Specific response generation and elaborative discourse structure

Abstract

When composing text, a writer has to carefully choose the discourse structure for coherence and effective communication. For each sentence, this involves deciding how much detail to provide (specificity), and what details to elaborate on. The goal of the thesis is to move towards better discourse structure in natural language generation by understanding how to control specificity in generation and what details to provide. The first part of the thesis explores techniques to control the specificity in dialogue response generation and tackles the plausibility issues that arise as more specific sentences are generated. The second part explores what to be specific about. We create question datasets which we apply to modeling the elaborative discourse structure between sentences, and question answering of openended questions.