Whose Line Is It? – Quote Attribution through Recurrent Neural Networks

This paper presents a recurrent neural network framework for the problem of attributing spoken lines to characters in a screenplay or novel. We study these quotes as a sequence in the absence of additional context, e.g. descriptions of scenes or actions, from the text surrounding them. Instead, attributions may only be made on the basis of learned expectations for how each character speaks, as well as an understanding of how they converse with each other. We use gated-feedback recurrent neural networks, trained in a supervised fashion, for modeling both of these aspects. We evaluate the prediction model on episodes of the television show Futurama and demonstrate improvement over simpler neural network constructions.