TaylorGAN: Neighbor-Augmented Policy Update Towards Sample-Efficient Natural Language Generation

Authors: Chun-Hsing Lin, Siang-Ruei Wu, Hung-yi Lee, Yun-Nung Chen

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments demonstrate that Taylor GAN achieves state-of-the-art performance without maximum likelihood pre-training. The dataset used in the experiments is EMNLP 2017 News... We evaluate our model s ability to generate realistic and diverse texts with n-gram based metric, Fréchet Embedding Distance and language model based metric.
Researcher Affiliation Academia Chun-Hsing Lin Siang-Ruei Wu Hung-Yi Lee Yun-Nung Chen National Taiwan University, Taipei, Taiwan
Pseudocode No The paper describes the proposed method verbally and mathematically but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The source code and data are available at https://github.com/Miu Lab/Taylor GAN/
Open Datasets Yes The dataset used in the experiments is EMNLP 2017 News, where sentences have a maximum length of 50 tokens and a vocabulary of 5.3k words after performing lowercase. The data is at https://github.com/pclucas14/Gans Falling Short/tree/master/real_data_experiments/data/news
Dataset Splits Yes Training and validation data consist of 269k and 10k sentences respectively.
Hardware Specification Yes Each training takes approximately 1 day on Nvidia Tesla V100 GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries).
Experiment Setup Yes All models (detailed in Appendix E) are trained for at most 200 epochs, namely 800k training steps with the batch size N = 64.