SentiGAN: Generating Sentimental Texts via Mixture Adversarial Networks

Authors: Ke Wang, Xiaojun Wan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on four datasets demonstrate that our model consistently outperforms several state-of-the-art text generation methods in the sentiment accuracy and quality of generated texts.
Researcher Affiliation Academia Ke Wang, Xiaojun Wan Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {wangke17, wanxiaojun}@pku.edu.cn
Pseudocode Yes Algorithm 1 The adversarial training process in Senti GAN
Open Source Code No The synthetic data and the oracle model (LSTM model) are publicly available at https://github.com/Lantao Yu/Seq GAN
Open Datasets Yes Movie Reviews (MR). We use Stanford Sentiment Treebank [Socher et al., 2013] which has two sentiment classes (negative and positive). The original dataset has a total of 9613 sentences. We select sentences containing at most 15 words, and the resulting dataset contains 2133 positive sentences and 2370 negative sentences. Beer Reviews (BR). We use the data scraped from Beer Advocate [Mcauley and Leskovec, 2013]. Customer Reviews (CR). We use customer reviews of various products [Hu and Liu, 2004].
Dataset Splits Yes We select sentences containing at most 15 words, and the resulting dataset contains 2133 positive sentences and 2370 negative sentences. (MR) ... and the resulting dataset contains 1437767 positive sentences and 11202 negative sentences. (BR) ... and the resulting dataset contains 1024 positive sentences and 501 negative sentences. (CR)
Hardware Specification Yes We implement our model based on Tensorflow and use a TITAN X graphic card for learning.
Software Dependencies No We implement our model based on Tensorflow
Experiment Setup Yes The generators are set as single-layer LSTM-RNNs with input/hidden dimension size of 300 and max sample length of 15 words. The CNN in our discriminator is the same as [Zhang and Lecun, 2015]. The N in Monte Carlo search is set as 15. In the per-training step, we pre-train generators for 120 steps, pre-train the discriminator for 50 steps. In adversarial training, the g-steps is 5 and d-steps is 1. The optimization algorithm is RMSProp.