The final speaker in this session at the DigiWorld network ECREA preconference on election campaigning in Brno is Patrick Parschan, whose focus is on the use of ideal point estimation for the analysis of party positions in campaign tweets. There are various text scaling algorithms for the analysis of political positions in tweets on a left-right scale, in comparison with party manifestos, but unfortunately these often do not agree with each other.
Patrick approached this with data for US and German elections across several years, using party manifestos and tweet corpora. Existing approaches have tended to draw on word frequency algorithms; topic modelling; word embeddings; and LLM-based categorisation: Patrick systematically compared their performance against each other, testing both whether they would work for short tweet texts at all, or (conversely) could cope with large manifesto corpora.
Using these approaches to order parties from left to right, for both manifestos and tweets, produces some fairly inconsistent results; while in Germany, for instance, the far-right AfD is usually positioned on the far right, but even this is not always the case. This is also dependent on whether they run on entire corporate or process each manifesto and tweet separately. Also, manifesto outputs cannot necessarily serve as the ground truth here: tweeting patterns may vary considerably from official party policies, of course.
