A Decade of News Forum Interactions: Threaded Conversations, Signed Votes, and Topical Tags

This paper introduces a large-scale, privacy-preserving dataset of ten years of user interactions on the Austrian newspaper DerStandard, comprising over 75 million comments and 400 million votes with anonymized identifiers and pre-computed vector embeddings to facilitate research on online discourse dynamics in the German language.

The Gist

Imagine a massive, digital town square that has been buzzing with conversation for ten straight years. In this town square, people don't just shout their opinions; they write them down, argue in organized lines, and vote on whether they agree or disagree with what others say.

This paper is essentially the blueprint and the map of that town square, which belongs to DerStandard, a major Austrian newspaper. The authors have spent a decade collecting every single interaction that happened there and turned it into a giant, safe-to-use dataset for scientists to study.

Here is the breakdown of what they did, using some everyday analogies:

1. The "Time Capsule" of Conversation

Think of this dataset as a 10-year time capsule (from 2013 to 2022) containing:

• 75 million comments: That's like every person in a large country writing a diary entry every single day for a decade.
• 400 million votes: Imagine a giant stadium where every single person in the audience raises a green hand for "I agree" or a red hand for "I disagree." This is rare because most social media sites (like Twitter/X) don't let you see exactly who voted for what, or they don't keep a perfect record of it.
• The Context: Every comment is tied to a specific news article (like a football match, a new law, or a pandemic update), so researchers know exactly what people were talking about.

2. The "Magic Translator" (Privacy Protection)

Here is the tricky part: You can't just hand over a diary full of people's real names and private thoughts to the public; that would be a privacy nightmare.

So, the authors used a digital magic trick:

• The Anonymizer: They took every user's name and ID and ran it through a "salted hash" machine. Think of this like turning a person's face into a unique, unrecognizable fingerprint. You can still tell that "Fingerprint A" is the same person across different conversations, but you have no idea who that person actually is in real life.
• The Secret Sauce (Embeddings): They didn't share the actual text of the comments (the words people wrote). Instead, they used a super-smart AI translator to turn every comment into a mathematical recipe (a vector of numbers).
  • Analogy: Imagine you have a book. Instead of giving someone the book, you give them a list of numbers that describes the book's "flavor," "mood," and "topics." If two books are about "football," their number-lists will look very similar. If one is about "football" and the other is about "baking," the lists will look totally different. This lets scientists study the meaning of the conversations without ever reading the actual words or seeing the users' names.

3. Why This Town Square is Special

Most online forums are like chaotic mosh pits where people drift in and out, or they disappear when a new app becomes popular (like when Twitter changed its name to X).

DerStandard is different:

• It's Stable: It's been running for decades, like a well-built library, not a pop-up tent.
• It's Organized: Unlike the chaotic comment sections on social media, this place has "threads." It's like a family tree of conversation where you can see exactly who replied to whom.
• It's Honest: The voting system (Green/Red hands) gives researchers a clear signal of agreement or disagreement, which is hard to find elsewhere.

4. What Can Scientists Do With This?

Because this data is so clean and structured, researchers can use it to answer big questions:

• How do arguments grow? They can watch a conversation start at the top and branch out like a tree, seeing how deep the arguments get.
• Who are the tribes? By looking at who votes for whom, they can map out the "factions" or political tribes within the community.
• How does the world change? They can track how the mood of the crowd shifted during big events, like the Coronavirus pandemic or elections, by looking at the "flavor" of the comments over time.

The Bottom Line

This paper is a gift to the scientific community. It's a huge, safe, and organized library of human conversation in the German language. It allows researchers to study how people argue, agree, and form communities online, without ever violating anyone's privacy. It's like having a microscope that can look at the DNA of a decade of public debate, without ever needing to see the faces of the people involved.

Technical Summary

Here is a detailed technical summary of the paper "A Decade of News Forum Interactions: Threaded Conversations, Signed Votes, and Topical Tags".

1. Problem Statement

The research addresses the growing scarcity of accessible, large-scale, longitudinal datasets for studying online human behavior and discourse.

• Data Access Restrictions: Major commercial social media platforms (e.g., Twitter/X) are increasingly restricting API access or shutting down public data streams.
• Instability of Social Platforms: Mainstream social platforms often suffer from rapid user migration or shifts in topical focus, making long-term longitudinal studies difficult.
• Language Bias: Most large-scale online datasets are English-centric, leaving "mid-resourced" languages (like German, spoken by ~100 million people) underrepresented in computational social science.
• Privacy vs. Utility: There is a tension between protecting user privacy (preventing re-identification) and providing raw text data necessary for semantic analysis.

2. Methodology

The authors constructed a massive, anonymized dataset from DerStandard, a major Austrian online newspaper, covering the period from 2013 to 2022.

Data Acquisition

• Source: Publicly available HTML pages on derstandard.at.
• Tools: Custom scripts using curl, xidel (XPath), and jq (JSON processing) to scrape article metadata, comments, voting records, and user interactions.
• Scope: The dataset spans 10 years, capturing over 75 million comments and 400 million votes.

Privacy and Anonymization

To comply with ethical standards and privacy regulations:

• Identifier Anonymization: All persistent identifiers (User IDs, Comment IDs) were replaced using a salted cryptographic hash function (BLAKE2s). This ensures that the same original ID maps to the same anonymized ID across all files (preserving relational links) without revealing the original identity.
• Text Handling: Raw comment texts are not released. Instead, the authors released pre-computed vector embeddings.
• Embedding Model: They utilized the "KaLM-embedding-multilingual-mini-v1" model (from Hugging Face), a specialized multilingual model.
  • Dimensions: 896-dimensional dense vectors.
  • Rationale: This allows for semantic analysis (clustering, similarity) without exposing the raw text, mitigating risks of re-identification or inference of sensitive attributes.

Data Structure & Pre-computation

The dataset is organized into several components to facilitate analysis:

• Articles: Yearly files with metadata and editorial topic tags (3-level hierarchy).
• Comments: Monthly files containing text embeddings, timestamps, and thread structure.
• Votes: Monthly files containing signed interactions (upvotes/downvotes).
• Users: A single comprehensive file with aggregated statistics (total comments, votes given/received, activity duration).
• Thread Reconstruction: Pre-computed fields for thread depth and root parent to allow rapid reconstruction of conversation hierarchies without re-processing raw data.

3. Key Contributions

• Scale and Scope: A unique dataset containing 75.6 million comments, 412.5 million votes, and 586,942 articles over a full decade.
• Signed Interactions: Unlike many datasets that only track presence, this includes explicit signed votes (up/down), enabling the study of agreement/disagreement dynamics and affective polarization.
• Mid-Resourced Language Focus: Provides a high-quality resource for German-language discourse analysis, a significant gap in current literature.
• Privacy-Preserving Semantic Data: The release of high-fidelity embeddings instead of raw text sets a precedent for sharing large-scale text data while protecting user privacy.
• Integration with External Data: The dataset is designed to align with political science datasets (e.g., Comparative Agendas Project, ParlaMint) and survey data (e.g., European Social Survey) due to its fine-grained temporal and topical structure.
• Faction Mapping: The authors provide a mapping of user IDs to two identified ideological factions (derived from prior work on affective polarization), enabling longitudinal studies of user alignment.

4. Results and Technical Validation

The authors performed rigorous technical validations to ensure data quality and the utility of the embeddings:

• Embedding Validity:
  • Structural Proximity: Cosine similarity analysis confirmed that embeddings of direct replies are semantically closer than non-replies, and comments within the same thread are closer than those across different threads.
  • Topical Cohesion: Comments sharing editorial topic tags (e.g., "football," "refugees") showed significantly higher semantic similarity than random pairs. Topics like "football" formed compact, distinct clusters in t-SNE visualizations.
• Data Consistency:
  • Missing Data: Approximately 8.46% of comment authors are missing (due to user deletion), which the authors retained to allow studies on user churn.
  • Anomalies: A small fraction of data (0.5% of comments) referenced deleted articles; timestamps for ~1,200 comments preceded their article timestamps. These were documented and flagged in summary files rather than discarded.
• Cross-Dataset Mapping: The authors successfully mapped 1,415 comments from the "One Million Post Corpus" and noted the limitations in linking the "GERMS-AT" dataset due to missing identifiers.

5. Significance

• Computational Social Science: The dataset offers a rare combination of scale, depth, and structure for analyzing online discourse, specifically enabling the study of network structures, temporal dynamics, and context-dependent interactions.
• Polarization Research: The availability of signed votes allows researchers to move beyond proxy measures (like edit wars or reply counts) to directly measure agreement and disagreement, providing empirical indicators of affective polarization.
• Reproducibility and Utility: By providing pre-computed embeddings and aggregated metadata, the authors lower the barrier to entry for researchers who cannot afford the computational resources to process 75 million raw texts.
• Ethical Benchmark: The approach of releasing embeddings rather than raw text demonstrates a viable path forward for sharing sensitive user-generated content in an era of strict privacy regulations.

Availability: The dataset is publicly available in the BSC Dataverse repository under a CC-BY license, accompanied by scripts for data processing and usage instructions.