Ten years of online conversations on DerStandard became a living lab in public discourse—the kind of data that only reveals itself when you watch long-term signals rather than the latest headline. A new dataset makes that arc legible, exposing not just what readers argued about, but how the arguments unfolded, who kept the conversation alive, and where the edges of opinion tend to gather after a game-changing event.
The study behind this dataset is led by Emma Fraxanet Morales and Vicenç Gómez, with Andreas Kaltenbrunner and Max Pellert, and conducted in collaboration with the Austrian Research Institute for Artificial Intelligence. It catalogs more than 75 million comments, over 400 million votes on those comments, and metadata on nearly 580 thousand articles spanning 2013 to 2022. It’s a rare blend of scale, structure, and German-language texture that invites researchers to map the social geography of online debate across a full decade, not just a single moment. DerStandard, founded in 1988 and online since 1995, was an early adopter of online discussion spaces, evolving from chat rooms to threaded comment forums with semi-automated moderation in collaboration with institutions like OFAI—the Austrian Research Institute for Artificial Intelligence.
What the dataset actually captures
At its core, the dataset is a ledger of threads, with each comment linked to an article, a timestamp, and the people who wrote or voted on it. The social signals—upvotes and downvotes—create a signed map of agreement and disagreement, letting researchers glimpse not only what people say, but who echoes what, when arguments intensify, and where a discussion fractures into subthreads. The site’s threaded structure provides a natural laboratory for studying how conversations evolve as they grow deeper or spread wider. It’s not a snapshot; it’s a living, breathing record of how communities argue over time.
Scale matters here. More than 75 million comments, 400 million votes, and nearly 580 thousand articles over a decade form a cross-section of a real community under pressure—from elections to pandemics to economic shifts. The platform’s editorial tags add a second axis, showing how topics migrate from broad categories into focused debates about refugees, climate policy, or the economy. To balance privacy with usefulness, the team also releases 896-dimensional embeddings for each comment, which capture meaning and context without by exposing the exact words themselves. This is not just data hoarding; it’s a carefully engineered compromise that keeps the human story legible while protecting individuals.
All of this is packaged in a form that respects privacy but remains usable. Anonymized user IDs, salted hashes, and the embeddings do the heavy lifting. The embeddings—KaLM-based—are designed to preserve semantic relations enough for tasks like topic classification and clustering while ensuring that the exact sentences aren’t retrievable. The setup is CPU-friendly, with data partitioned into monthly and yearly resolutions to make it accessible to researchers with limited computing power. It’s a deliberate attempt to widen access without sacrificing ethics, a blueprint for how to scale social science in an era of data abundance.
What the data reveals about polarized conversations
One of the most striking features is the explicit sign information in votes. With upvotes and downvotes attached to each comment, researchers can build signed networks that reveal not only who tends to agree with whom, but where disagreements cluster and how a debate travels across time. It’s a map of social friction—blue zones of alignment, red zones of opposition, and the gray swirls where opinions bend and reform. These signed interactions are the bones of polarization in a real community.
For a long-running community like DerStandard, prior work had glimpsed polarization through patterns of language and topic choice. The new dataset preserves a way to link those patterns to persistent factions and to study their behavior across a decade. It’s not merely a binary split; it’s a mosaic of issue-based alignments. Some readers cluster around immigration and security concerns; others circle around economic policy or Europe-wide politics. The study also makes it possible to connect individual users to the factions previously identified and to study how attachment to those positions changes—or stubbornly sticks—as events unfold. The result is a rich, dynamic portrait of how online publics are shaped by, and in turn shape, the real world.
Beyond who agrees with whom, the data reveal how topics themselves connect. The embeddings show that comments about focused topics such as football form tight semantic clusters, while broader policy topics yield looser but still recognizable coherence. Interestingly, semi-related topics such as refugees and policy issues display higher cross-topic similarity than unrelated domains like football and distant international topics. In short, the map of ideas on DerStandard reflects a web where some ideas travel together more readily than others, and where the social fabric of debate shapes the path of semantics as much as words do.
Thread structure matters too. The three views the researchers bring together—a thread tree, a reply network, and a signed vote network—reveal different kinds of influence. Deep threads tend to harbor more sustained exchanges, sometimes slowing cross-cutting dialogue, while fast, shallow exchanges can spur rapid, cross-cutting disagreements that cross topic boundaries. The data allow scientists to quantify these dynamics and link them to real-world events, offering a more nuanced picture of how online forums adapt to external shocks over years rather than days. The long arc matters: it’s where collective minds shift, or stubbornly resist, in the face of headlines and policy twists.
Why privacy-preserving data can unlock public insight
Privacy is not an obstacle here; it is a design choice. The DerStandard dataset deliberately avoids releasing raw text, instead offering embeddings and a fully anonymized user map. This preserves the roar of discussion—the timing, frequency, and social reach—without exposing private words that could be misused or traced back to a person. It’s a pragmatic compromise that keeps research viable and readers safer, while preserving the analytical backbone that social science needs. The result is a resource that invites researchers to ask big questions without turning the forum into an identifiable archive of private speech.
Why does that matter in practice? Because data-sharing is often the bottleneck of modern science. Major platforms pull the plug on public access or clamp down on who can see what. The DerStandard release demonstrates that you can have a data resource that scales to tens of millions of comments while staying ethically constrained and scientifically productive. More importantly, it proves a path forward for non-English communities, expanding the field beyond the English-dominated datasets that have long steered the conversation. In a world of linguistic diversity, this kind of resource is a reminder that insight does not have to come at the expense of people’s privacy.
In the future, datasets like this could guide the design of healthier digital forums. If moderators know which topics are likely to fracture conversations, and which social ties tend to bridge gaps, they can tailor interventions that reduce polarization without silencing voices. If researchers can trace how moderation affects conversation lifecycles, platforms can balance openness with civility. And for policymakers, such maps offer a more grounded sense of how online public opinion actually forms and shifts over time, beyond press-release snapshots. The decennial scope is the key here: it lets us see whether interventions have lasting effects or merely patch the next flare-up.
As the online public sphere continues to evolve, decade-spanning resources like this DerStandard dataset could become a compass for designing forums where passionate voices meet listening ears—and where data-informed design nudges discourse toward curiosity rather than combat.
