Summary of existing disentanglement-based methods and the setting they adopted, with a reference of their performance on the Yelp dataset. For the settings, we include the encoder-decoder training method (Enc-Dec) in Section 5.1.1, the disentanglement method (Disen.) in Section 5.1.2, and the loss types used to control style (Style Control) and content (Content Control) in Section 5.1.3. For the model performance, we report automatic evaluation scores including BLEU with the one human reference (BL-Ref) provided by Li et al. (2018), accuracy (Acc.), BLEU with the input (BL-Inp), and perplexity (PPL). * marks numbers reported by Liu et al. (2020). Readers can refer to Hu, Lee, and Aggarwal (2020) for more complete performance results on Yelp.
| . | Settings . | Performance on Yelp . | ||||||
|---|---|---|---|---|---|---|---|---|
| Enc-Dec . | Disen. . | Style Control . | Content Control . | BL-Ref . | Acc. (%) . | BL-Inp . | PPL↓ . | |
| Mueller, Gifford, and Jaakkola (2017) | VAE | LRE | – | – | – | – | – | – |
| Hu et al. (2017) | VAE | ACC | ACO | – | 22.3 | 86.7 | 58.4 | – |
| Shen et al. (2017) | AE&GAN | ACC | AdvR∥AdvO | – | 7.8 | 73.9 | 20.7 | 72* |
| Fu et al. (2018) | AE | ACC | AdvR | – | 12.9 | 46.9 | 40.1 | 166.5* |
| Prabhumoye et al. (2018) | AE | ACC | ACO | – | 6.8 | 87.2 | – | 32.8* |
| Zhao et al. (2018) | GAN | ACC | AdvR | – | – | 73.4 | 31.2 | 29.7 |
| Yang et al. (2018) | AE | ACC | LMO | – | – | 91.2 | 57.8 | 47.0&60.9 |
| Logeswaran, Lee, and Bengio (2018) | AE | ACC | AdvO | Cycle | – | 90.5 | – | 133 |
| Tian, Hu, and Yu (2018) | AE | ACC | AdvO | Noun | 24.9 | 92.7 | 63.3 | – |
| Liao et al. (2018) | VAE | LRE | – | – | – | 88.3 | – | – |
| Romanov et al. (2019) | AE | LRS | ACR&AdvR | – | – | – | – | – |
| John et al. (2019) | AE&VAE | LRS | ACR&AdvR | BoW&AdvBoW | – | 93.4 | – | – |
| Bao et al. (2019) | VAE | LRS | ACR&AdvR | BoW&AdvBoW | – | – | – | – |
| Dai et al. (2019) | AE | ACC | ACO | Cycle | 20.3 | 87.7 | 54.9 | 73 |
| Wang, Hua, and Wan (2019) | AE | LRE | – | – | 24.6 | 95.4 | – | 46.2 |
| Li et al. (2020) | GAN | ACC | ACO&AdvR | – | – | 95.5 | 53.3 | – |
| Liu et al. (2020) | VAE | LRE | – | – | 18.8 | 92.3 | – | 18.3 |
| Yi et al. (2020) | VAE | ACC | ACO | Cycle | 26.0 | 90.8 | – | 109 |
| Jin et al. (2020a) | AE | LRE | – | – | – | – | – | – |
| . | Settings . | Performance on Yelp . | ||||||
|---|---|---|---|---|---|---|---|---|
| Enc-Dec . | Disen. . | Style Control . | Content Control . | BL-Ref . | Acc. (%) . | BL-Inp . | PPL↓ . | |
| Mueller, Gifford, and Jaakkola (2017) | VAE | LRE | – | – | – | – | – | – |
| Hu et al. (2017) | VAE | ACC | ACO | – | 22.3 | 86.7 | 58.4 | – |
| Shen et al. (2017) | AE&GAN | ACC | AdvR∥AdvO | – | 7.8 | 73.9 | 20.7 | 72* |
| Fu et al. (2018) | AE | ACC | AdvR | – | 12.9 | 46.9 | 40.1 | 166.5* |
| Prabhumoye et al. (2018) | AE | ACC | ACO | – | 6.8 | 87.2 | – | 32.8* |
| Zhao et al. (2018) | GAN | ACC | AdvR | – | – | 73.4 | 31.2 | 29.7 |
| Yang et al. (2018) | AE | ACC | LMO | – | – | 91.2 | 57.8 | 47.0&60.9 |
| Logeswaran, Lee, and Bengio (2018) | AE | ACC | AdvO | Cycle | – | 90.5 | – | 133 |
| Tian, Hu, and Yu (2018) | AE | ACC | AdvO | Noun | 24.9 | 92.7 | 63.3 | – |
| Liao et al. (2018) | VAE | LRE | – | – | – | 88.3 | – | – |
| Romanov et al. (2019) | AE | LRS | ACR&AdvR | – | – | – | – | – |
| John et al. (2019) | AE&VAE | LRS | ACR&AdvR | BoW&AdvBoW | – | 93.4 | – | – |
| Bao et al. (2019) | VAE | LRS | ACR&AdvR | BoW&AdvBoW | – | – | – | – |
| Dai et al. (2019) | AE | ACC | ACO | Cycle | 20.3 | 87.7 | 54.9 | 73 |
| Wang, Hua, and Wan (2019) | AE | LRE | – | – | 24.6 | 95.4 | – | 46.2 |
| Li et al. (2020) | GAN | ACC | ACO&AdvR | – | – | 95.5 | 53.3 | – |
| Liu et al. (2020) | VAE | LRE | – | – | 18.8 | 92.3 | – | 18.3 |
| Yi et al. (2020) | VAE | ACC | ACO | Cycle | 26.0 | 90.8 | – | 109 |
| Jin et al. (2020a) | AE | LRE | – | – | – | – | – | – |