Generalizable Negativity: Classifying Negative Campaigning across German Elections

Negative campaigning (NC) has become a prevalent campaign strategy in recent years, especially in interactive media such as Facebook, Twitter/X or Instagram where politicians can communicate without any mediation by gatekeepers like journalists. Even though the identification of NC is a crucial precondition for understanding modern election campaigns, the automated identification of NC is still in its infancy, even with state-of-the-art text analysis methods. Especially when researchers want to identify NC across elections and different levels of a polity, they need to develop and validate specific classifiers. As actor constellations and topics vary considerably across different elections, applying automated measures to lower-level elections requires a thorough validation of how well trained models generalize across contexts. Using several federal and state level elections as a test case, this paper investigates the generalizability of transformer models for text classification in the complex German multilevel and multiparty system. Based on over 40,000 social media posts of candidates in eight state and federal elections between 2013 and 2021 that were annotated by human coders, the classifier can identify negative campaigning across space and time. A leave-one-out classification shows that the model can accurately predict data even for unknown elections with moderately sized training data. We demonstrate how a classifier for a demanding theoretical concept can be trained and validated in multidimensional contexts and provide orientation for similar projects.