MARDY: Modeling Argumentation Dynamics in Political Discourse (Phase 2)

点击翻译按钮获取中文摘要

This interdisciplinary collaboration project involving Computational Linguistics, Machine Learning and Political Science has the aim of developing computational models and methods for analyzing argumentation in political discourse - specifically capturing the dynamics of discursive exchanges on controversial issues over time. The goal is to support analysis of the possible impact of arguments advanced by different political actors. Here, factors besides the substance of the claim and its justification need to be considered: the development and structure of discourse coalitions, overlapping and competing use of justifications and frames, the status of and relations between actors, etc. Such factors are to be integrated into effective, scalable computational models which support both professional analysts (such as political scientists) and informed laypersons researching current and past debates. Our modeling approach combines state-of-the-art analysis methods from language technology with powerful machine learning techniques (in particular joint inference and deep learning) and analytical insights from political science to produce globally coherent discourse networks relating actors, claims, and justifications. We do so by solving the necessary computational component tasks of extracting the above factors from texts and combining them in more complex models. As part of model development, the project will create high-quality annotations (coding) for a corpus of past debates in German political discourse as covered in daily newspapers, capturing issue-related content of claims, argumentative structure and relevant discourse references. The annotated newspaper corpus will be made freely available for academic research. To complement empirical access to debates mediated through newspapers, we will use additional sources (available corpora of parliamentary proceedings) to access actors' original contributions and for semi-structured systematization of argumentative positions.By applying dynamic network models to datasets of real debates, retrospective predictive modeling experiments can be used to test hypotheses about empirically relevant factors driving political discourse and to learn how model parameters relate to specific elements of theoretical interpretation and/or to intuitions that expert analysts have developed through experience. Using interactive visualization techniques and diagnostic tools that we will develop alongside the modeling, (families of) predictive models applied on new data can contribute to the identification of unexpected turns in an ongoing debate, and to other schemes of explorative or systematic analysis of argumentation dynamics in political discourse. The close connection among the participating groups ensures that the computational models transfer into useful tools for political scientists, and we aim at establishing a "best-practice" methodology for this exchange over the course of the project.