Using Multilingual LLMs to Detect Group Appeals

• In Event: Group Appeals

Abstract

A growing literature in political science studies parties’ group-based appeals, a prominent feature in parties' electoral strategies. To date, such literature has faced serious issues in the availability, quality and cross-national comparability of such data.
This paper shows how we can resolve these issues by detecting group appeals using multi-lingual LLMs (specifically a fine-tuned mDeBERTa-3 model) to detect group appeals, classify the groups into meaningful categories and detect the stance that (positive, negative or neutral) parties take towards the social group. These models perform impressively well in classifying group appeals. We analyze election manifestos of major parties in the UK, the Netherlands, Belgium, Switzerland, Germany, Austria, Spain and Ireland since the 1970s, showing LLMs to be valuable in extracting data on group appeals, tracing trends in appeals to different groups within the electorate over a long time period.

Authors

• Will Horne, Clemson University

• Alona Olga Dolinsky, Vrije Universiteit Amsterdam

• Lena Maria Huber, University of Mannheim