Identifying Party Families with Time Series Multilingual Word Embedding

• In Event: Embeddings in Multiple Languages

Abstract

While party family is an important and widely used concept in the study of European Politics, much of the identification of party families relies on expert surveys. In this paper, we develop a new way of automatically identifying party families and their change over time using party manifestos and parliamentary speeches. Built on recent developments of machine learning and natural language processing, especially time series word embedding and our previous work on multilingual word embedding, we create vector representations of each party in each time period t for parties in 7 European countries. These party vectors exist in a shared vector space based on the manifestos and speeches given by the party in the time period. Since all parties are positioned in the same vector space, we apply clustering analysis to all parties to identify party clusters (families). Our approach allows us to analyze party families across countries (languages) and over time without having to translate all documents into a single language. Among other things, we examine whether some party families are more ideologically homogenous than others, whether membership in particular party families has changed over time, and whether party families identified by cluster analysis match those identified by more traditional expert surveys.

Authors

• Yaoyao Dai, University of North Carolina, Charlotte

• Matt Golder, Pennsylvania State University

• Benjamin J Radford, University of North Carolina at Charlotte