparisons use P-Stance and SemEval-2016 in-target tests and VAST zero-shot tests, with both a fine-tuned RoBERTa-base model and ChatGPT.

    The authors report that retrieval plus filtering effectively selects sample-relevant background knowledge and improves stance detection, while redundant or irrelevant Wikipedia content can hurt classification. This experiment evaluates the retrieval-and-filtering pipeline against alternatives, rather than isolating only the classifier’s performance.

  8. Knowl 8 — Human review finds high quality in ChatGPT-generated knowledge

    empirical result

    Three evaluators who were not involved in the work assessed 500 randomly selected samples from SemEval-2016, P-Stance, COVID-19-Stance, and VAST. For episodic knowledge, they judged whether the output was relevant to the sample and whether filtering avoided removing required knowledge; for discourse knowledge, they judged consistency with the original content. The reported percentages of “yes” judgments, averaged across evaluators, were 96.00% for episodic-knowledge relevance, 95.13% for not removing required knowledge during filtering, and 96.87% for discourse-knowledge consistency. These evaluations measure judged output quality, not stance-classification accuracy.

  9. Knowl 9 — The knowledge base depends on pre-crawled Wikipedia

    limitation

    KASD requires relevant Wikipedia pages to be crawled in advance to construct its knowledge base. The paper does not evaluate alternative, potentially more timely sources such as news websites. Consequently, the reported method does not establish how well its knowledge acquisition would support timely stance detection when target-relevant facts change or emerge after the Wikipedia material was collected.

Coverage note — The per-target topic-count/perplexity diagnostics and illustrative case studies are omitted because they support method interpretation but do not establish independent, generalizable findings.

References

  1. 1.Emily Allaway and Kathleen R. 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 2020, Online, November 16-20, 2020, pages 8913–8931. Association for Computational Linguistics.
  2. 2.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.
  3. 3.Pengyuan Chen, Kai Ye, and Xiaohui Cui. 2021. Integrating n-gram features into pre-trained model: a novel ensemble model for multi-target stance detection. In International Conference on Artificial Neural Networks, pages 269–279. Springer.
  4. 4.Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2022. Incorporating stock market signals for twitter stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 4074–4091. Association for Computational Linguistics.
  5. 5.Javid Ebrahimi, Dejing Dou, and Daniel Lowd. 2016. Weakly supervised tweet stance classification by relational bootstrapping. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1012–1017.
  6. 6.Tianqing Fang, Hongming Zhang, Weiqi Wang, Yangqiu Song, and Bin He. 2021. DISCOS: bridging the gap between discourse knowledge and commonsense knowledge. In WWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021, pages 2648–2659. ACM / IW3C2.
  7. 7.Shalmoli Ghosh, Prajwal Singhania, Siddharth Singh, Koustav Rudra, and Saptarshi Ghosh. 2019. Stance detection in web and social media: A comparative study. In Experimental IR Meets Multilinguality, Multimodality, and Interaction - 10th International Conference of the CLEF Association, CLEF 2019, Lugano, Switzerland, September 9-12, 2019, Proceedings, volume 11696 of Lecture Notes in Computer Science, pages 75–87. Springer.
  8. 8.Kyle Glandt, Sarthak Khanal, Yingjie Li, Doina Caragea, and Cornelia Caragea. 2021. Stance detection in COVID-19 tweets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 1596–1611. Association for Computational Linguistics.
  9. 9.Kazuaki Hanawa, Akira Sasaki, Naoaki Okazaki, and Kentaro Inui. 2019. Stance detection attending external knowledge from wikipedia. J. Inf. Process., 27:499–506.
  10. 10.Kazi Saidul Hasan and Vincent Ng. 2013. Stance classification of ideological debates: Data, models, features, and constraints. In Proceedings of the Sixth International Joint Conference on Natural Language Processing, pages 1348–1356.
  11. 11.Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71–77, Dublin, Ireland. Association for Computational Linguistics.
  12. 12.Matthew D. Hoffman, David M. Blei, and Francis R. Bach. 2010. Online learning for latent dirichlet allocation. In Advances in Neural Information Processing Systems 23: 24th Annual Conference on Neural Information Processing Systems 2010. Proceedings of a meeting held 6-9 December 2010, Vancouver, British Columbia, Canada, pages 856–864. Curran Associates, Inc.
  13. 13.Hu Huang, Bowen Zhang, Yangyang Li, Baoquan Zhang, Yuxi Sun, Chuyao Luo, and Cheng Peng. 2023. Knowledge-enhanced prompt-tuning for stance detection. ACM Transactions on Asian and Low-Resource Language Information Processing.
  14. 14.Yang Li and Jiawei Yuan. 2022. Generative data augmentation with contrastive learning for zero-shot stance detection. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 6985–6995. Association for Computational Linguistics.
  15. 15.Yingjie Li, Tiberiu Sosea, Aditya Sawant, Ajith Jayaraman Nair, Diana Inkpen, and Cornelia Caragea. 2021. P-stance: A large dataset for stance detection in political domain. In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021, volume ACL/IJCNLP 2021 of Findings of ACL, pages 2355–2365. Association for Computational Linguistics.
  16. 16.Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, pages 2738–2747.
  17. 17.Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. 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), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 81–91. Association for Computational Linguistics.
  18. 18.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.
  19. 19.Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.
  20. 20.Yunpu Ma, Volker Tresp, and Erik A. Daxberger. 2019. Embedding models for episodic knowledge graphs. J. Web Semant., 59.
  21. 21.Saif M. Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiao-Dan Zhu, and Colin Cherry. 2016. Semeval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2016, San Diego, CA, USA, June 16-17, 2016, pages 31–41. The Association for Computer Linguistics.
  22. 22.Dat Quoc Nguyen, Thanh Vu, and Anh Tuan Nguyen. 2020. Bertweet: A pre-trained language model for english tweets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, EMNLP 2020 - Demos, Online, November 16-20, 2020, pages 9–14. Association for Computational Linguistics.
  23. 23.Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 4222–4235. Association for Computational Linguistics.
  24. 24.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.
  25. 25.Peter Stefanov, Kareem Darwish, Atanas Atanasov, and Preslav Nakov. 2020. Predicting the topical stance and political leaning of media using tweets. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 527–537, Online. Association for Computational Linguistics.
  26. 26.Qingying Sun, Zhongqing Wang, Qiaoming Zhu, and Guodong Zhou. 2018. Stance detection with hierarchical attention network. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2399–2409, Santa Fe, New Mexico, USA. Association for Computational Linguistics.
  27. 27.Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971.
  28. 28.Hanzi Xu, Slobodan Vucetic, and Wenpeng Yin. 2022. Openstance: Real-world zero-shot stance detection. CoRR, abs/2210.14299.
  29. 29.Bowen Zhang, Xianghua Fu, Daijun Ding, Hu Huang, Yangyang Li, and Liwen Jing. 2023. Investigating chain-of-thought with chatgpt for stance detection on social media. CoRR, abs/2304.03087.
  30. 30.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.
  31. 31.Kai Zheng, Qingfeng Sun, Yaming Yang, and Fei Xu. 2022. Knowledge stimulated contrastive prompting for low-resource stance detection. In Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 1168–1178. Association for Computational Linguistics.
  32. 32.Lixing Zhu, Zheng Fang, Gabriele Pergola, Robert Procter, and Yulan He. 2022a. Disentangled learning of stance and aspect topics for vaccine attitude detection in social media. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 1566–1580. Association for Computational Linguistics.
  33. 33.Qinglin Zhu, Bin Liang, Jingyi Sun, Jiachen Du, Lanjun Zhou, and Ruifeng Xu. 2022b. Enhancing zero-shot stance detection via targeted background knowledge. In SIGIR ’22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022, pages 2070–2075. ACM.

Citation

MLA
Li, A., et al. “Stance Detection on Social Media with Background Knowledge”. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 15703–17, https://doi.org/10.18653/V1/2023.EMNLP-MAIN.972.
APA
Li, A., Liang, B., Zhao, J., Zhang, B., Yang, M., & Xu, R. (2023). Stance Detection on Social Media with Background Knowledge. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 15703–15717. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.972
Chicago
Li, A., B. Liang, J. Zhao, B. Zhang, M. Yang, and R. Xu. 2023. “Stance Detection on Social Media with Background Knowledge”. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 15703–17. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.972.
Harvard
Li, A. et al. (2023) “Stance Detection on Social Media with Background Knowledge”, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp. 15703–15717. Available at: https://doi.org/10.18653/V1/2023.EMNLP-MAIN.972.
Vancouver
1. Li A, Liang B, Zhao J, Zhang B, Yang M, Xu R (2023) Stance Detection on Social Media with Background Knowledge. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp 15703–15717

BibTeX

@inproceedings{Li_2023, title={Stance Detection on Social Media with Background Knowledge}, url={http://dx.doi.org/10.18653/V1/2023.EMNLP-MAIN.972}, DOI={10.18653/v1/2023.emnlp-main.972}, booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing}, publisher={Association for Computational Linguistics}, author={Li, Ang and Liang, Bin and Zhao, Jingqian and Zhang, Bowen and Yang, Min and Xu, Ruifeng}, year={2023}, pages={15703–15717} }
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/