PaperlayerView exact revision on arXiv

Source-linked AI summary

A Decade of Climate Polarization on Brazilian YouTube using Language Models

Daniel Morais, Diego H. M. Magalhaes, Gabriel H. Silva, Andrea Failla, Valeria de C. Santos, Helen C. S. C. Lima, Carlos H. G. Ferreira

arXiv:2609.27811v1cs.SIcs.CLcs.CY

TL;DR

Evidence on long-term climate discourse in Portuguese-language YouTube remains limited, despite the platform’s persistent comment threads. The study analyzes a decade of Brazil-oriented comments with scalable self-trained stance classification. Denialist comments are less prevalent but more associated with cross-stance contestation, while pro-consensus discourse is more reinforced within stance-homogeneous threads.

  • Problem

    Longitudinal evidence on climate stances and reply-based interactions on YouTube, especially in Portuguese-language discussions, remains limited.

    Source: §1 · p. 2, §2 · p. 4

  • Method

    The study analyzes a Brazil-oriented decade-long corpus of Portuguese-language YouTube comments using scalable stance detection and interaction- and frame-focused analyses.

    Source: §1 · p. 2, §5 · p. 13

  • Results

    Denialist comments are less prevalent but show more cross-stance contestation, while pro-consensus discourse is more reinforced within stance-homogeneous threads.

    Source: §1 · p. 3, §5 · p. 13

  • Takeaways & Limitations

    The findings suggest that countering denial requires more than scientific information, including trust-building, institutional transparency, and context-sensitive engagement.

    Source: §5 · p. 13

    Because annotation relies only on comment text, context-dependent stance expression may be underestimated.

    Source: §3.2 · p. 6

Abstract

from arXiv · show

Online platforms have become arenas for the public contestation of climate change, shaping how scientific knowledge, denial, and uncertainty are expressed and disputed. Yet longitudinal evidence remains limited for YouTube, especially for Portuguese-language discourse. Addressing this gap, we characterize how climate stances are expressed and contested over time in a large corpus of Portuguese-language YouTube comments retrieved through Brazil-oriented climate-related searches. To support this analysis in a noisy, imbalanced, and low-resource setting, we collect more than 240,000 comments posted between 2014 and 2024 and formulate stance detection as a three-way classification task (Believer, Denier, and Inconclusive). We operationalize stance attribution through a scalable self-training pipeline based on Llama 3.1, using Low-Rank Adaptation (LoRA) and hybrid instance selection to expand the training set with high-confidence pseudo-labeled examples while preserving class diversity. This approach improves coverage and class balance for minority and rhetorically complex classes, enabling large-scale stance attribution without extensive manual annotation. Our results show that polarisation is marked by interactional asymmetries: denialist comments are less prevalent, but they are associated with a comparatively higher share of cross-stance contestation, while pro-consensus discourse is more strongly reinforced within stance-homogeneous threads.

1 Introduction

The study addresses limited evidence on long-term Portuguese-language climate discourse on Brazilian YouTube, where stance and contestation unfold in persistent comment threads. It analyzes a decade of comments using scalable stance detection and examines how positions and frames appear in reply interactions.

1 Introduction

  • Prior evidence is concentrated on Twitter/X and short observation windows, leaving less known about long-term dynamics and reply-based interactions on YouTube.Persistent video-anchored threads may involve both reinforcement and conflict over time.Source: §1 · p. 2
  • The study examines more than 240,000 Portuguese-language comments from Brazil-oriented climate-related YouTube videos, posted between 2014 and 2024.It studies stance prevalence, temporal change, reply interactions, and discursive framing.Source: §1 · p. 2
  • The analysis classifies comments as Believer, Denier, or Inconclusive and uses Llama 3.1 self-training with LoRA and hybrid selection of high-confidence pseudo-labels.The selection approach aims to expand training data while preserving class diversity.Source: §1 · p. 2
  • Denialist comments are less frequent but show a larger share of cross-stance contestation, while pro-consensus discourse is more strongly reinforced in stance-homogeneous threads.The analysis also associates pro-consensus discourse with governance and responsibility frames, and denialist discourse with institutional distrust and conspiratorial reasoning.Source: §1 · p. 3

2 Related Work

Prior climate-discourse research documents polarization and recurring frames, but is concentrated on a limited set of platforms and relatively short observation windows. This study addresses the resulting evidence gap with a longitudinal, interaction-focused analysis of Portuguese-language YouTube discourse.

2 Related Work

  • Prior work establishes polarized, frame-dependent climate discourse but concentrates on Twitter/X and other short-text settings, with relatively short observation windows.Studies also cover Reddit and other communities, including scientific, political, and conspiratorial frames.Source: §2 · p. 3
  • YouTube’s persistent, video-anchored comment threads support sustained discussion, yet large-scale longitudinal climate-stance analyses remain scarce, especially in Portuguese-language and Global South contexts.The gap includes limited evidence on stance evolution and how replies structure reinforcement or conflict.Source: §2 · pp. 3–4
  • Computational climate-discourse methods include deep learning, topic modeling, fine-grained stance classification, and LLM-assisted interpretation, but often involve costly annotation, limited temporal scope, or short-text platforms.Source: §2 · p. 4
  • This work combines scalable stance detection with interaction-focused and frame-oriented analysis of a decade-long Brazil-oriented Portuguese-language YouTube corpus.It examines both stance prevalence and reply-based contestation in an underexplored linguistic and regional context.Source: §2 · p. 4

3 Methodology

The methodology collects and filters Brazil-oriented Portuguese-language YouTube comments, manually annotates a subset, and builds a self-trained stance classifier. The resulting labels support analysis of stance prevalence, reply interactions, and recurring frames over the ten-year period.

3.1 Data Collection and Pre-processing

  • The corpus contains 247,514 Portuguese comments from 1,020 videos and 478 channels, collected from 2014–2024 using Brazil-oriented, climate-related searches.The team filters for Portuguese and climate relevance and retains videos with more than 30 comments; API relevance ranking is platform-defined and not externally reproducible as a relevance measure.Source: §3.1 · pp. 4–5
  • The dataset cannot be publicly redistributed under YouTube API policy; the authors plan to provide video IDs, query configuration, and scripts for authorized researchers to rehydrate it.Source: §3.1 · p. 5

3.2 Comment Annotation

  • Two authors annotate individual comments as Believer, Denier, or Inconclusive; the final labeled set contains 4,008 comments, with Cohen’s kappa of 0.88.Believer labels cover agreement with climate science, Denier labels skepticism or rejection of anthropogenic responsibility, and Inconclusive labels apply when stance cannot be determined from text alone.Source: §3.2 · pp. 5–6
  • Text-only annotation excludes video, title, channel, parent-comment, and thread context, so context-dependent stance expression may be underestimated.This conservative protocol avoids using information unavailable to the classifier; generic agreement, irony, and sarcasm are labeled Inconclusive when the target stance is unclear.Source: §3.2 · p. 6

3.3 Model Training and Evaluation

  • The classifier uses Llama 3.1 8B with a task-specific classification head, LoRA adaptation, and imbalance-aware training; evaluation uses stratified five-fold cross-validation and macro F1 as the primary metric.The annotated dataset is substantially imbalanced, and class weights are inversely proportional to class frequencies.Source: §3.3 · pp. 6–7
  • Self-training adds high-confidence pseudo-labels using class-wise thresholds and controlled random sampling, then samples inversely to original class proportions to increase minority-class representation.The added pseudo-labeled instances equal 50% of the manual dataset; the resulting classifier is intended for corpus attribution, not evaluation against an independent gold-standard test set.Source: Self-Training and Hybrid Instance… · pp. 7–8

3.4 Large-Scale Stance Characterization

  • The final self-trained classifier labels the inference corpus for analysis of stance distribution, reply interactions, and discursive framing across the decade.The analysis measures yearly reply composition to Believer- and Denier-initiated comments and examines high-engagement polarized contexts and representative exchanges.Source: §3.4 · p. 8

4 Results

The enhanced classifier improves macro-level and minority-class performance, while the corpus is mostly inconclusive and explicitly polarized threads show asymmetric cross-stance engagement. Believer discourse centers on governance and responsibility, whereas Denier discourse invokes distrust and institutional control.

4.1 Classifier Diagnostics

  • 0.677 ± 0.029 to 0.789 ± 0.020: Macro-F1 increased for the enhanced model over the supervised baseline (p < 0.001).The absolute gain was 0.112, with higher diagnostic scores across all evaluated macro-level metrics.Source: Table 1 · p. 9, §4.1 · p. 9
  • 0.610 ± 0.026 to 0.729 ± 0.024 for Believer F1 and 0.611 ± 0.060 to 0.810 ± 0.026 for Denier F1; Inconclusive gains were not statistically significant.Believer and Denier gains were significant across precision, recall, and F1, while the Inconclusive improvements were positive but not significant.Source: §4.1 · p. 9

4.2 Large-Scale Stance Distribution and Interaction Dynamics

  • 80% to 88% of replies in Believer-initiated threads were Believer replies from 2016 onward, while Believer replies made up 29% to 49% of Denier-initiated threads from 2015 onward.Denier replies generally remained below 20% in Believer-initiated threads; overall, Denier comments attracted a comparatively higher share of cross-stance contestation.Source: Figure 1 · p. 10, Figure 2 · p. 10, §4.2 · pp. 10–11
  • 64.9% of the 242,367-comment inference corpus was Inconclusive, compared with 19.3% Believer and 15.7% Denier comments.Despite the majority of comments being inconclusive, 84,953 comments expressed a clear stance toward climate change.Source: §4.2 · p. 10

4.3 Discursive Frames and High-Engagement Polarization

  • Believer discourse frames climate change around governance, economic responsibility, and environmental justice, while Denier discourse invokes institutional distrust and conspiratorial anti-systemic narratives.Believer comments often treat scientific consensus as a premise for claims about political inaction and accountability; Denier rhetoric shifts attention toward intent, power, and legitimacy.Source: §4.3 · p. 11
  • In the top comment-volume decile, Believer videos commonly emphasize governance failure and climate impacts, while Denier videos contest human causality or frame climate science as politically motivated.Videos presenting concrete visual evidence of climate impacts can still become battlegrounds attracting denialist scrutiny rather than producing consensus.Source: §4.3 · pp. 11–12
  • A relatively small subset of polarized discourse can structure conflict and visibility within a much larger, predominantly inconclusive communicative landscape.Believers mobilize scientific authority to demand responsibility and collective action, whereas Deniers frame climate science as ideological or institutional control.Source: §4.3 · p. 12, Table 2 · p. 12

5 Conclusion and Future Work

The decade-long analysis finds that polarized stances disproportionately structure debate, with denialist comments drawing more cross-stance contestation and distinct frames separating pro-consensus from denialist discourse. The authors propose future comparisons across Global South countries and links to salient events and policy debates.

5 Conclusion and Future Work

  • Examples were translated from Portuguese and paraphrased for privacy while retaining their original meaning, stance, and argumentative structure.Source: §5 · p. 12
  • Most comments were Inconclusive, but polarized stances disproportionately structured debate; Denier comments were less prevalent yet attracted more cross-stance contestation.The authors describe denial as a recurring catalyst of conflict.Source: §5 · pp. 12–13
  • Pro-consensus discourse emphasizes governance, responsibility, and collective action, whereas denialist discourse reflects institutional distrust, conspiratorial reasoning, and suspicion toward expert authority.The authors suggest that responding to denial requires combining scientific evidence with trust-building, institutional transparency, and context-sensitive engagement.Source: §5 · p. 13
  • The self-training approach enables scalable stance detection in noisy, imbalanced, low-resource settings, supporting analyses difficult to conduct through manual annotation alone.Source: §5 · p. 13
  • Future work proposes comparing Global South countries, tracking evolving narratives, and linking online discourse to environmental disasters, elections, policy debates, and media coverage.These comparisons could assess whether asymmetries recur under different political and environmental conditions.Source: §5 · p. 13

A Search Queries for Data Collection

The data collection strategy uses Brazil-oriented climate-related searches, mostly in Portuguese, to retrieve YouTube content spanning scientific consensus and climate skepticism. Queries cover scientific, environmental, political, institutional, and denial-related themes.

A Search Queries for Data Collection

Loading 2609.27811v1…