Back

Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis

Hui WuXiaodong Shi

article2022ACL107 citations

Proposes an adversarial soft prompt tuning framework that combines domain-specific continuous prompts with adversarial training to learn domain-invariant representations, setting new performance benchmarks for both single- and multi-source cross-domain sentiment classification.

Listen

Analyzing customer sentiment across different product categories is essential for modern business intelligence. However, artificial intelligence models trained on one product domain often perform poorly when deployed to another because vocabulary and phrasing differ significantly. While prompt tuning using pre-trained language models has emerged as an efficient alternative to traditional model fine-tuning, standard fixed text templates fail across varied domains because sentiment words naturally shift across contexts (for example, "useful" in book reviews versus "real" in video reviews).

The article develops and evaluates Adversarial Soft Prompt Tuning (AdSPT), a novel framework designed to improve cross-domain sentiment analysis. The objective is to demonstrate that combining learnable, domain-specific continuous vectors ("soft prompts") with adversarial training effectively bridges the gap between source and target domains without requiring manually engineered text templates.

To test this approach, the authors conducted experiments using the benchmark Amazon review dataset spanning four distinct product domains: Books, DVDs, Electronics, and Kitchen appliances. The evaluation covered both single-source adaptation (training on one product category and testing on another) and multi-source adaptation (training on three categories to test on a fourth). The model paired continuous soft prompt vectors for individual domains with domain discriminators that engaged in a minimax game to encourage the language model to learn domain-invariant sentiment features.

The findings confirm that AdSPT delivers superior classification performance. First, AdSPT achieved new state-of-the-art results, reaching an average accuracy of 93.14% in single-source domain adaptation and 93.75% in multi-source adaptation, outperforming established baselines such as BERT-DAAT and SENTIXFix. Second, soft prompt tuning consistently outperformed fixed text templates, yielding an average performance improvement of over 1.3 percentage points. Third, multi-source adaptation generally exceeded single-source performance, except when the source and target domains were already highly similar (such as Electronics and Kitchen appliances). Finally, sensitivity testing revealed that shorter prompt lengths (specifically, three prompt tokens) yielded optimal performance, minimizing computational complexity.

These results demonstrate that organizations can reduce the labor and expertise costs associated with manual prompt engineering while maintaining high analytical accuracy across changing data domains. By combining domain-specific prompts with adversarial learning, systems can reliably extract transferable features from existing labeled datasets. Organizations implementing cross-domain text classification should adopt soft prompt architectures and leverage multi-source data where available, keeping prompt lengths minimal.

While the reported performance establishes high confidence across the tested consumer product reviews, the evaluation is limited to binary sentiment classification within standard e-commerce datasets. Stakeholders should conduct pilot validation when applying the framework to highly complex, multi-class tasks or specialized technical domains before full-scale operational deployment.

sentiment analysistuning methodadversarial soft prompt tuningcross-domain sentiment analysisdomain adaptation
Wu et al (2022).pdf

No sufficiently relevant recommendations were found.

Cover for Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis

Abstract

Cross-domain sentiment analysis has achieved promising results with the help of pre-trained language models. As GPT-3 appears, prompt tuning has been widely explored to enable better semantic modeling in many natural language processing tasks. However, directly using a fixed predefined template for cross-domain research cannot model different distributions of the [MASK] token in different domains, thus making underuse of the prompt tuning technique. In this paper, we propose a novel Adversarial Soft Prompt Tuning method (AdSPT) to better model cross-domain sentiment analysis. On the one hand, AdSPT adopts separate soft prompts instead of hard templates to learn different vectors for different domains, thus alleviating the domain discrepancy of the [MASK] token in the masked language modeling task. On the other hand, AdSPT uses a novel domain adversarial training strategy to learn domain-invariant representations between each source domain and the target domain. Experiments on a publicly available sentiment analysis dataset show that our model achieves new state-of-the-art results for both single-source domain adaptation and multi-source domain adaptation.

Topics

sentiment analysistuning method

Keywords

adversarial soft prompt tuningcross-domain sentiment analysisdomain adaptationpre-trained language models

Table of Contents

  • Abstract
  • 1 Introduction
  • 2 Related Work
  • 3 Problem Formulation
  • 4 Method
  • 4.1 Soft Prompt Tuning for Sentiment Classification
  • 4.2 Domain Adversarial Training
  • 4.3 Learning Procedure
  • 5 Experiments
  • 5.1 Experimental Setup
  • 5.2 Baselines
  • 5.3 Main Results
  • 5.4 Analysis
  • 6 Conclusion
  • Acknowledgements
  • References

Knowls

  1. Knowl 1 — AdSPT uses domain-specific continuous prompts with a shared language model

    model/method

    Adversarial Soft Prompt Tuning (AdSPT) addresses sentiment classification when labeled reviews come from one or more source domains and unlabeled reviews come from a target domain. It uses a shared masked-language-model pre-trained language model (PLM), but assigns each domain its own independently learned sequence of kk soft-prompt embedding vectors. For a tokenized review xx from domain dd, the PLM input is

    Xd(x)=[e([CLS]),e(x),ud,1,…,ud,k,e([MASK]),e([SEP])],X_d(x)=[e([\mathrm{CLS}]),e(x),u_{d,1},\ldots,u_{d,k},e([\mathrm{MASK}]),e([\mathrm{SEP}])],

    where e(⋅)e(\cdot) is the PLM embedding function, ud,ju_{d,j} is the jjth trainable prompt vector for domain dd, and [CLS][\mathrm{CLS}], [MASK][\mathrm{MASK}], and [SEP][\mathrm{SEP}] are special tokens. The PLM and masked-language-model head are shared across domains; the domain-specific prompts provide trainable domain-aware context. The method combines these prompts with adversarial training of the masked-token representations to address domain differences.

  2. Knowl 2 — Sentiment classification is performed by predicting label words at the mask

    model/method

    AdSPT converts binary sentiment classification into masked-language-model prediction. Let xx be a review, dd its domain, and zd(x)∈R∣V∣z_d(x)\in\mathbb{R}^{|V|} the masked-language-model logits produced at the [MASK][\mathrm{MASK}] position using the prompt for domain dd; VV is the PLM vocabulary. The positive and negative labels are represented by the vocabulary words “good” and “bad,” respectively. The class probabilities are

    p(y∣x,d)=exp⁡(zd(x)v(y))∑y′∈{+,−}exp⁡(zd(x)v(y′)),p(y\mid x,d)=\frac{\exp(z_d(x)_{v(y)})}{\sum_{y'\in\{+, -\}}\exp(z_d(x)_{v(y')})},

    where y∈{+,−}y\in\{+,-\} is the sentiment label and v(+)=goodv(+)=\text{good} and v(−)=badv(-)=\text{bad}. For labeled source examples (xi,yi)(x_i,y_i), the sentiment objective is binary cross-entropy, −∑ilog⁡p(yi∣xi,di)-\sum_i\log p(y_i\mid x_i,d_i), where did_i is the source domain of example ii. The masked-token representation and fixed label words let the classifier use the PLM’s vocabulary prediction rather than a newly initialized sentiment-classification head.

  3. Knowl 3 — Each source domain is adversarially aligned with the target through its own discriminator

    model/method

    For mm source domains S1,…,SmS_1,\ldots,S_m and an unlabeled target dataset TT, AdSPT uses one binary domain discriminator glg_l for each source–target pair (Sl,T)(S_l,T). Each discriminator receives the PLM representation at the [MASK][\mathrm{MASK}] position and predicts whether the review came from source domain ll (domain label 00) or the target domain (label 11). In particular, target examples are used in each of the mm pairwise discrimination tasks. If pl(d∣x)p_l(d\mid x) is discriminator glg_l’s probability for domain label dd on review xx, the summed domain cross-entropy is

    Ldomain=−∑l=1m[∑x∈Sllog⁡pl(0∣x)+∑x∈Tlog⁡pl(1∣x)].L_{\mathrm{domain}}=-\sum_{l=1}^{m}\left[\sum_{x\in S_l}\log p_l(0\mid x)+\sum_{x\in T}\log p_l(1\mid x)\right].

    Here SlS_l denotes the source reviews from domain ll, TT the target reviews, and each review is represented using its own domain’s soft prompt. The discriminators minimize this loss, while the PLM and prompts are trained to maximize it. Thus the discriminators learn to distinguish each source from the target as the representation learner is encouraged to make that distinction harder.

  4. Knowl 4 — Joint training trades off source sentiment prediction against domain invariance

    equation

    AdSPT combines the labeled-source sentiment loss LclassL_{\mathrm{class}} with the pairwise source–target domain loss LdomainL_{\mathrm{domain}}. Its stated joint objective is

    min⁡M,p,f{λLclass−min⁡gLdomain},\min_{M,p,f}\left\{\lambda L_{\mathrm{class}}-\min_g L_{\mathrm{domain}}\right\},

    where MM is the shared PLM, pp denotes all domain-specific soft-prompt vectors, ff is the masked-language-model head, gg denotes the collection of source–target domain discriminators, and λ\lambda is the trade-off weight. The sentiment term trains the model to predict source polarity labels; the negative domain term makes the PLM and prompts oppose the discriminators’ ability to identify source versus target. The discriminators, in turn, are optimized to minimize their domain-classification loss.

  5. Knowl 5 — AdSPT alternates source sentiment updates with pairwise adversarial updates

    algorithm

    The training procedure uses labeled reviews from every source domain and unlabeled reviews from the target domain. The sentiment loss is cross-entropy over the two masked-language-model label words; each domain loss is binary cross-entropy for distinguishing one source domain from the target. Initialize the shared PLM MM, domain-specific prompt vectors pp, masked-language-model head ff, and one discriminator glg_l per source domain. For each source domain, update the prediction head and discriminator to reduce their respective losses, then update the PLM and prompts to reduce sentiment loss while increasing domain loss.

    Input: Labeled source datasets S_1, ..., S_m; unlabeled target dataset T; training schedule
    Output: Trained PLM M, prompts p, masked-language-model head f, and discriminators g_1, ..., g_m
    Initialize M, p, f, and each g_l
    Repeat until the training schedule ends:
        For each source domain l from 1 to m:
            Compute L_class on labeled reviews from S_l
            Compute L_domain for the pair (S_l, T) using discriminator g_l
            Update f to minimize L_class
            Update g_l to minimize L_domain
            Update M and p to minimize lambda * L_class - L_domain
    Return M, p, f, and g_1, ..., g_m

    In the reported experiments, training ran for 10 epochs with batch size 2 and Adam; the learning rate was 2×10−52\times10^{-5} for PLM optimization and 5×10−55\times10^{-5} for the domain discriminators.

  6. Knowl 6 — Amazon review experiments evaluate 12 single-source and 4 multi-source transfers

    experimental setup

    The experiments use the Amazon reviews dataset with four domains: Books (B), DVDs (D), Electronics (E), and Kitchen appliances (K). Each domain has 2,000 manually labeled reviews—1,000 positive and 1,000 negative—and 4,000 unlabeled reviews. The authors randomly selected 20% of each domain’s examples for development and report mean accuracy over five-fold cross-validation. Single-source evaluation uses all 12 ordered pairs of distinct domains. Multi-source evaluation uses three domains as sources and the remaining domain as target, yielding BDE→K, BDK→E, BEK→D, and DEK→B. The PLM is a 12-layer Transformer initialized from RoBERTaBASE. Training uses Adam for 10 epochs, batch size 2, and learning rates of 2×10−52\times10^{-5} for the PLM and 5×10−55\times10^{-5} for domain discriminators. Accuracy is the evaluation metric.

  7. Knowl 7 — Single-source results show AdSPT’s average accuracy is 93.14%

    data/table

    The following mean accuracies (%) compare single-source transfer methods on the 12 directed Amazon-domain pairs, averaged over five-fold cross-validation. FT is source-only fine-tuning; AT is adversarial training; PT(HARD) uses the manually specified template “It is [MASK]”; PT(SOFT) uses soft prompts; AdSPT is soft-prompt tuning with adversarial training. AdSPT has the best average, 93.14%, exceeding SENTIX-Fix by 0.46 percentage points and BERT-DAAT by 3.02 points. It is best on most transfer pairs; for E→D, PT(SOFT) scores 93.25% versus AdSPT’s 93.15%.

    Source →\to target BERT-DAAT SENTIX-Fix FT FT+AT PT(HARD) PT(HARD)+AT PT(SOFT) AdSPT
    B →\to D 89.70 91.30 88.96 89.70 89.75 90.75 90.50 92.00
    B →\to E 89.57 93.25 86.15 87.30 91.75 92.45 93.05 93.75
    B →\to K 90.75 96.20 89.05 89.55 91.90 92.70 92.75 93.10
    D →\to B 90.86 91.15 89.40 89.55 90.90 91.50 91.75 92.15
    D →\to E 89.30 93.55 86.55 86.05 91.75 92.75 93.55 94.00
    D →\to K 87.53 96.00 87.53 87.69 91.05 92.35 92.50 93.25
    E →\to B 88.91 90.40 86.50 87.15 90.00 91.90 91.90 92.70
    E →\to D 90.13 91.20 87.98 88.20 92.10 92.55 93.25 93.15
    E →\to K 93.18 96.20 91.60 91.91 92.90 93.55 93.95 94.75
    K →\to B 87.98 89.55 87.55 87.65 89.15 90.75 91.75 92.35
    K →\to D 88.81 89.85 87.30 87.72 90.05 91.00 91.35 92.55
    K →\to E 91.72 93.55 90.45 90.25 92.15 92.50 93.10 93.95
    Average 90.12 92.68 88.25 88.56 91.12 92.06 92.45 93.14

    The table also shows that soft prompts improve over their hard-prompt counterparts: PT(SOFT) averages 92.45% versus 91.12% for PT(HARD), and AdSPT averages 93.14% versus 92.06% for PT(HARD)+AT.

  8. Knowl 8 — Multi-source results give AdSPT a 93.75% average accuracy

    data/table

    These mean accuracies (%) compare fine-tuning and prompt-tuning variants on the four three-source-to-one-target tasks in the Amazon reviews experiments. AdSPT is highest in each task and averages 93.75%, compared with 92.94% for soft prompts without adversarial training, 92.00% for hard prompts with adversarial training, and 89.67% for ordinary fine-tuning. The results show gains from both soft prompts and adversarial training in this multi-source setting.

    Source →\to target FT FT+AT PT(HARD) PT(HARD)+AT PT(SOFT) AdSPT
    BDE →\to K 89.70 91.30 91.50 92.25 93.25 93.75
    BDK →\to E 90.57 91.25 91.30 93.00 93.75 94.25
    BEK →\to D 88.56 89.05 90.75 91.25 92.00 93.50
    DEK →\to B 89.86 91.75 92.00 92.25 92.75 93.50
    Average 89.67 90.84 91.39 92.00 92.94 93.75

    FT denotes source-only fine-tuning, AT adversarial training, PT(HARD) hard-prompt tuning, PT(SOFT) soft-prompt tuning, and AdSPT soft-prompt tuning with adversarial training. The domain abbreviations are B: Books, D: DVDs, E: Electronics, and K: Kitchen appliances.

  9. Knowl 9 — Multi-source transfer usually improves on the best single-source result

    empirical result

    The reported cross-setting comparison finds higher AdSPT accuracy with three source domains than with the best single source for target domains B, D, and E: 93.50% versus 92.70% for B, 93.50% versus 93.15% for D, and 94.25% versus 94.00% for E. Kitchen appliances (K) is the exception: single-source E→K reaches 94.75%, above the multi-source BDE→K result of 93.75%. The authors attribute this exception to the feature distributions of Electronics and Kitchen appliances being closer to each other. The plotted comparison on page 8 likewise shows the multi-source advantage for B, D, and E, but not K.

  10. Knowl 10 — Prompt length affects performance, with three soft-prompt tokens best in the tested sweep

    empirical result

    AdSPT’s prompt-length analysis varies the number kk of soft-prompt tokens on the multi-source BDE→K and single-source B→K tasks. The plotted sweep on page 8 evaluates six lengths, k∈{2,3,4,5,6,10}k\in\{2,3,4,5,6,10\}, and reports the best performance at k=3k=3 for both settings. The results indicate that performance is sensitive to prompt length; adding more prompt tokens does not necessarily improve transfer. The main experiments therefore use three soft-prompt tokens.

Coverage note — No substantial contributed component was omitted; the paper states no separate limitations, and background, related work, and acknowledgements are outside the contribution.

References

  1. 1.Eyal Ben-David, Nadav Oved, and Roi Reichart. 2021. Pada: A prompt-based autoregressive approach for adaptation to unseen domains.
  2. 2.Shai Ben-David, John Blitzer, Koby Crammer, Fernando Pereira, et al. 2007. Analysis of representations for domain adaptation. Advances in Neural Information Processing Systems, 19:137.
  3. 3.John Blitzer, Mark Dredze, and Fernando Pereira. 2007. Biographies, bollywood, boom-boxes and blenders: Domain adaptation for sentiment classification. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 440–447.
  4. 4.Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. pages 1877–1901.
  5. 5.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.
  6. 6.Chunning Du, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. 2020. Adversarial and domain-aware bert for cross-domain sentiment analysis. In Proceedings of the 58th annual meeting of the Association for Computational Linguistics, pages 4019–4028.
  7. 7.Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030.
  8. 8.Tianyu Gao, Adam Fisch, and Danqi Chen. 2020. Making pre-trained language models better few-shot learners. pages 3816–3830.
  9. 9.Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. 2021. Ptr: Prompt tuning with rules for text classification. arXiv preprint arXiv:2105.11259.
  10. 10.Shengding Hu, Ning Ding, Huadong Wang, Zhiyuan Liu, Juanzi Li, and Maosong Sun. 2021. Knowledgeable prompt-tuning: Incorporating knowledge into prompt verbalizer for text classification. arXiv preprint arXiv:2108.02035.
  11. 11.Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691.
  12. 12.Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. pages 4582–4597.
  13. 13.Zheng Li, Xin Li, Ying Wei, Lidong Bing, Yu Zhang, and Qiang Yang. 2019. Transferable end-to-end aspect-based sentiment analysis with selective adversarial learning. arXiv preprint arXiv:1910.14192.
  14. 14.Zheng Li, Ying Wei, Yu Zhang, and Qiang Yang. 2018. Hierarchical attention transfer network for cross-domain sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.
  15. 15.Zheng Li, Yun Zhang, Ying Wei, Yuxiang Wu, and Qiang Yang. 2017. End-to-end adversarial memory network for cross-domain sentiment classification. In IJCAI, pages 2237–2243.
  16. 16.Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2021a. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602.
  17. 17.Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021b. Gpt understands, too. arXiv:2103.10385.
  18. 18.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. arXiv preprint arXiv:1907.11692.
  19. 19.Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. 2009. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430.
  20. 20.Sinno Jialin Pan and Qiang Yang. 2009. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359.
  21. 21.Minlong Peng, Qi Zhang, Yu-gang Jiang, and Xuan-Jing Huang. 2018. Cross-domain sentiment classification with target domain specific information. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2505–2513.
  22. 22.Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. 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 2227–2237, New Orleans, Louisiana. Association for Computational Linguistics.
  23. 23.Xiaoye Qu, Zhikang Zou, Yu Cheng, Yang Yang, and Pan Zhou. 2019. Adversarial category alignment network for cross-domain sentiment classification. 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 2496–2508.
  24. 24.Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training.
  25. 25.Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.
  26. 26.Sebastian Ruder and Barbara Plank. 2017. Learning to select data for transfer learning with bayesian optimization. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 372–382.
  27. 27.Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. 2018. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723–3732.
  28. 28.Timo Schick, Helmut Schmid, and Hinrich Schütze. 2020. Automatically identifying words that can serve as labels for few-shot text classification. pages 5569–5578.
  29. 29.Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few shot text classification and natural language inference. pages 255–269.
  30. 30.Timo Schick and Hinrich Schütze. 2021. It’s not just size that matters: Small language models are also few-shot learners. pages 2339–2352.
  31. 31.Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts.
  32. 32.Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008.
  33. 33.Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou, and Daniel Cer. 2021. Spot: Better frozen model adaptation through soft prompt transfer. arXiv preprint arXiv:2110.07904.
  34. 34.Jianfei Yu and Jing Jiang. 2016. Learning sentence embeddings with auxiliary tasks for cross-domain sentiment classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 236–246.
  35. 35.Han Zhao, Shanghang Zhang, Guanhang Wu, José MF Moura, Joao P Costeira, and Geoffrey J Gordon. 2018. Adversarial multiple source domain adaptation. Advances in Neural Information Processing Systems, 31:8559–8570.
  36. 36.Jie Zhou, Junfeng Tian, Rui Wang, Yuanbin Wu, Wenming Xiao, and Liang He. 2020. Sentix: A sentiment-aware pre-trained model for cross-domain sentiment analysis. In Proceedings of the 28th International Conference on Computational Linguistics, pages 568–579.
  37. 37.Xiaojin Zhu and Andrew B Goldberg. 2009. Introduction to semi-supervised learning. Synthesis lectures on artificial intelligence and machine learning, 3(1):1–130.
  38. 38.Yftah Ziser and Roi Reichart. 2018. Pivot based language modeling for improved neural domain adaptation. 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 1241–1251.

Citation

MLA
Wu, H., and X. Shi. “Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 2438–47, https://doi.org/10.18653/V1/2022.ACL-LONG.174.
APA
Wu, H., & Shi, X. (2022). Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2438–2447. https://doi.org/10.18653/V1/2022.ACL-LONG.174
Chicago
Wu, H., and X. Shi. 2022. “Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2438–47. https://doi.org/10.18653/V1/2022.ACL-LONG.174.
Harvard
Wu, H. and Shi, X. (2022) “Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis”, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 2438–2447. Available at: https://doi.org/10.18653/V1/2022.ACL-LONG.174.
Vancouver
1. Wu H, Shi X (2022) Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 2438–2447

BibTeX

@inproceedings{Wu_2022, title={Adversarial Soft Prompt Tuning for Cross-Domain Sentiment Analysis}, url={http://dx.doi.org/10.18653/V1/2022.ACL-LONG.174}, DOI={10.18653/v1/2022.acl-long.174}, booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, publisher={Association for Computational Linguistics}, author={Wu, Hui and Shi, Xiaodong}, year={2022}, pages={2438–2447} }
Metadata:Crossref

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/