Audience Engagement and Narrative Resistance on Telegram: A Dataset of Sentiment Shifts, Semantic Similarity, and Zero-Shot Classification

I. Data Files (.csv)

• Merged_Post_Comments_Data_anonymized.csv: The master dataset. It contains merged metrics for articles and comments, including views, reactions, and the probability scores for the six narrative classes (e.g., Military Success, Economic Hardship).

• article_sentiment_scores_all_extended.csv: Contains sentiment scores and engagement metadata (total reactions, comment counts) for the source posts/articles across different Telegram portals.

• comment_sentiment_scores_all_extended_ssot.csv: It includes individual sentiment scores, total reactions per comment, and semantic similarity scores relative to the parent article.

• semantic_similarity_over_time_all.csv: A longitudinal dataset tracking how the semantic alignment of the audience’s conversation shifts in the hours following an article's publication.

II. Analysis Scripts (.py)

• Zero-Shot_Narrative_Classification.py: The Python pipeline used to categorize Telegram posts into six distinct narrative frames using a valhalla/distilbart-mnli-12-1 (or similar) transformer model.

• semantic_analysis_4tg_gpu.py: The core computational script used to generate sentiment scores and calculate cosine similarity between article embeddings and comment embeddings using GPU acceleration.