Scholarlink

This work introduces Stance-EC, the first large-scale Spanish-language dataset for stance detection developed in Ecuador. It is centered on the highly polarized 2022 protests. Utilizing X (formerly Twitter)’s API, we downloaded tweets relevant to the event using specific hashtags. After deduplication, we obtained a corpus of 232,609 tweets, from which a random sample of 25,454 was selected for manual annotation. Tweets were categorized as ‘favor’, ‘against’, or ‘neutral’ towards the protest, totaling 6,756, 7,066, and 11,632 in each category. We used machine learning classifiers including Random Forest, SVM, Logistic Regression, Decision Tree, CatBoost, XGBoost, and LightGBM for unigram and bigram analysis, to test the dataset’s capability, evaluating performance with Accuracy, Precision, Recall, and F1-score. Results show strong performance across all metrics, typically around 70% for both unigrams and bigrams, and even exceeding 80% in LightGBM with unigrams, highlighting the dataset’s reliability for future research. Dataset is available at: https://github.com/Leo-Thomas/Stance-EC.git .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Stance-EC: A Spanish Dataset of Political Discourse on X/Twitter for Stance Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mike Bermeo,
  • Leo Thomas Ramos,
  • Silvana Escobar-Córdova,
  • Diego Morales-Navarrete,
  • Erick Cuenca

摘要

This work introduces Stance-EC, the first large-scale Spanish-language dataset for stance detection developed in Ecuador. It is centered on the highly polarized 2022 protests. Utilizing X (formerly Twitter)’s API, we downloaded tweets relevant to the event using specific hashtags. After deduplication, we obtained a corpus of 232,609 tweets, from which a random sample of 25,454 was selected for manual annotation. Tweets were categorized as ‘favor’, ‘against’, or ‘neutral’ towards the protest, totaling 6,756, 7,066, and 11,632 in each category. We used machine learning classifiers including Random Forest, SVM, Logistic Regression, Decision Tree, CatBoost, XGBoost, and LightGBM for unigram and bigram analysis, to test the dataset’s capability, evaluating performance with Accuracy, Precision, Recall, and F1-score. Results show strong performance across all metrics, typically around 70% for both unigrams and bigrams, and even exceeding 80% in LightGBM with unigrams, highlighting the dataset’s reliability for future research. Dataset is available at: https://github.com/Leo-Thomas/Stance-EC.git .