Professional Documents
Culture Documents
Secondpaper
Secondpaper
19
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 19–22
July 5, 2020.
2020
c Association for Computational Linguistics
https://doi.org/10.18653/v1/P17
allel corpus is available (Xu et al., 2012; Rao and rial also investigates stylized generative models in
Tetreault, 2018). This is similar to machine trans- non-NLP domains, and thus would inspire future
lation, but semi-supervised learning is adopted to NLP studies in this direction.
address small-data training (Wang et al., 2019b).
(2) Non-parallel style transfer, where each sen- 2 Tutorial Outline
tence is annotated by a style label (e.g., positive PART I: Introduction (20 min)
or negative sentiment). This setting is the most
explored setting in previous style transfer litera- • Definition of style
ture. We will discuss classification losses to dis-
tinguish different styles (John et al., 2019), and • Settings and Problem formulations
adversarial losses/cycle consistency to match con-
• Examples of style (e.g., sentiment, artistic
tent information (Shen et al., 2017). We will
style, grammatical style)
also present an editing-based approach that ed-
its style-specific words and phrases into the de- PART II: Style-Conditioned Text Generation
sired style (Li et al., 2018). (3) Unsupervised style (50 min)
transfer, where the entire corpus is unlabeled (no
parallel pairs or style labels). In recent studies, • Techniques
researchers have applied auxiliary losses (such as – Encoding style in embeddings: sequence-
orthogonality penalty) to detect the most prevalent to-sequence models with style embed-
variation of text in a corpus, and are sometimes dings, style conditioned language models,
able to accomplish style transfer in a purely unsu- Variational Autoencoder (VAE) with style
pervised fashion. Since unsupervised style trans- embeddings;
fer is new to NLP and less explored, we will also
– Encoding style and content in latent space:
introduce several studies in the computer vision
Conditional Variational Autoencoder
domain, bringing future opportunities to text gen-
(CVAE) Gaussian Mixture Variational
eration in this setting (Gatys et al., 2016; Chen
Autoencoder (GMM-VAE), Adversarially
et al., 2016).
Regularized Autoencoder (ARAE).
Next, we will discuss style adversarial text gen- – Models with multiple style-specific de-
eration (Zhang et al., 2019). The setting of ad- coders
versarial attacks is similar to style transfer in that
it aims to change the style classifier’s prediction. • Applications
However, the synthesized sentence in this setting
– Creative text generation (e.g., poetry com-
should in fact keep the actual style as humans per-
position)
ceive, but “fool” the style classifier. Thus, it is
known as the adversarial attack. We will dis- – Persona and emotion conditioned dialogue
cuss style adversarial generation in the charac- models
ter level, the word level, as well as the sentence – Stylized image caption generation
level. Techniques include discrete word manipu-
• Evaluation measures
lation and continuous latent space manipulation.
Finally, we will conclude our tutorial by pre- – Stylistic adherence
senting the challenges of stylized text generation – Content preservation
and discussing future directions, such as small- – Language fluency
data training, non-categorical style modeling, and – Novelty and diversity
a generalized scope of style transfer (e.g., control-
ling the syntax as a style (Bao et al., 2019)). PART III: Style-Transfer Text Generation
By the end of the tutorial, the audience will have (60 min)
a systematic view of different settings of stylized • Parallel supervised style transfer
text generation, understand common techniques to
model and manipulate the style of text, and be – Sequence-to-sequence learning
able to apply existing approaches to new scenar- – Semi-supervised training with limited par-
ios that require stylized text generation. Our tuto- allel data
20
– Applications: Shakespearean–modern En- interests include deep learning applied to natural
glish transfer, formality style transfer language processing as well as programming lan-
guage processing. Recently, he has been focusing
• Non-parallel supervised style transfer more on text generation, from both continuous la-
– Auxiliary classification for style modeling tent space and discrete word space. He has more
– Adversarial learning for matching content than 30 papers published at top-tier conferences
– Cycle consistency for content matching and journals, including AAAI, ACL, CIKM, COL-
– Edit-based style transfer ING, EMNLP, ICASSP, ICLR, ICML, IJCAI, IN-
TERSPEECH, NAACL-HLT, and TACL. He pre-
– Applications: Sentiment, genre and gram-
sented a tutorial “Discreteness in Neural Natural
matical style transfer
Language Processing” at EMNLP-IJCNLP’19.
• Unsupervised style transfer
Olga Vechtomova
– Approaches: Mutual information penal-
ovechtomova@uwaterloo.ca
ties and correlation penalties for automatic
https://ov-research.uwaterloo.ca
style detection
– A brief introduction of unsupervised style Dr. Olga Vechtomova is an Associate Professor
transfer in image domain (e.g., color, in the Department of Management Sciences, Fac-
shape, angle) ulty of Engineering, cross-appointed in the School
PART IV: Style-Advsersarial Text Generation of Computer Science at the University of Water-
(30 minutes) loo. Olga leads the Natural Language Processing
Lab, affiliated with the Waterloo.AI Institute. Her
• Style adversarial vs. style transfer research has been supported by a number of indus-
• Approaches try and government grants, including Amazon Re-
search Award and Natural Sciences and Engineer-
– Character-level attack ing Research Council (NSERC). The research in
– Word-level attack her Lab is mainly focused on designing deep neu-
– Sentence-level attack ral networks for natural language generation tasks.
Her current and recent projects include controlled
PART IV: Conclusion, Future Work, and Q&A
text generation, text style transfer, and designing
(20 min)
text generative models for creative applications.
• Challenges: non-categorical style, small-data She has over 50 publications in NLP and Informa-
training tion Retrieval conferences and journals, including
NAACL-HLT, COLING, ACL, ACM SIGIR, and
• A broader view of “style”: text summariza-
CIKM. She and her colleagues recently received
tion/simplification as style transfer, syntax-
the ACM SIGIR 2019 Test of Time Award.
semantic disentanglement
Acknowledgments
3 Instructors
Lili Mou We acknowledge the support of the Natural
doublepower.mou@gmail.com Sciences and Engineering Research Council of
https://lili-mou.github.io Canada (NSERC), under grant Nos. RGPIN-2019-
04897 and RGPIN-2020-04465. Lili Mou is also
Dr. Lili Mou is an Assistant Professor at the supported by AltaML, the Amii Fellow Program,
Department of Computing Science, University of and the Canadian CIFAR AI Chair Program.
Alberta. He is also an Amii Fellow and a Cana-
dian CIFAR AI Chair. Lili received his BS and References
PhD degrees in 2012 and 2017, respectively, from
School of EECS, Peking University. After that, Yu Bao, Hao Zhou, Shujian Huang, Lei Li, Lili Mou,
Olga Vechtomova, Xin-yu Dai, and Jiajun Chen.
he worked as a postdoctoral fellow at the Univer- 2019. Generating sentences from disentangled syn-
sity of Waterloo and a research scientist at Adept- tactic and semantic spaces. In ACL, pages 6008–
mind (a startup in Toronto, Canada). His research 6019.
21
Cheng-Kuan Chen, Zhufeng Pan, Ming-Yu Liu, and Bakhtiyar Syed, Gaurav Verma, Balaji Vasan Srini-
Min Sun. 2019. Unsupervised stylish image de- vasan, Vasudeva Varma, et al. 2019. Adapting
scription generation via domain layer norm. In Pro- language models for non-parallel author-stylized
ceedings of the AAAI Conference on Artificial Intel- rewriting. arXiv preprint arXiv:1909.09962.
ligence, volume 33, pages 8151–8158.
Alexey Tikhonov and Ivan P Yamshchikov. 2018.
Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Guess who? multilingual approach for the auto-
Ilya Sutskever, and Pieter Abbeel. 2016. InfoGAN: mated generation of author-stylized poetry. In IEEE
Interpretable representation learning by information Spoken Language Technology Workshop, pages
maximizing generative adversarial nets. In NIPS, 787–794.
pages 2172–2180.
Olga Vechtomova, Hareesh Bahuleyan, Amirpasha
Ghabussi, and Vineet John. 2018. Generating lyrics
Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan
with variational autoencoder and multi-modal artist
Zhao, and Rui Yan. 2018. Style transfer in text: Ex-
embeddings. arXiv preprint arXiv:1812.08318.
ploration and evaluation. In AAAI, pages 663–670.
Wenlin Wang, Zhe Gan, Hongteng Xu, Ruiyi Zhang,
Leon A Gatys, Alexander S Ecker, and Matthias Guoyin Wang, Dinghan Shen, Changyou Chen, and
Bethge. 2016. Image style transfer using convolu- Lawrence Carin. 2019a. Topic-guided variational
tional neural networks. In CVPR, pages 2414–2423. autoencoders for text generation. arXiv preprint
arXiv:1903.07137.
Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan
Salakhutdinov, and Eric P Xing. 2017. Toward con- Yunli Wang, Yu Wu, Lili Mou, Zhoujun Li, and Wen-
trolled generation of text. In ICML, pages 1587– han Chao. 2019b. Harnessing pre-trained neural
1596. networks with rules for formality style transfer. In
EMNLP-IJCNLP.
Vineet John, Lili Mou, Hareesh Bahuleyan, and Olga
Vechtomova. 2019. Disentangled representation Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xu-
learning for non-parallel text style transfer. In ACL, ancheng Ren, Houfeng Wang, and Wenjie Li. 2018.
pages 424–434. Unpaired sentiment-to-sentiment translation: A cy-
cled reinforcement learning approach. In ACL,
Jiwei Li, Michel Galley, Chris Brockett, Georgios Sp- pages 979–988.
ithourakis, Jianfeng Gao, and Bill Dolan. 2016. A
Peng Xu, Yanshuai Cao, and Jackie Chi Kit Che-
persona-based neural conversation model. In ACL,
ung. 2019. Unsupervised controllable text genera-
pages 994–1003.
tion with global variation discovery and disentangle-
ment. arXiv preprint arXiv:1905.11975.
Juncen Li, Robin Jia, He He, and Percy Liang. 2018.
Delete, retrieve, generate: a simple approach to sen- Wei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and
timent and style transfer. In NAACL-HLT, pages Colin Cherry. 2012. Paraphrasing for style. In COL-
1865–1874. ING, pages 2899–2914.
Yuan Li, Chunyuan Li, Yizhe Zhang, Xiujun Li, Guo- Xiaopeng Yang, Xiaowen Lin, Shunda Suo, and Ming
qing Zheng, Lawrence Carin, and Jianfeng Gao. Li. 2017. Generating thematic chinese poetry using
2020. Complementary auxiliary classifiers for label- conditional variational autoencoders with hybrid de-
conditional text generation. In AAAI. coders. arXiv preprint arXiv:1711.07632.
Sudha Rao and Joel Tetreault. 2018. Dear sir or Huangzhao Zhang, Hao Zhou, Ning Miao, and Lei Li.
madam, may I introduce the GYAFC dataset: Cor- 2019. Generating fluent adversarial examples for
pus, benchmarks and metrics for formality style natural languages. In ACL, pages 5564–5569, Flo-
transfer. In NAACL-HLT, pages 129–140. rence, Italy.
22