A Textual Time Capsule: A Text-Based Approach to Tracing Swiss Elite Polarization Across 130 Years
Chair of Systems Design, ETH Zurich
14 Sep 2023, 14:45–15:10
Increased polarization has long been lamented as a sign of democratic instability and erosion. It is a phenomenon constantly postulated to have increased in the past few decades. However, this has yet to be studied over a prolonged period, and Switzerland is no exception. Thus, we ask whether this claim still holds when studying polarization over the past 130 years in the Swiss parliament.
To achieve this, we employ text embeddings to compute multi-dimensional representations of Swiss parliament members based on the legislative bills introduced, from 1891 to today. We build upon the SwissBERT model, a Switzerland-specific model based on the XMOD architecture. Furthermore, to enhance the model’s understanding of language structure and context, we fine-tune SwissBERT using masked language modeling (MLM) on parliamentary texts, encompassing speeches and bill texts. Additionally, we train the embeddings to predict party class. Finally, we employ a multi-dimensional variance-based polarization measure to study polarization trends in the Swiss Parliament over 120 years. Our findings indicate that, while polarization has indeed increased, this trend began in the 1970s and reached its peak in the 1990s.