Scholarlink

The excessive growth of opinions or emotions in different social networks has enabled the optimization of sentiment analysis (SA). However, the data labeling phase usually takes a lot of time and resources if manual labeling is performed, due to the amount of data. The present study seeks to use bidirectional transformer models (pysentimiento) to label tweets in Spanish, considering political and health events in Peru. For this purpose, a methodology for the construction of a dataset is proposed, which includes the extraction of tweets by Twint, the filtering and cleaning of data, the labeling of tweets by pysentimiento, the balancing of classes, and the manual verification of tweets. Through the experimental phase, where traditional supervised models-naive Bayes (NB), super vector machine (SVM), random forest (RF), and multilayer perceptron (MP)-are trained with the proposed datasets, the improvement in the performance of each model is shown by cross validation and the stratification of classes. The results show that manual verification contributes to an optimal classification of tweets in the positive and negative categories, but not in the neutral category.

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

Tagging with pysentimiento and its Evaluation in Traditional Supervised Learning Models for Sentiment Analysis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Claudia Francesca Suarez Mariscal,
  • Yeshica Isela Ormeño Ayala,
  • Jose Mauro Pillco Quispe,
  • Lino Prisciliano Flores Pacheco

摘要

The excessive growth of opinions or emotions in different social networks has enabled the optimization of sentiment analysis (SA). However, the data labeling phase usually takes a lot of time and resources if manual labeling is performed, due to the amount of data. The present study seeks to use bidirectional transformer models (pysentimiento) to label tweets in Spanish, considering political and health events in Peru. For this purpose, a methodology for the construction of a dataset is proposed, which includes the extraction of tweets by Twint, the filtering and cleaning of data, the labeling of tweets by pysentimiento, the balancing of classes, and the manual verification of tweets. Through the experimental phase, where traditional supervised models-naive Bayes (NB), super vector machine (SVM), random forest (RF), and multilayer perceptron (MP)-are trained with the proposed datasets, the improvement in the performance of each model is shown by cross validation and the stratification of classes. The results show that manual verification contributes to an optimal classification of tweets in the positive and negative categories, but not in the neutral category.