Title: Document-level Modeling of Discourse
Abstract: Discourse models capture a number of complex language processes, such as describing our experiences or how and what we talk about in a conversation.
Exponential family models provide a reliable foundation and building block for
usable and interpretable discourse models. More generally, they are an
indispensable technique within human language processing, human language
technologies, and other fields. In this talk, we will explore a number of
exponential family discourse models, along with effective inference strategies,
that can achieve consistent 10% improvement in capturing latent meaning
expectations and nearly an order of magnitude improvement in intrinsic and
holistic probabilistic evaluations.