Latent variable models can help us conceptualize an unobserved data-generating process. In the ideal case, posterior estimates of these variables can reveal meaningful structure in the data. However, during inference, the latent variables may not obey the roles for which they have been designed. In this talk, I will outline two examples of nonstandard inference procedures used to uncover structure in data. First, I will discuss an unsupervised technique to discover gendered language in a large corpus of text, examining the differences in the words used to describe men and women. Then, I will describe a multi-view VAE used to combine disparately coded sentiment lexica into a common representation. I will end with an overview of ongoing work that continues this effort to guide the latent space.
Alexander Hoyle is a PhD student in the Computational Linguistics and Information Processing (CLIP) Lab at the University of Maryland, advised by Philip Resnik. His (still nascent) research interests involve using interpretable unsupervised methods to study social and cultural phenomena as manifested in language.