Towards controllable neural generation of arguments
Xinyu Hua · 2021
Argumentation is an essential cognitive skill that we utilize to achieve various communicative goals. It has great impact on all aspects of our lives, ranging from policymaking to conflict resolutions. The automatic construction of persuasive arguments thus presents great opportunities, as it can significantly reduce the workload required to search evidence and existing opinions of the issue at hand. Yet the research on argument generation is relatively under-explored. Existing methods either rely on manually crafted rules tailored for specific domains and language styles, or built on pre-indexed argument inventory for retrieval. It remains unclear whether the retrieved arguments can be organically integrated into the underlying text generation stack in a controlled manner. This dissertation presents a line of research aimed at more controllable neural text generation systems for arguments. We focus on two key aspects: (1) the proper inclusion and organization of diverse external knowledge that are pertinent to the given topic and stance, and (2) the generation modeling design that enables effective and interpretable text planning and conditional realization. We experiment with a newly collected Reddit ChangeMyView corpus and highlight the great impact of retrieval results as well as the quality of text planning. We then showcase how the controlled generation framework can be generalized to other domains with distinct language styles. Finally, we describe two novel formulations designed for pre-trained Transformers to achieve improved fluency and controllability for argument generation and beyond. --Author's abstract