All work
News articles
1,054
Sentiment accuracy, SVM
68.72%
Four-stance accuracy, SVM
62.56%
Majority-class baseline
50%

Data and labels

Articles came from NewsAPI, MediaStack and Google News RSS. Each one was cleaned, tokenized, stemmed and lemmatized, then turned into count and TF-IDF features. A BERT sentiment model assigned each article a stance.

Bar chart of stance labels for 1,054 articles: 528 anti-regulation, 277 strongly pro-regulation, 189 weakly pro-regulation and 60 neutral.
Half of the coverage (528 of 1,054 articles) argues against holding platforms responsible.

Exploring the coverage

  • K-Means with silhouette analysis, and hierarchical clustering with cosine distance in R.
  • LDA topic modeling with five topics, which separates regulatory, technological, political and user-focused threads.
  • Apriori association rules in R.

Predicting stance

Naive Bayes, decision trees (entropy and Gini at several depths) and SVMs with linear, RBF and polynomial kernels were trained to predict each article's label.

Bar chart of accuracy on held-out articles: SVM RBF on sentiment 68.72%, SVM RBF on four stances 62.56%, linear SVM 59.24%, Naive Bayes 56.40%, best decision tree 51.66%.
Accuracy on held-out articles.

With four uneven classes, always guessing "anti-regulation" scores about 50%. The four-stance models beat that by a modest margin, and the three-way sentiment model does clearly better.

Limitations

The labels come from a BERT model, not from human annotators, so these accuracies measure agreement with that model's labels. The study maps how coverage frames the question; it does not judge whether any article is true.