By PAUL VON BÜNAU
The astonishing rise of the “Alternative für Deutschland” (AfD, translated: alternative for Germany) has shaken up the political establishment in Germany. Despite a double-digit finish in three state elections in March, the political elite in Germany still struggles on how to deal with the new political landscape. In late April, the AfD is scheduled to vote on its first party program, outlining its basic policies on issues such as migration and internal security. While there is plenty of journalistic analysis of the draft program, we have taken a different route and applied a machine learning approach to the AfD program. What we found, has the potential to trigger serious political debate.
How is it possible to apply machine learning to political texts?
We build up our analysis on a data set of the “Manifesto Project” , where sociologists have manually tagged sentences in political documents from various parties with information such as political orientation and policy area. This is a highly valuable historic dataset, which we utilized to apply machine learning techniques. Training a “classifier” on the existing data, we are able to predict the political orientation on a left-to-right-spectrum of any given sentence and paragraph. Additionally, we have leveraged various approaches to determine the similarity of paragraphs and the uniqueness of statements.
Finding #1: The AfD party program shares some similar paragraphs with other parties
Within the AfD party program draft, there are quite a few paragraphs, which reveal a stark similarity with other party programs. Some with the right-wing extremist party NPD, others with moderate parties such as CDU and SPD. This does not come as a surprise, if one looks at the composure of AfD membership. Most party members had affiliations with other parties before they joined the AfD. The AfD, one could claim, is indeed a melting pot of different ideas from across the spectrum, which makes the endeavor to unite the party under one common umbrella a tough undertaking.
Finding #2: The AfD has a similar Left-Right-Profile as German mainstream parties
When moving beyond the fine-grained paragraph-by-paragraph comparison, looking at the party program profile from a macro perspective, we observed striking similarities between the AfD and SPD and CDU, which are commonly referred to as people’s parties (“Volksparteien”) in political discourse. This analysis does have interesting implication: Either, the Left-Right spectrum is not capable to display the uniqueness of the AfD (which would then render it obsolete), or new political parties are so unique in their syntax and wording, that a machine learning approach, trained on documents of other parties, is not able to “comprehend” the new political language.
Finding #3: AfD has quite a few “unique” statements
While the left-right-profile didn’t reveal any AfD-uniqueness, we approached the issue from yet another perspective, trying to single out “unique” statements. We found that quite a few issues were uniquely represented in the AfD program, while the manifesto of the other German parties – of the SPD in particular – did not merely bear as many unique statements. The AfD thus appears to clearly contrast itself as opposed to SPD, CDU and the Greens, which are slowly converging in their political beliefs. This observation is actually mirrored by day-to-day politics: SPD and CDU form the reigning coalition in Germany, while CDU and the Green Party are in negotiations about cooperation in Baden-Württemberg, one of Germany’s largest states.
The full analysis (with graphs, but in German) can be found here. Additionally, we have incorporated a tool, which allows for easy navigation of the different party programs. Thus, the interested reader can start to explore similarities and differences on his or her own, critically reflecting on our analysis.
This analysis was conducted during an “after-work hackathon” (involving alcoholic beverages, at times) and we do not claim to have conducted a holistic and comprehensive analysis. On the contrary, we would like to spark the debate within the machine learning community about how to identify the political direction of any given paragraph in a more reliable manner. Please feel free to reach out to us at firstname.lastname@example.org.
 Volkens, Andrea / Lehmann, Pola / Matthieß, Theres / Merz, Nicolas / Regel, Sven / Werner, Annika (2015): The Manifesto Data Collection. Manifesto Project (MRG/CMP/MARPOR). Version 2015a. Berlin: Wissenschaftszentrum Berlin für Sozialforschung (WZB)