This paper addresses the challenge of generating plausible text sentences with controlled attributes, such as sentiment and tense, by learning disentangled latent representations. The authors propose a new neural generative model that combines variational autoencoders (VAEs) and holistic attribute discriminators to effectively impose semantic structures. The model can be seen as enhancing VAEs with the wake-sleep algorithm, enabling the use of fake samples as additional training data. The model learns interpretable representations from word annotations and produces sentences with desired attributes. Quantitative experiments using trained classifiers validate the accuracy of short sentence and attribute generation. The paper discusses the challenges of discrete text samples and the need for disentangled latent representations, and presents the model's architecture, including the generator and discriminators. The model's effectiveness is demonstrated through experiments on sentiment and tense generation, showing improved accuracy and interpretability compared to existing methods.This paper addresses the challenge of generating plausible text sentences with controlled attributes, such as sentiment and tense, by learning disentangled latent representations. The authors propose a new neural generative model that combines variational autoencoders (VAEs) and holistic attribute discriminators to effectively impose semantic structures. The model can be seen as enhancing VAEs with the wake-sleep algorithm, enabling the use of fake samples as additional training data. The model learns interpretable representations from word annotations and produces sentences with desired attributes. Quantitative experiments using trained classifiers validate the accuracy of short sentence and attribute generation. The paper discusses the challenges of discrete text samples and the need for disentangled latent representations, and presents the model's architecture, including the generator and discriminators. The model's effectiveness is demonstrated through experiments on sentiment and tense generation, showing improved accuracy and interpretability compared to existing methods.