Stylistic Control for Neural Natural Language Generation

With the rise of conversational assistants, it has become more critical for dialog systems to keep users engaged by responding in a natural, interesting, and often personalized way, even in a task-oriented setting. Recent work has thus focused on stylistic control for natural language generation (NLG) systems in order to jointly control response semantics and style. In this talk, I will describe our work on automatic data curation and modeling approaches to facilitate style control for both personality-specific attributes of style (based on Big-Five personality traits), and other style attributes that are helpful for personalization, e.g., response length, descriptiveness, point-of-view, and sentiment. I will present work that incorporates these attributes into the training and generation pipelines for different NLG architectures, and will show how our data curation and modeling approaches are generalizable to new domains and style choices. Finally, I will describe how we use a combination of automatic and human evaluation methods to measure how well models successfully hit multiple style targets without sacrificing semantics.