In this paper, we empirically investigate how communications by the European Central Bank (ECB) alter the market’sbeliefs about the market deflation outlook. We determine events when the Executive Board of the ECB express theirviews about deflation in the Eurozone within the observation period from October 2009 to October 2016. In doingso, we apply a textual analysis of the contents of formal ECB communications as well as related news and Tweets.We use options on the inflation index to extract market deflation expectations. In this deflationary environment, werun our quasi-natural experiment to understand the ability of ECB to shape market expectations through their voices.Our main findings show that within a Granger causality relation between traditional media and social media, formalECB outlooks of deflation have diverse effects on market deflation expectations. The long-term and short-term marketdeflation expectations are mostly derived by high influential members of social media and online type of news mediawhich carry the ECB messages, respectively. Taken together our results suggest that ECB communications underminethe anchoring of inflation expectations in the Euro area.

Relevance and Research Question:

While the measurement of political positions from text has a long tradition in political science, the rapid developments in machine learning and natural language processing offer new opportunities to extract more information. Especially, it is now much easier to move beyond a bag-of-words approach at the large scale, extracting more information from political text. The article proposes an application and evaluation of cross-lingual topical scaling on political manifestos, asking whether it allows a valid measurement of populist rhetoric from political text.

Methods and Data:

Populist rhetoric is estimated relying on several hundred election manifestos (source: polidoc.net) from eight European countries across five languages (English, French, Spanish, Italian and German). We adopt word embeddings, which are vector representations of the meaning of words in their context. Using Python, these are utilized to assess the similarity of sentences from election manifestos to a set of core populist keywords, aggregating at the manifesto level using harmonic function label propagation. The results are assessed for face, convergent and construct validity. The approach features cross-lingual word embeddings and additional model tweaks to accommodate differences in populist discourses across contexts.

Results:

Analyses on parts of the sample (German and English manifestos) reveal that the approach works reasonably well within single countries. The resulting measure places well-known populist parties at the correct end of the spectrum, correlates with existing data and detects diverse populist statements in manifestos based on a simple query on word embeddings. At the same time, the cross-lingual scaling exercise appears challenging, with country clusters emerging from the analysis.

Added Value:

In addition to moving beyond a bag-of-words approach, the paper uses word embeddings for topical scaling in a way that does not require the qualitative coding of a training set. If the approach proofs valid, it can safe even more human effort in the estimation of levels of populist – or other – rhetoric in political text. The method delivers continuous measures of sparse, here populist rhetoric for all kinds of parties which can be utilized in the comparative analysis of politics.