Style Transfer in Text: Exploration and Evaluation

Authors: Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, Rui Yan

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark our models and the evaluation metrics on two style transfer tasks: paper-news title transfer, and positive-negative review transfer. Results show that the proposed content preservation metric is highly correlate to human judgments, and the proposed models are able to generate sentences with similar content preservation score but higher style transfer strength comparing to autoencoder.
Researcher Affiliation Academia Zhenxin Fu,1 Xiaoye Tan,1 Nanyun Peng,2 Dongyan Zhao,1,3 Rui Yan1,3 1Institute of Computer Science and Technology, Peking University, Beijing, China 2Information Science Institute, University of Southern California, California, USA 3Beijing Institute of Big Data Research, Beijing, China {fuzhenxin, txye, zhaodongyan, ruiyan}@pku.edu.cn, npeng@isi.edu
Pseudocode No The paper describes its models and algorithms in prose and mathematical equations but does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper mentions that a dataset is
Open Datasets Yes We compose a dataset1 of paper-news titles to facilitate the research in language style transfer. We composed the first dataset ourselves and used the data released by He and Mc Auley (2016) as the second dataset. 1Available at https://github.com/fuzhenxin/textstyletransferdata
Dataset Splits Yes For both datasets, we divided them into three parts: training, validation, and test data. The size of the validation and test data is 2,000 sentences, and the rest are used as training data.
Hardware Specification No The paper does not specify the hardware used for running the experiments. It does not mention any specific GPU models, CPU models, or other computing resources.
Software Dependencies No The paper mentions using
Experiment Setup Yes We use Adadelta (Zeiler 2012) with the initial learning rate 0.0001 and batch size 128 to learn the parameters for all models. The best parameters are decided based on the perplexity on the validation data with a maximum of 50 training epochs for paper-news task and 10 training epochs for positive-negative task. For paper-news title transfer, we explored word embedding size of 64, encoder hidden vector size among {32,64,128}, and style embedding size among {32,64,128}. For positive-negative review transfer, we explored word embedding size of 64 for multi-decoder and {64,128} for style-embedding model, encoder hidden vector size among {16,32,64}, and style embedding size among {16,32,64}.