Abstract
Argumentation plays an important role in a wide range of human activities, ranging from casual discussions to legal negotiation, where multiple parties form reasons and draw conclusions throughout this process. Computational argumentation, as a newly emerging research field, aims to automatically identify and extract the argument components presented in natural language and also predict the relations among them. Understanding argumentative structures allows us to ascertain people’s positions towards a controversial topic and their reasons for justifying the opinions, providing valuable insights into various fields. Current works mainly focus on primary tasks in the emerging wider field of argument mining. In this thesis, we aim at understanding and leveraging argument mining more comprehensively. Specifically, to better understand argument mining, our research works explore more integrated and more practical argument mining tasks in various domains, including the debating system and the peer review and rebuttal process. We also propose new datasets and methods to facilitate the study of these new tasks. Converting the unstructured natural language into structured argument segments together with relations among them makes it possible to have a deeper understanding of natural language and a clearer picture of each party’s positions and why they are holding the positions. With the understanding of argument mining, we then investigate how to leverage argumentative structures for generation. Expressing opinions in natural language in a convincing way benefits the acceptability of the standpoints for the readers. Motivated by generating such convincing opinions, we propose a structure-controllable text generation task with an in-depth understanding of the text structure, which offers stronger generation flexibility and applicability for practical use cases.