Recent study in auto-encoder framework has led to advances in natural language processing. One challenging but critical task in modeling text data is to learn the document level representation. We shed some light on the possiblity of combining

traditional lingural features with a probabilistic auto-encoder model (VAE), on a level of document representation, to improve the quality and robustness of a document classification system. For more details about the project, check our poster here!

One approach to build unsupervised representation learning framework is through generative modeling. The nature of text data, including discreteness or an recurrsive structure, makes it more difficult to be modeled, particular the models delievered from other communities such as vision. We hereby focus on collecting

and summarizing recent developments on such challenging task. The methods being studied include variational autoencoder, NVDM for document modelling, generating sentence with VAE, paragraph vector model, skip-thought model, sequence autoencoder, and multi-task sequence to sequence learning. For more details, check the paper here!