Semantic-Map-Based Assistant for Creative Text Generation

Abstract A weak semantic map of English words is used here as an intuitive interface to a pattern generator, which produces paragraphs of text in the style of a specific writer. For this purpose, a recurrent neural net based on the Long Short-Term Memory (LSTM) model is used. The objective is to generate a poem in the style of Byron. The result is a guided generation of sequences of words, based on the correspondence between latent neural network representations and the semantic map. Future applications of this technique in the form of cognitive assistants are discussed, including an automated poetry-writing assistant.