About this project and app!

Contributing for more diversified and informative online political communication in Europe!

The PolarscopEU is an app that measures and maps the polarization of online social networks in EU countries. Addressing the major impact that online political communication is having on democratic institutions, as well as individuals’ unawareness of the bias of their online environments, this tool allows its users to better understand how they are positioned, in terms of the online content that they create and consume, in two scales of content diversity and bias. Additionally, PolarScopEU also offers its users different options to assess automatically the politicisation of relevant political issues made by different media and political actors. This tool not only facilitates comparisons of issue politicisation across countries, time periods, and actors. The data will will be used to create an innovative polarization index of online political communication in European Union (EU) countries.

How it works

At the moment, this application analyzes the political content of a Bluesky user and their network to provide an interactive “bias profile.” After entering a handle, the app collects a sample of posts from that user and the accounts they follow. It then applies automated text analysis to detect policy topics, measure sentiment, and identify references to the European Union. Topic classification is performed using a fine-tuned transformer model based on BERT, which assigns each post to a predefined policy category. Sentiment is measured using a transformer-based sentiment analysis model (XLM-RoBERTa), classifying posts as positive, neutral, or negative. EU mentions are detected using a structured dictionary-based approach that searches for relevant keywords and abbreviations in multiple languages. The results are aggregated to generate a visual “bias plot,” showing the relationship between tone and topic diversity for the user and their network. The app also computes similarity scores using cosine similarity to compare how closely other accounts align with the original user in terms of policy focus and tone. Together, these tools provide an accessible way to explore patterns in political communication and network alignment.

Technical Details

The topic classification component relies on a fine-tuned transformer model developed by the poltextLAB Artificial Intelligence Laboratory and distributed via Hugging Face. The model is trained to assign short political texts to a predefined set of policy categories. For each post, the classifier produces a ranked set of candidate labels with associated probability scores. The application applies a structured post-processing procedure to select the most appropriate category, incorporating calibrated confidence thresholds to ensure stability and consistency in topic assignment across heterogeneous text inputs. Sentiment analysis is conducted using a multilingual transformer model based on the XLM-RoBERTa architecture. The model outputs a probability distribution over three sentiment classes (positive, neutral, negative), from which the highest-probability label is selected as the dominant tone. The corresponding confidence score is retained for analytical transparency. Detection of references to the European Union is implemented through a rule-based dictionary approach that combines curated keyword lists and abbreviation patterns in multiple languages, allowing for systematic identification of EU-related content. Network alignment is operationalized using cosine similarity over normalized policy distributions. For each account, a vector representation is constructed in which each dimension corresponds to the proportion of posts assigned to a specific policy category. These vectors are normalized at the account level to control for differences in activity volume. Cosine similarity is then computed between the original user and each network account, providing a scale-invariant measure of thematic proximity. An extended specification incorporates tone by constructing joint topic–sentiment vectors (e.g., Topic A–Positive, Topic A–Negative), thereby capturing not only thematic alignment but also similarity in evaluative orientation.