💻 computer science

Who Shapes Brazil's Vaccine Debate? Semi-Supervised Modeling of Stance and Polarization in YouTube's Media Ecosystem

This study presents the largest longitudinal analysis of Brazil's vaccine discourse on YouTube, utilizing a semi-supervised framework to classify 1.4 million comments and reveal how polarization spikes during crises while becoming fragmented in the post-pandemic era, with science communicators and digital-native outlets emerging as key drivers of both supportive and oppositional engagement.

Original authors: Geovana S. de Oliveira, Ana P. C. Silva, Fabricio Murai, Carlos H. G. Ferreira

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Geovana S. de Oliveira, Ana P. C. Silva, Fabricio Murai, Carlos H. G. Ferreira

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 Brazil's YouTube channel as a massive, bustling town square. For seven years, this square has been the main place where people gather to talk about vaccines. Sometimes they are cheering, sometimes they are shouting, and sometimes they are just chatting about the weather.

This paper is like a team of digital detectives who decided to map out exactly what's happening in this town square. They wanted to answer three big questions: How can we teach a computer to understand these messy conversations? How has the mood changed over time? And who are the people actually leading the shouting matches?

Here is the story of their investigation, broken down simply:

1. The Problem: A Mountain of Confusing Noise

The researchers had a huge problem. They had 1.4 million comments about vaccines. That's like trying to read every single letter in a library overnight.

Most of these comments weren't clear "Yes" or "No" answers. Many were just "Hmm, interesting," or "I don't know," or "My cousin said..." (which the researchers call "Inconclusive"). Because there were so many "maybe" comments and so few clear "yes/no" ones, a standard computer program would get confused and just guess "maybe" for everything to be safe.

2. The Solution: The "Smart Tutor" Method

To fix this, the researchers didn't just hire a computer; they gave it a tutor.

Step 1 (The Manual Lesson): A small group of humans read about 3,500 comments and taught the computer what "Pro-Vaccine," "Anti-Vaccine," and "Inconclusive" actually look like.
Step 2 (The Self-Teaching): The computer then looked at the remaining 1.4 million comments. But instead of guessing blindly, it only picked the ones it was 100% sure about. It said, "I'm so confident this is a 'Yes' that I'll teach myself from it."
The Result: This "self-teaching" method made the computer much smarter. It learned to spot the subtle differences between a skeptic and a supporter, even when they were using tricky language.

3. The Timeline: From Calm to Chaos to Fragmentation

The researchers looked at the town square through three different time lenses:

Before the Pandemic (2018–2020): The square was relatively calm. People mostly agreed with each other if they were pro-vaccine. The "anti-vaccine" crowd was quiet and didn't argue back much.
During the Pandemic (2020–2023): Chaos! The square exploded. Everyone started shouting at once. The pro-vaccine and anti-vaccine groups started fighting back and forth with equal intensity. It was a perfect storm of arguments.
After the Pandemic (2023–2024): The shouting didn't stop, but the crowd changed. The "anti-vaccine" group started sticking together in tight, loud cliques, reinforcing their own beliefs. The "pro-vaccine" group, however, seemed to lose its unity and started arguing less with each other. The debate became more fragmented.

The Analogy: Think of it like a sports game. Before the pandemic, the home team (pro-vaccine) was winning easily. During the pandemic, it became a fierce, tied match with both sides screaming. After the pandemic, the visiting team (anti-vaccine) started huddling together and cheering louder, while the home team scattered a bit.

4. The Cast of Characters: Who is Leading the Shouting?

The researchers wanted to know: Who is the loudest voice in the square?

You might guess it's the big news stations (like CNN or BBC). But the study found something surprising: The news stations are just the stage; the real actors are the scientists and the internet personalities.

The News Stations (Legacy Media): They set the agenda (they say, "Let's talk about the flu shot"), but they aren't the ones getting into the biggest fights.
The Science Communicators (The "Cool Teachers"): Channels run by doctors and science YouTubers (like Olá, Ciência!) are the main battlegrounds. They are the most popular, but they are also the places where the fiercest arguments happen. People trust them, but they also attack them the most.
The Digital Native Commentators: These are independent creators who aren't part of the big news networks. They are surprisingly powerful in spreading both support and skepticism.

The Metaphor: Imagine a school. The Principal (the News) announces a new rule. But the real drama happens in the hallways with the Science Teachers (who are trying to explain the rule) and the Student Influencers (who are either cheering for the rule or organizing a protest). The Principal is just watching from the office.

5. The Big Takeaway

The most important lesson from this paper is that polarization isn't just about who is watching; it's about who is talking to whom.

Misinformation isn't the only villain: Sometimes, videos that are trying to be neutral or explain complex medical issues (like "Does this vaccine cause blood clots?") actually create the most fighting because they touch on fear and uncertainty.
The "Inconclusive" crowd is huge: Most people aren't screaming "Yes" or "No." They are just watching, confused. This is a big opportunity for public health officials to reach out before people get pulled into the shouting matches.

In a nutshell: The study shows that to understand the vaccine debate in Brazil, you can't just look at the news. You have to look at the "cool teachers" and the "internet influencers" on YouTube, because that's where the real conversation—and the real fighting—is happening. And the best way to understand it is to use a smart computer that knows how to listen to the nuances of human speech.

1. Problem Statement

The paper addresses the critical challenge of understanding long-term vaccine discourse dynamics in non-English, high-impact contexts, specifically Brazil. While previous computational studies on vaccine hesitancy have focused on:

English-language data: Limiting applicability to global contexts like Brazil.
Narrow time windows: Often restricted to the peak of the COVID-19 pandemic.
Specific vaccines: Focusing on single vaccines (e.g., COVID-19, HPV) rather than the full National Immunization Program (PNI).
Lack of structural analysis: Failing to map how different media actors (legacy news, science communicators, digital-native outlets) shape polarization.

The authors aim to fill these gaps by investigating how public discourse and polarization evolved across Brazil's YouTube ecosystem from 2018 to 2024, covering the pre-pandemic, pandemic, and post-pandemic periods.

2. Methodology

A. Data Collection and Preprocessing

Dataset: A longitudinal dataset comprising 1,422,406 comments from 14,318 videos across 3,897 channels.
Scope: Covers 19 vaccines from the Brazilian National Immunization Schedule (PNI).
Timeframe: January 2018 to July 2024.
Filtering:
- Excluded educational content (e.g., "exam," "course") to retain public discourse.
- Filtered for Portuguese language using a hybrid approach (fine-tuned RoBERTa + langdetect).
- Used the YouTube Data API v3 to retrieve metadata, comments, likes, and view counts.

B. Stance Detection Framework (RQ1)

To classify the massive, imbalanced dataset, the authors developed a semi-supervised pipeline:

Manual Annotation: A reference set of 3,476 comments was manually labeled by three annotators into three classes: Favorable, Against, and Inconclusive. The dataset exhibited severe class imbalance (74.4% Inconclusive).
Supervised Baseline: A Llama 3.1 8B model was fine-tuned using QLoRA (Quantized Low-Rank Adaptation) with a weighted cross-entropy loss to handle class imbalance.
Semi-Supervised Self-Training:
- The baseline model generated predictions for the remaining ~1.3M unlabeled comments.
- Confidence-based Selection: Predictions were filtered using Shannon Entropy. Only low-entropy (high-confidence) predictions were selected as pseudo-labels.
- Inverse-Frequency Sampling: To counter the dominance of the Inconclusive class, the selection process prioritized high-confidence samples from minority classes (Favorable and Against).
- Retraining: A second-stage classifier was trained on the combined dataset (original manual labels + high-confidence pseudo-labels).

C. Discourse and Ecosystem Analysis (RQ2 & RQ3)

Temporal Dynamics: Analyzed stance distribution across three phases (Pre-pandemic, Pandemic, Post-pandemic) and mapped comments to specific vaccines using a custom dictionary.
Interaction Analysis: Calculated conditional reply probabilities ( $P(s_{reply} | s_{parent})$ ) to measure antagonistic vs. supportive interactions.
Channel Taxonomy: Classified channels into Certified (ANJ-registered legacy media) and Non-Certified (Science communicators, Digital-native commentary). Metrics included engagement rates and polarization density per channel type.

3. Key Contributions

Methodological Advancement: Demonstrated that semi-supervised self-training with entropy-based filtering significantly improves stance detection on highly imbalanced social media data, outperforming purely supervised baselines, particularly for minority classes.
Longitudinal Scope: Provided the first large-scale, seven-year analysis of vaccine discourse in Brazil, covering the full immunization schedule rather than just COVID-19.
Structural Mapping of Polarization: Mapped the "hybrid media ecosystem," revealing that polarization is not driven primarily by legacy journalism but by science communicators and digital-native channels.
Open Science: Released the trained stance detection model and the dataset to the community for reproducibility.

4. Key Results

A. Model Performance (RQ1)

The semi-supervised approach yielded substantial improvements over the supervised baseline:

Favorable Class: F1-score increased from 0.57 to 0.91.
Against Class: F1-score increased from 0.67 to 0.88.
Inconclusive Class: Maintained high performance (F1 ~0.94) with reduced variance.
Conclusion: The low-entropy self-training strategy effectively mitigated class imbalance and improved robustness.

B. Temporal Dynamics and Polarization (RQ2)

Volume: The Inconclusive stance dominates (74.4%), indicating that most interactions are informational or neutral.
Polarization Spikes: Polarization (explicit Favorable vs. Against) nearly doubled during the pandemic (13.6% $\to$ 26.0%).
Interaction Patterns:
- Pre-pandemic: Asymmetric; pro-vaccine users actively corrected skeptics.
- Pandemic: Symmetric contestation; both sides engaged equally (0.43 probability of cross-stance replies).
- Post-pandemic: Asymmetric again, but shifted. Skeptical groups showed strong in-group reinforcement (0.67), while pro-vaccine cohesion weakened (0.50).
Vaccine Specificity: Crisis events (e.g., Yellow Fever outbreak in 2018, COVID-19 in 2020) concentrated attention on specific vaccines, while post-crisis periods saw fragmented discourse across the schedule.

C. Sources of Dissemination (RQ3)

Non-Certified Dominance: 82.4% of anti-vaccine comments occurred on non-certified channels.
Science Communicators as Battlegrounds: Channels like Julio Pereira – Neurocirurgião and Olá, Ciência! were the top hubs for both supportive and opposing discourse. They function as arenas where scientific authority is actively contested.
Legacy Media Role: Traditional news outlets (e.g., BBC Brasil, CNN Brasil) set the agenda but were not the primary sites of intense polarized engagement.
Video Content: Highly polarized videos often addressed ambiguous or controversial topics (e.g., side effects, AstraZeneca thrombosis) rather than overtly anti-vaccine manifestos. "Polarity magnets" included videos staging debates between skeptics and experts.

5. Significance and Implications

Public Health Strategy: The findings suggest that public health interventions cannot rely solely on traditional media. Science communicators on platforms like YouTube are critical leverage points; they are trusted but also the primary sites of epistemic conflict.
Platform Governance: Moderation strategies focusing only on "fake news" may be insufficient. The study highlights that polarization is embedded in the interactional structure of conversations (e.g., how replies reinforce in-group identity) and is amplified by algorithmic visibility of controversial topics.
Computational Social Science: The paper validates semi-supervised learning as a scalable solution for analyzing massive, imbalanced, and ambiguous text corpora in low-resource languages, offering a blueprint for future longitudinal studies of misinformation and polarization.

In summary, the paper reveals that Brazil's vaccine debate is shaped by a hybrid ecosystem where digital-native and science communication channels drive polarization, and that crisis periods fundamentally restructure the interactional dynamics between pro- and anti-vaccine communities.