💬 NLP

A Longitudinal, Multinational, and Multilingual Corpus of News Coverage of the Russo-Ukrainian War

This paper introduces DNIPRO, a multilingual corpus of 246,000 news articles from five nations covering the Russo-Ukrainian War, which features human-annotated metadata on stance, sentiment, and framing to enable the empirical study of divergent geopolitical narratives and information warfare.

Original authors: Dikshya Mohanty, Taisiia Sabadyn, Jelwin Rodrigues, Chenlu Wang, Abhishek Kalugade, Ritwik Banerjee

Published 2026-03-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dikshya Mohanty, Taisiia Sabadyn, Jelwin Rodrigues, Chenlu Wang, Abhishek Kalugade, Ritwik Banerjee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Russo-Ukrainian war not just as a battle of tanks and soldiers, but as a massive, global storytelling contest. Every country is a narrator, sitting in a different room, telling the same story but with completely different characters, motives, and endings.

This paper introduces dnipro (which means "Dnieper," a major river in Ukraine), a massive digital library built to help us understand how these different stories are being told.

Here is the breakdown of what the researchers did, using some simple analogies:

1. The "Time-Traveling Library"

Imagine you wanted to study how a story changes over time. You'd need a library that has every newspaper, every day, for every country involved.

What they built: A collection of 246,000 news articles collected over 31 months (from Feb 2022 to Aug 2024).
The Scope: It's like having a window into five different living rooms: Russia, Ukraine, the USA, the UK, and China.
The Languages: They didn't just read English; they translated articles from Russian and Chinese into English so everyone could compare them side-by-side.

2. The "Translator and Detective" Team

Reading 246,000 articles is impossible for a human. So, the researchers used a team of AI detectives (specifically advanced language models like Phi-4) to do the heavy lifting.

The Translator: First, the AI translated all the Russian and Chinese articles into English.
The Labeler: Then, the AI acted like a librarian tagging books. It went through every article and attached three specific tags:
1. Sentiment: Is the article happy, sad, angry, or neutral about a specific person (like "Ukraine" or "Putin")?
2. Stance: Is the article for or against a specific group?
3. Framing: What is the main theme? Is the story about Security (war, bombs), Economy (sanctions, oil prices), Politics (elections, leaders), or Humanitarian issues (refugees, hospitals)?

3. The "Echo Chamber" Discovery

The most exciting part of the paper is what they found when they compared the stories. They discovered that the countries aren't just lying to each other; they are building parallel realities.

Think of it like two people watching the same car crash:

Person A (Russian Media): Says, "The car was driving safely, but a pedestrian jumped in front of it. The pedestrian is at fault."
Person B (Ukrainian/Western Media): Says, "The car was speeding and hit a pedestrian who was standing still. The driver is at fault."

Crucially, neither side explicitly says, "You are lying." They just tell their own version of the story so convincingly that their audiences believe their version is the only truth. The paper calls this "constructing incompatible realities."

4. What the Data Revealed

By using their "library," the researchers saw some fascinating patterns:

The Security Blanket: Every single country, no matter who they are, talks mostly about Security (war, defense, violence). It's the dominant theme everywhere.
The Economic Lens: China and the UK focused heavily on Economy (trade, sanctions), while the US and UK focused more on Politics and Policy.
The Moral Silence: Surprisingly, very few articles talked about Morality (right vs. wrong, ethics). The war was discussed in terms of strategy and survival, not ethics.
The Shift: They tracked how opinions changed over time. For example, Chinese media started slightly pro-Ukraine but slowly shifted to be more neutral or pro-Russia as the war dragged on. Russian media went from "mildly negative" to "overwhelmingly hostile."

5. Why This Matters

In the past, if you wanted to study war propaganda, you might have looked at social media (Twitter/X), which is full of noise, bots, and short posts.
dnipro is different because it looks at professional news outlets. It shows us that the "information war" isn't just about fake news; it's about framing.

The Big Takeaway:
The paper gives us a tool to see how nations use news to shape what their citizens believe. It shows that in a modern war, the battle isn't just for territory; it's for the narrative. By having this dataset, researchers can now mathematically prove how these different realities are built, helping us understand why people in different countries can look at the same event and see two completely different worlds.

In short: The researchers built a giant, multilingual mirror that reflects how the world sees the war, revealing that we are all looking at the same event but seeing different pictures.

1. Problem Statement

The Russo-Ukrainian War (escalating Feb 2022) has been characterized by a "battle for narrative control," where competing geopolitical actors construct incompatible realities through media framing. While existing datasets exist, they suffer from significant limitations:

Scope: Most are limited to social media (Twitter/X, Telegram), which lacks the formal structure and editorial positioning of traditional news outlets.
Language & Geography: Existing resources are often monolingual or restricted to specific regions (e.g., only Russian or only Western media).
Temporal Coverage: Few datasets offer continuous, longitudinal coverage spanning the entire duration of the conflict.
Annotation Depth: Many lack comprehensive, human-evaluated annotations for stance, sentiment, and topical framing across multiple languages and sources.

There is a critical need for a structured, multilingual, and multinational corpus to empirically study how nations frame the war, track narrative evolution, and detect implicit contradictions without relying on explicit rebuttals.

2. Methodology

Data Acquisition and Construction

The authors constructed dnipro, a corpus of 246,229 news articles collected over 31 months (February 1, 2022 – August 31, 2024).

Sources: 11 influential media outlets aligned with 5 geopolitical actors:
- Russia: TASS, Sputnik (English), Izvestia (Russian).
- Ukraine: Censor.net, The Kyiv Independent, European Pravda (English).
- USA: CNN.
- UK: The Financial Times.
- China: Xinhua, Global Times, China Daily (Mandarin and English).
Languages: English, Russian, and Mandarin Chinese.
Collection Strategy: A tiered approach was used. For outlets with dedicated war sections, all articles were collected. For others, keyword filtering or manual browser-based retrieval was employed to overcome API restrictions.
Preprocessing: Articles were stripped of boilerplate, duplicates removed, and non-English texts (Russian/Chinese) translated to English using the M2M100-418M model (selected via human evaluation for fidelity and fluency).

Annotation Pipeline

Since full text cannot be distributed due to copyright, the corpus provides structured metadata and derived annotations:

Named Entity Recognition (NER): Performed using spaCy (OntoNotes 5.0 taxonomy) to identify actors and locations.
Sentiment Analysis: Sentence-level sentiment scores for key entities (e.g., "Ukraine," "Zelensky") generated using a pretrained RoBERTa-large model, aggregated daily/monthly.
Stance Detection: Using Phi-4 (14B) via few-shot prompting to classify sentences as favor, against, or none toward specific targets.
Topical Framing: Using Phi-4 to assign articles to one of nine frames based on Habermas's discourse ethics (e.g., Security, Economy, Morality, Legal, Politics, etc.), outputting labels with confidence scores.
Claim Typology: In use-case experiments, claims were categorized (Factual, Attributed, Disputed, Reported, Command/Policy, Analytical) using GPT-3.5.

Quality Control

Human Evaluation: Two native English annotators validated a stratified sample of 150–200 instances.
Metrics:
- Stance: Human inter-annotator agreement ( $\kappa$ ) was 0.77; LLM-human agreement was moderate ( $\kappa \approx 0.60$ ).
- Topical Framing: Prevalence-adjusted bias-adjusted kappa (pabak) averaged 0.71. Frames like Welfare, Public Opinion, and Legal showed near-perfect agreement ( $>0.8$ ), while Policy and Politics showed moderate agreement due to conceptual overlap.

3. Key Contributions

The dnipro Corpus: The largest longitudinal, multinational, and multilingual news corpus on the Russo-Ukrainian war to date, covering 31 months and 5 nations.
Rich Annotation Schema: Provides entity-level stance, sentiment, and multi-label topical framing, enabling comparative analysis across linguistic and geopolitical divides.
Methodological Framework: Demonstrates a pipeline for constructing large-scale, copyright-compliant datasets using machine translation and LLM-based annotation validated by human experts.
Empirical Insights: Proofs-of-concept showing how the corpus reveals "implicit contradictions" where outlets construct incompatible realities through divergent attribution and framing rather than direct refutation.

4. Results and Findings

The paper presents several exploratory analyses demonstrating the corpus's utility:

Thematic Salience:
- Security dominates all outlets, but Economy is prominent in Chinese media and The Financial Times.
- Russian/Chinese outlets emphasize institutional terminology ("special military operation") and security concerns.
- Ukrainian outlets focus on humanitarian dimensions and specific conflict locations.
- Western outlets balance humanitarian topics with institutional responses (sanctions, ICC).
Temporal Stance Shifts:
- China: Shifted from mildly pro-Ukrainian to mildly oppositional, reflecting strategic alignment with Russia.
- Russia: Moved from mildly negative to overwhelmingly hostile.
- USA: Became increasingly supportive.
- UK: Showed high variability, likely due to war fatigue and domestic pressures.
Narrative Divergence:
- Contradictions are rarely explicit rebuttals. Instead, outlets construct "parallel realities."
- Example: Regarding the Bucha Massacre, Russian media attributed atrocities to Ukrainian nationalists, while Ukrainian media attributed them to Russian forces. Both used "factual" or "attributed" claims without directly negating the other's narrative, relying instead on divergent attribution and selective emphasis.

5. Significance

For Research: dnipro enables the study of information warfare and narrative evolution at a scale previously impossible. It allows researchers to move beyond "fake news" detection to understanding how competing realities are constructed through framing and omission.
For Policy & Journalism: The dataset helps identify how geopolitical blocs align or diverge over time and how specific events trigger narrative shifts.
Methodological Impact: It validates the use of LLMs (Phi-4) for large-scale annotation when combined with rigorous human validation, offering a blueprint for future conflict-related corpus construction.
Accessibility: The corpus is publicly available (CC-BY-NC 4.0) with metadata and annotation files, though full text is restricted to URLs to respect copyright.

Limitations

Temporal Scope: Ends in August 2024; does not capture post-2024 developments.
Geographic Gaps: Limited NATO representation (only US/UK); excludes key regional actors like Germany, Hungary, or India.
Translation Nuances: Machine translation may lose culturally specific rhetoric or connotations.
Annotation Reliability: While substantial, automated annotations are not perfect for fine-grained, high-stakes nuance without human review.

In conclusion, dnipro is a foundational resource for computational journalism and political science, offering a unique lens into how major global powers construct and disseminate conflicting narratives during a prolonged geopolitical conflict.