Electronic Theses and Dissertation

Mobil otonom merupakan salah satu inovasi teknologi yang berpotensi mentransformasi sistem transportasi global, namun tingkat penerimaan dan keberhasilannya sangat dipengaruhi oleh persepsi publik. Analisis sentimen terhadap opini masyarakat di media sosial menjadi pendekatan penting untuk memahami tingkat penerimaan tersebut. Penelitian ini bertujuan membangun dan mengevaluasi model klasifikasi sentimen berbasis pre-trained Transformer terhadap cuitan mengenai mobil otonom pada platform X. Dataset penelitian terdiri atas 99.349 cuitan berbahasa Inggris yang diperoleh melalui web scraping dan diberi label sentimen positif, negatif, dan netral menggunakan metode VADER. Selanjutnya, tiga model pre-trained Transformer, yaitu BERT, RoBERTa, dan DistilBERT, di-fine-tune untuk tugas klasifikasi sentimen menggunakan skema pembagian data holdout (80:10:10). Untuk meningkatkan kemampuan model dalam mengenali seluruh kelas sentimen secara proporsional, proses pelatihan menerapkan class-weighted cross-entropy loss. Evaluasi dilakukan menggunakan metrik Accuracy, Precision, Recall, dan Macro F1-score. Hasil penelitian menunjukkan bahwa model BERT memberikan performa terbaik dengan Macro Precision sebesar 0,9539, Macro Recall sebesar 0,9623, dan Macro F1-score sebesar 0,9578, diikuti oleh DistilBERT dengan Macro F1-score sebesar 0,9565 dan RoBERTa sebesar 0,9529. Meskipun selisih performa di antara ketiga model tergolong kecil, BERT dipilih sebagai model dengan performa terbaik untuk konteks penelitian ini, didukung oleh tingkat overfitting yang paling rendah dan jumlah kesalahan klasifikasi yang paling sedikit dibandingkan kedua model lainnya. Penelitian ini memberikan dasar empiris bagi pemanfaatan model Transformer dalam pemantauan persepsi publik terhadap teknologi kendaraan otonom serta dapat menjadi referensi bagi penelitian lanjutan yang memanfaatkan teknik pelabelan yang lebih akurat, seperti anotasi manusia atau semi-supervised learning, maupun model Transformer generasi terbaru.

Autonomous vehicles represent one of the technological innovations with the potential to transform global transportation systems; however, their level of acceptance and success is strongly influenced by public perception. Sentiment analysis of public opinion on social media has become an important approach for understanding this level of acceptance. This study aims to build and evaluate pre-trained transformer-based sentiment classification models applied to tweets about autonomous vehicles on the X platform. The research dataset consists of 99,349 English-language tweets obtained through web scraping, labeled with positive, negative, and neutral sentiment using the VADER method. Subsequently, three pre-trained transformer models BERT, RoBERTa, and DistilBERT were fine-tuned for the sentiment classification task using a holdout (80:10:10) data-splitting scheme. To improve the models' ability to recognize all sentiment classes proportionally, the training process applied class-weighted cross-entropy loss. Evaluation was conducted using Accuracy, Precision, Recall, and Macro F1-score metrics. The results show that the BERT model achieved the best performance, with a Macro Precision of 0.9539, Macro Recall of 0.9623, and Macro F1-score of 0.9578, followed by DistilBERT with a Macro F1-score of 0.9565 and RoBERTa with 0.9529. Although the performance difference among the three models was relatively small, BERT was selected as the model with the best performance for the context of this study, supported by the lowest overfitting level and the fewest classification errors compared to the other two models. This study provides an empirical foundation for the use of transformer models in monitoring public perception of autonomous vehicle technology and can serve as a reference for future research employing more accurate labeling techniques, such as human annotation or semi-supervised learning, as well as newer generations of transformer models.