Traditional Natural Language Generation (NLG) methods incorporated a planning stage where the content of the text to be generated was planned before its surface realization. However, with the advent of sequence-2-sequence models, modern NLG methods have gravitated towards directly generating text from the input without an explicit planning stage. This compromises the narrative flow and quality of the text and makes the generation process difficult to steer.

 

In this talk, I discuss our approaches to text generation that incorporate content-plans. In the first part of the talk, I describe our approach to the problem of data-2-text generation. A key challenge in this problem is the structural gap between the input graph and the output (sequential) text. We bridge this structural gap by inserting an automatically generated content-plan between the input and the output. In the second part of the talk, I describe our approach to incorporating a user-provided plan for the task of story generation. We use Reinforcement Learning to encourage the generation model to adhere to the plan and show that our approach results in text that fits the user's plan while maintaining its coherence.