Text Style Transfer via Learning Style Instance Supported Latent Space
Authors: Xiaoyuan Yi, Zhenghao Liu, Wenhao Li, Maosong Sun
IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three transfer tasks, sentiment modification, formality rephrasing, and poeticness generation, show that Sty Ins obtains a better balance between content and style, outperforming several recent baselines. |
| Researcher Affiliation | Academia | Xiaoyuan Yi1,2,3 , Zhenghao Liu1,2,3 , Wenhao Li1 and Maosong Sun1,2,4 1Department of Computer Science and Technology, Tsinghua University 2Institute for Artificial Intelligence, Tsinghua University 3State Key Lab on Intelligent Technology and Systems, Tsinghua University 4Jiangsu Collaborative Innovation Center for Language Ability, Jiangsu Normal University {yi-xy16, liu-zh16, liwh16}@mails.tsinghua.edu.cn, sms@tsinghua.edu.cn |
| Pseudocode | Yes | Algorithm 1 Training Process 1: for number of iterations do 2: Sample a source style si and a target style sj; 3: Sample instances Φi K from Di and Φj K from Dj; 4: Sample x from Di, x / Φi K; 5: Accumulate Lrecon, Lcycle, Lstyle; 6: if y exits then 7: Accumulate Lsuper; 8: end if 9: Update the parameters of G; 10: for n C steps do 11: Update the parameters of C with LC; 12: end for 13: end for |
| Open Source Code | Yes | Our source code is available at github.com/Xiaoyuan Yi/Sty Ins. |
| Open Datasets | Yes | We use the Yelp dataset processed by [Li et al., 2018], which consists of restaurant reviews with two sentiments, namely negative and positive. Formality Rephrasing. The recently released dataset GYAFC [Rao and Tetreault, 2018] contains paired formal and informal sentences in two domains. We use the Family & Relationships domain. Poeticness Generation. We also consider Chinese poeticness generation, as in [Shang et al., 2019], which seeks to transfer a vernacular sentence to a classical poetic one. ... We build a corpus called Chinese Poetic and Vernacular Text (CPVT)... |
| Dataset Splits | Yes | Table 1: Data Statistics. Dataset Styles Paired Unpaired Train Valid Test Train Valid Yelp Neg. N/A N/A 500 180k 2,000 Pos. 500 270k 2,000 GYAFC Inf. 52k 2,788 1,332 N/A N/A For. 52k 2,247 1,019 CPVT Ver. 4k 1,000 2,000 200k 10k Poe. 4k 1,000 2,000 200k 10k |
| Hardware Specification | No | The paper does not provide specific details on hardware used for experiments, such as GPU/CPU models or memory specifications. |
| Software Dependencies | No | The paper mentions software components like NLTK, Ken LM, and BERT, but does not provide specific version numbers for these or other dependencies required for replication. |
| Experiment Setup | Yes | We set word embedding size, hidden state size, the number of style instances K and the length of generative flow chain T to 256, 512, 10 and 6 respectively. The encoder and decoder share the same word embedding. The prior and posteriori distributions of z in Eq. (13) share parameters to reduce model size. The discriminator is a Convolutional Neural Network (CNN) based classifier with Spectral Normalization [Miyato et al., 2018]. To handle the discrete nature of sentences, as in [Dai et al., 2019], we multiply the softmax distribution by the word embedding matrix, to get a soft generated word and feed this weighted embedding to the discriminator. Adam with mini-batches (batch size=64) is used for optimization. |