Back

Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation

Kexin YangDayiheng LiuWenqiang LeiBaosong YangMingfeng XueBoxing ChenJun Xie

article2023ACL57 citations

Proposes a parameter-efficient framework that controls language models for single- and multi-attribute text generation by learning lightweight continuous prompts and prompt connectors while adding only 0.08% extra parameters to a frozen GPT-2.

Listen

Controlling the output of large language models to satisfy specific attributes, such as tone, emotion, or topic, is crucial for modern automated writing systems. However, current techniques present significant operational trade-offs. Fully fine-tuning a base model requires extensive computing and storage costs because a separate copy of the model must be maintained for every attribute. Conversely, guiding models at run-time with external attribute classifiers slows down text generation and frequently damages output fluency.

The article introduces Tailor, a lightweight continuous soft-prompt framework designed to steer a frozen GPT-2 model across single-attribute and multi-attribute generation tasks. The primary objective is to demonstrate that parameter-efficient soft prompts—continuous task-specific input vectors—can steer generation accurately while retaining fluency, boosting inference speed, and enabling multi-attribute control without fine-tuning full models.

The researchers evaluated Tailor through empirical experiments across single-attribute and multi-attribute generation benchmarks using Yelp restaurant reviews and cross-domain datasets (SST-2 and AG News). In Tailor, each attribute is trained as a single continuous vector prefix while keeping the underlying base model parameters fixed. For multi-attribute tasks, the authors tested a non-training combination approach—utilizing an attention mask (MAP mask) and a re-indexing position sequence (RP sequence) to avoid cross-attention and order sensitivity—alongside a training-based Multi-Attribute Prompt connector (MAP connector) trained using pseudo-labeled attribute data. The evaluation benchmarked correctness, fluency, perplexity, text diversity, inference speed, and human quality ratings against standard fine-tuning, adapter modules, and classifier-guided methods like GeDi and PPLM.

The evaluation yielded several key findings. First, Tailor achieves highly competitive attribute control using only 0.08% of the trainable parameters needed for full fine-tuning. In multi-attribute generation, Tailor's connector approach achieved an average correctness score of 87.15%, significantly outperforming full fine-tuning at 69.80% and adapters at 69.10%. Second, Tailor operates dramatically faster during generation than classifier-guided alternatives, processing a sample in 0.758 seconds compared to 1.680 seconds for GeDi and 15.553 seconds for PPLM (roughly a 20-fold speedup over PPLM). Third, the non-training approach effectively eliminated position sensitivity during prompt concatenation, boosting multi-attribute correctness from 76.20% to 78.82%. Finally, Tailor demonstrated strong generalization in low-data regimes and on previously unseen attribute pairings, such as positive sentiment combined with Mexican food topics.

These findings indicate that organizations can implement highly customizable, multi-attribute text generation without incurring high infrastructure, storage, or computational costs. By freezing the underlying base model and training only modular soft prompts, engineering teams reduce retraining overhead and runtime latency while preserving text quality and fluency.

Decision-makers and practitioners deploying controllable language models should adopt modular soft-prompt strategies like Tailor for multi-attribute generation workflows. If zero extra training is preferred, the non-training mask and position re-indexing approach serves as an effective plug-and-play solution. Where maximum attribute fidelity is required, investing minimal compute to train a prompt connector provides the best performance trade-off.

The primary limitation of the article is its focus on two-attribute combinations and evaluations centered mainly on the GPT-2 base architecture. While confidence in the reported performance metrics is high given the combined automatic and human evaluations, further testing across diverse model architectures and scaling to combinations of three or more simultaneous attributes are necessary before deploying in broader production settings.

Alibaba GroupNational University of Singaporeattribute dataclassifiersinferenceattribute-based controlled text generationprompt tuningmulti-attribute generation
arXiv: 2204.13362yangkexin/Tailor

No sufficiently relevant recommendations were found.

Cover for Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation

Abstract

Attribute-based Controlled Text Generation (CTG) refers to generating sentences that satisfy desirable attributes (e.g., emotions and topics). Existing work usually utilize fine-tuning or resort to extra attribute classifiers, yet suffer from increases in storage and inference time. To address these concerns, we explore attribute-based CTG in a parameter-efficient manner. In short, the proposed Tailor represents each attribute as a pre-trained continuous vector (i.e., single-attribute prompt), which guides the generation of a fixed pre-trained language model (PLM) to satisfy a pre-specified attribute. These prompts can be simply concatenated as a whole for multi-attribute CTG without any re-training. Nevertheless, this may raise problems of fluency downgrading and position sensitivity. To solve this, Tailor provides two solutions to enhance the combination. The former contains a multi-attribute prompt mask and a re-indexing position sequence to bridge the gap between the training (one single-attribute prompt for each task) and the testing stage (concatenating two prompts). The latter introduces a trainable prompt connector to further enhance the combinations. Experiments demonstrate that, only requiring 0.08% extra training parameters of the GPT-2, Tailor can achieve effective and general improvements on eleven attribute-specific generation tasks.

Topics

attribute dataclassifiersinferencelanguage modelprobabilistic inferencetuning methodcontrol parameters

Keywords

attribute-based controlled text generationprompt tuningmulti-attribute generationsingle-attribute promptsmulti-attribute prompt maskattribute classifiersattribute combination

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Methodology
  • 3.1 Tailor for Single-Attribute CTG
  • 3.2 Tailor for Multi-Attribute CTG
  • 3.2.1 Non-Training Method
  • 3.2.2 Training Method
  • 4 Experiments
  • 4.1 Experimental Setup
  • 4.2 Main Results
  • 4.3 Further Discussions
  • 5 Conclusions
  • Limitations
  • Ethics Statement
  • References
  • A Implement Details
  • B Yelp Dataset
  • C Classifiers For Correctness Evaluation
  • D Human Evaluation Details
  • E Experiments on Cross Domain Dataset
  • F Ablations
  • G Case Study
  • ACL 2023 Responsible NLP Checklist

Knowls

  1. Knowl 1 — Attribute-specific continuous prompts control a frozen language model

    model/method

    Tailor represents each desired attribute with its own trainable continuous prompt and uses those prompts to control a fixed GPT-2 language model. For an attribute-specific training sentence with tokens x1,…,xnx_1,\ldots,x_n, the prompt SkS_k is a matrix of lkl_k vectors, each of dimension dembd_{\mathrm{emb}}, the GPT-2 embedding dimension. The prompt vectors are randomly initialized and prepended to the sentence embeddings. Training maximizes the sentence log-likelihood while updating only the prompt parameters; the GPT-2 weights remain fixed:

    max⁡Sk  ∑t=1nlog⁡Pθg(xt∣Sk,x<t),\max_{S_k}\;\sum_{t=1}^{n}\log P_{\theta_g}(x_t\mid S_k,x_{<t}),

    where θg\theta_g denotes the frozen GPT-2 parameters and x<tx_{<t} denotes the sentence tokens preceding xtx_t. A trained prompt can be used as a plug-in for generating text with its associated attribute. The reported learning rate for the single-attribute prompts was 5×10−55\times10^{-5} with zero warm-up steps.

  2. Knowl 2 — MAP connector trains prompts to compose two attributes

    model/method

    Tailor's Multi-Attribute Prompt (MAP) connector is a trainable continuous prompt intended to help two single-attribute prompts work together. During connector training, each single-attribute sentence is paired with its real attribute prompt and a pseudo-prompt for another attribute, and the sequence of the two prompts, the connector, and the sentence is fed to a frozen GPT-2. Only the connector parameters are updated, using the sentence language-modeling objective:

    max⁡C  ∑t=1nlog⁡Pθg(xt∣Su,Sv,C,x<t),\max_C\;\sum_{t=1}^{n}\log P_{\theta_g}(x_t\mid S_u,S_v,C,x_{<t}),

    where SuS_u and SvS_v are the two attribute prompts, CC is the connector, x1,…,xnx_1,\ldots,x_n are the training sentence tokens, and θg\theta_g denotes the fixed GPT-2 parameters. At inference, the connector is placed between the two prompts and the input prefix; the resulting sequence conditions generation for the requested attribute pair. The connector length used in the experiments was 128 vectors, and its learning rate was 5×10−55\times10^{-5} with zero warm-up steps.

  3. Knowl 3 — Pseudo-prompts enable connector training without paired multi-attribute examples

    algorithm

    Tailor trains the MAP connector from single-attribute data by constructing a pseudo-prompt to pair with each sentence's real attribute prompt. A classifier trained on the single-attribute data predicts a probability distribution over the relevant prompt set S={S1,…,Sm}S=\{S_1,\ldots,S_m\}. For classifier probabilities p1,…,pmp_1,\ldots,p_m, the argmax method selects the prompt associated with the most probable class, while the weighted method forms an element-wise weighted sum:

    Sargmax=Sargmax⁡zpz,Sweighted=∑z=1mpzSz.S_{\mathrm{argmax}}=S_{\operatorname{argmax}_z p_z}, \qquad S_{\mathrm{weighted}}=\sum_{z=1}^{m}p_zS_z.

    The selected or weighted pseudo-prompt is paired with the sentence's real prompt to train the connector. This procedure supplies connector training with prompt pairs even though examples of the desired multi-attribute combinations are unavailable.

  4. Knowl 4 — MAP mask blocks attention from one attribute prompt to the other

    model/method

    When two single-attribute prompts are concatenated, Tailor's non-training method adds an additive mask to the GPT-2 attention logits. For the sequence consisting of prompt SuS_u, prompt SvS_v, and an input prefix, the mask assigns −∞-\infty to attention entries where a position in the second prompt queries a position in the first prompt; all other mask entries are zero. Thus, the second prompt cannot attend to the first, matching the separate-prompt training condition, while the input tokens can attend to the prompts under the usual causal attention rule. The mask is applied in the attention softmax, so the blocked attention weights are zero. This method requires no additional training.

  5. Knowl 5 — Re-indexing position IDs makes prompt order swaps position-stable

    model/method

    Tailor's Re-indexing Position (RP) sequence addresses the change in GPT-2 position embeddings that otherwise occurs when two prompts are concatenated in a different order. For two prompts of equal length ll followed by an input prefix of nn tokens, Tailor assigns position IDs 1,…,l1,\ldots,l to the first prompt, reuses 1,…,l1,\ldots,l for the second prompt, and assigns l+1,…,l+nl+1,\ldots,l+n to the prefix. Reusing the same position IDs for both prompts keeps their position information stable when their order is swapped. In Tailor's non-training method, RP is used together with the MAP mask.

  6. Knowl 6 — MAP training improves multi-attribute Yelp generation with few trainable parameters

    empirical result

    On Yelp restaurant reviews, Tailor was evaluated on six combinations of positive or negative sentiment with Mexican, American, or Asian food topics. Each attribute had 30,000 training and 3,000 validation sentences; evaluation used 15 attribute-unrelated prefixes with 100 generated continuations per prefix. Correctness was the fraction classified as containing the requested attributes; text quality was measured by grammaticality probability (GRAM) and perplexity (PPL), and diversity by distinct nn-gram rates (Dist-1/2/3). TailorArgmax, which trains the MAP connector with argmax pseudo-prompts, achieved 87.15% average correctness (92.97% for sentiment and 81.32% for food topic), versus 69.80% for sequential fine-tuning and 81.71% for an Adapter trained with pseudo-labels. TailorWeight achieved 83.98% average correctness. TailorArgmax used 0.08% of the fine-tuning parameter count; its GRAM was 0.69, PPL was 52.73, and Dist-1/2/3 were 0.05/0.33/0.69. The corresponding scores for TailorConcat, the non-training variant, were 78.82% average correctness, GRAM 0.63, PPL 52.76, and diversity 0.05/0.32/0.68. In human ratings on a 1–5 scale, TailorArgmax scored 4.57 for quality and 2.37 for attribute alignment, compared with 4.79 and 1.91 for Adapter with pseudo-labels and 4.67 and 1.74 for fine-tuning.

  7. Knowl 7 — Single-attribute prompts improve correctness and human-rated attribute alignment

    empirical result

    In Yelp single-attribute generation, TailorSingle used 0.08% of the parameters used by full fine-tuning. For food-topic generation it achieved 83.89% correctness, compared with 74.70% for an Adapter with 0.10% trainable parameters; for sentiment generation it achieved 93.80%, compared with the Adapter's 93.32%. TailorSingle's GRAM/PPL scores were 0.71/45.79 for food and 0.71/46.20 for sentiment; its Dist-1/2/3 scores were 0.05/0.35/0.71 and 0.06/0.35/0.70, respectively. Human evaluation on a 1–5 scale gave TailorSingle quality and attribute-alignment scores of 4.62 and 3.04, versus 4.69 and 2.97 for fine-tuning; the summed scores were 7.66 for both. Thus, the human ratings favored TailorSingle for attribute alignment even though its automatic correctness was below fine-tuning (87.53%/97.95% for food/sentiment).

  8. Knowl 8 — Both components of non-training prompt composition contribute to correctness

    empirical result

    An ablation on the six Yelp attribute combinations compared TailorConcat, which uses both the MAP mask and RP sequence, with versions that remove one or both components. Average correctness was 78.82% with both components, 78.36% without the MAP mask, 77.77% without RP, and 76.20% without either. The corresponding sentiment/food correctness scores were 87.54%/70.10%, 87.39%/69.34%, 88.33%/67.21%, and 87.88%/64.52%. TailorConcat also returned the same scores after swapping the prompt order, whereas simple concatenation was position-sensitive. These results support using the mask and re-indexing together to improve combination performance and reduce order sensitivity.

  9. Knowl 9 — The trained connector transfers to a held-out attribute combination

    empirical result

    To test whether MAP connector training generalized beyond combinations seen during training, the positive-sentiment plus Mexican-food combination was removed from connector training and evaluated separately on Yelp. TailorArgmax achieved 89.89% average correctness on this unseen combination, including 97.07% sentiment correctness and 82.70% food-topic correctness. TailorConcat achieved 87.54% average correctness (95.60% sentiment and 79.47% food), so TailorArgmax was higher by 2.35 percentage points on average. TailorArgmax scored 91.64% average correctness when evaluated on combinations without a held-out pair.

  10. Knowl 10 — Cross-domain experiments show mixed correctness results and stronger diversity

    empirical result

    Tailor was also tested on cross-domain combinations pairing sentiment prompts trained on SST-2 with topic prompts trained on AG News. Scores were averaged over eight combinations: two sentiment attributes and four news topics. TailorConcat achieved 63.38% average correctness (68.08% sentiment and 58.67% news-topic correctness), GRAM 0.59, PPL 36.82, and Dist-1/2/3 of 0.11/0.48/0.80. TailorArgmax achieved 61.42% average correctness (63.65% sentiment and 59.18% news-topic correctness), GRAM 0.68, PPL 35.33, and diversity 0.13/0.53/0.84, using 0.08% of the fine-tuning parameter count. The results show that connector training increased grammaticality and diversity relative to TailorConcat, while average correctness was not higher than the fine-tuning baseline's 62.54%.

  11. Knowl 11 — TailorSingle has lower measured inference time than classifier-guided baselines

    empirical result

    In the reported inference-speed comparison, generation took 0.758 seconds per sample for TailorSingle, 1.680 seconds per sample for GeDi, and 15.553 seconds per sample for PPLM. The paper reports these as relative speeds of 1.00×, 0.45×, and 0.05×, respectively, with higher relative speed indicating faster generation. This comparison supports Tailor's use of prompt-based control without per-generation classifier-guidance procedures.

  12. Knowl 12 — Evaluation is limited to two-attribute combinations and a narrow set of language models

    limitation

    The experiments study multi-attribute generation by combining two attributes at a time, so they do not establish how Tailor behaves when more than two attributes must be composed. The authors also identify evaluation on a wider variety of pretrained language models as an open need; the reported experiments primarily use GPT-2. Consequently, the results do not establish that the approach scales to larger attribute combinations or generalizes across a broad range of language-model architectures.

Coverage note — The few-shot evaluation and qualitative generation examples are omitted because they provide supplementary robustness checks and illustrations rather than distinct core methods or findings.

References

  1. 1.Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In ICLR 2020. OpenReview.net.
  2. 2.Zi-Yi Dou, Pengfei Liu, Hiroaki Hayashi, Zhengbao Jiang, and Graham Neubig. 2021. Gsum: A general framework for guided neural abstractive summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 4830–4842. Association for Computational Linguistics.
  3. 3.Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.
  4. 4.Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. 2021. PPT: pre-trained prompt tuning for few-shot learning. CoRR, abs/2109.04332.
  5. 5.Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. 2021. PTR: prompt tuning with rules for text classification. CoRR, abs/2105.11259.
  6. 6.Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomás Mikolov. 2017. Bag of tricks for efficient text classification. In EACL 2017, pages 427–431. Association for Computational Linguistics.
  7. 7.Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard Socher. 2019. CTRL: A conditional transformer language model for controllable generation. CoRR, abs/1909.05858.
  8. 8.Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq R. Joty, Richard Socher, and Nazneen Fatema Rajani. 2021. Gedi: Generative discriminator guided sequence generation. In Findings of EMNLP 2021, pages 4929–4952. Association for Computational Linguistics.
  9. 9.Guillaume Lample, Sandeep Subramanian, Eric Michael Smith, Ludovic Denoyer, Marc’Aurelio Ranzato, and Y-Lan Boureau. 2019. Multiple-attribute text rewriting. In ICLR 2019. OpenReview.net.
  10. 10.Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2015. A diversity-promoting objective function for neural conversation models. arXiv preprint arXiv:1510.03055.
  11. 11.Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In ACL 2021, pages 4582–4597. Association for Computational Linguistics.
  12. 12.Zhaojiang Lin, Andrea Madotto, and Pascale Fung. 2020. Exploring versatile generative language model via parameter-efficient transfer learning. In Findings of EMNLP 2020, volume EMNLP 2020 of Findings of ACL, pages 441–459. Association for Computational Linguistics.
  13. 13.Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2021a. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. CoRR, abs/2107.13586.
  14. 14.Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In ACL 2022.
  15. 15.Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021b. GPT understands, too. CoRR, abs/2103.10385.
  16. 16.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.
  17. 17.Yiwei Lyu, Paul Pu Liang, Hai Pham, Eduard H. Hovy, Barnabás Póczos, Ruslan Salakhutdinov, and Louis-Philippe Morency. 2021. Styleptb: A compositional benchmark for fine-grained controllable text style transfer. In NAACL-HLT 2021, pages 2116–2138. Association for Computational Linguistics.
  18. 18.Jing Qian, Li Dong, Yelong Shen, Furu Wei, and Weizhu Chen. 2022. Controllable natural language generation with contrastive prefixes. arXiv preprint arXiv:2202.13257.
  19. 19.Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.
  20. 20.Elisabeth Rudolph. 1989. The role of conjunctions and particles for text connexity. In Text and discourse connectedness, page 175. John Benjamins.
  21. 21.Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. 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 6830–6841.
  22. 22.Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642.
  23. 23.Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625–641.
  24. 24.Johannes Welbl, Amelia Glaese, Jonathan Uesato, Sumanth Dathathri, John Mellor, Lisa Anne Hendricks, Kirsty Anderson, Pushmeet Kohli, Ben Coppin, and Po-Sen Huang. 2021. Challenges in detoxifying language models. In Findings of EMNLP 2021, pages 2447–2469. Association for Computational Linguistics.
  25. 25.Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics.
  26. 26.Kevin Yang and Dan Klein. 2021. FUDGE: controlled text generation with future discriminators. In NAACL-HLT 2021, pages 3511–3535. Association for Computational Linguistics.
  27. 27.Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. 2022. A survey of controllable text generation using transformer-based pre-trained language models. CoRR, abs/2201.05337.
  28. 28.Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28.
  29. 29.Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul F. Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. CoRR, abs/1909.08593.

Citation

MLA
Yang, K., et al. “Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation”. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 410–27, https://doi.org/10.18653/v1/2023.acl-long.25.
APA
Yang, K., Liu, D., Lei, W., Yang, B., Xue, M., Chen, B., & Xie, J. (2023). Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 410–427. https://doi.org/10.18653/v1/2023.acl-long.25
Chicago
Yang, K., D. Liu, W. Lei, et al. 2023. “Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation”. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 410–27. https://doi.org/10.18653/v1/2023.acl-long.25.
Harvard
Yang, K. et al. (2023) “Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation”, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 410–427. Available at: https://doi.org/10.18653/v1/2023.acl-long.25.
Vancouver
1. Yang K, Liu D, Lei W, Yang B, Xue M, Chen B, Xie J (2023) Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 410–427

BibTeX

@inproceedings{yang-etal-2023-tailor,
    title = "Tailor: A Soft-Prompt-Based Approach to Attribute-Based Controlled Text Generation",
    author = "Yang, Kexin  and
      Liu, Dayiheng  and
      Lei, Wenqiang  and
      Yang, Baosong  and
      Xue, Mingfeng  and
      Chen, Boxing  and
      Xie, Jun",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.25/",
    doi = "10.18653/v1/2023.acl-long.25",
    pages = "410--427"
}
Metadata:ACL Anthology

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/