<![CDATA[The advancement of digital technology has transformed how society accesses and responds to political information, particularly through platforms like YouTube, which serve as arenas for public discourse. Comments on political speech videos often contain complex sentiments such as irony, slang, and code-mixing, which are difficult to identify using traditional sentiment analysis methods. This study aims to analyze public sentiment toward the Indonesian President’s political speeches on YouTube from 2014 to 2024 using the FastText word embedding approach and to compare its performance with the TF-IDF + Logistic Regression method. The evaluation was conducted on three sentiment classes using automatically labeled data and oversampling experiments to address class imbalance. The results show that FastText achieved an accuracy of 76.82%, slightly higher than TF-IDF + Logistic Regression at 74.11%. Although the difference in accuracy is relatively small, the FastText model demonstrated more stable performance on informal texts and varied contexts. The use of oversampling helped balance predictions across classes without significantly improving accuracy. This study highlights the potential of FastText to enhance the effectiveness of Indonesian-language sentiment analysis, particularly for political comments on social media, while also revealing the limitations of automatic labeling that may affect classification outcomes.]]>
Copyrights © 2025
