Analisis Perbandingan TF-IDF, FastText, dan BERT pada Model Support Vector Machine untuk Klasifikasi Sentimen Komentar Folkative

Abstract

Sentiment analysis on social media presents its own challenges due to the use of informal language, slang, sarcasm, and typos commonly found on platforms like Instagram. This study aims to apply and compare three text feature representation methods, namely TF-IDF as a baseline and FastText word embedding and BERT (Bidirectional Encoder Representations from Transformers), on a Support Vector Machine (SVM) model for sentiment classification of comments on the Folkative Instagram account. The data used amounted to 11,914 comments that had gone through a pre-processing process. For data labeling, a semi-automatic lexicon-based method was used that was validated by Indonesian language experts. Model optimization was performed using GridSearchCV and model validation through Stratified 5-Fold Cross Validation. The results showed that TF-IDF with SVM achieved the best performance with an accuracy of 87.61% using the RBF kernel, followed by FastText with SVM at 87.54% using the RBF kernel, and BERT with SVM at 86.38% using the RBF kernel. These results demonstrate that simpler methods can still provide superior results if they match the characteristics of the data. TF-IDF was more effective in classifying comments because comments on the Folkative Instagram account tend to be short and contain explicit sentiment, while FastText and BERT were unable to provide significant performance improvements on this dataset.