Neural language models have recently shown impressive performance in generating realistic text that can be hard to distinguish from human writing. However, controllable generation and manipulation of text remain difficult. In this talk, we’ll explore latent variable approaches for language modeling to enable control, and discuss the challenges and theoretical insights into molding a meaningful latent space geometry for discrete text data. As a preliminary result, we’ll demonstrate the success of our model in sentence interpolation and style transfer through simple vector arithmetics.
Tianxiao Shen is a third-year PhD student at MIT, advised by Prof. Regina Barzilay and Prof. Tommi Jaakkola. Her research interests are in natural language processing and machine learning, with a focus on text generation. Before coming to MIT, she received her B.E. in Computer Science from Tsinghua University.