Analogy is a powerful form of ideation and therefore an automated or semiautomated analogical method is a potentially useful way to develop, or at least inspire, new and possibly patentable ideas.The last few years have shown significant developments in the training and use of latent spaces for text generation using Variational Autoencoders (VAE), though many problems remain including preventing ‘collapse’ of the latent space during its training and successfully disentangling the latent variables, including the syntax from the semantics.A hierarchical sentence and document variational denoising autoencoder architecture is presented, in which the encoded sentence vectors are first generated and then an encoding and decoding is performed of the sequence (in the document) of these sentence vectors. The latent vectors for both sentences and documents are structured into ‘syntactic’ and ‘semantic’ subsections based on their use in auxiliary training tasks. A large dataset of patent titles and abstracts, along with their IPC6 codes, is used to train the VAE networks.The resulting document latent space is used to perform analogy transforms to seek to generate/inspire useful and potentially novel patent concepts.