Electronic Theses and Dissertation

Analisis sentimen ulasan film secara manual menjadi tidak efisien akibat volume data yang terus meningkat di platform seperti IMDb. Penelitian ini bertujuan untuk menerapkan dan membandingkan performa model Transformer yaitu BERT, RoBERTa, dan DistilBERT dalam melakukan klasifikasi sentimen ulasan film ke dalam tiga kategori yaitu positif, netral, dan negatif, serta mengimplementasikan model terbaik ke dalam aplikasi web berbasis Streamlit. Data dikumpulkan menggunakan teknik web scraping pada situs IMDb dan menghasilkan 11.380 data ulasan film. Pelabelan data dilakukan secara otomatis menggunakan model LLM LLaMA 3.1 8B Instant melalui API Groq dengan tingkat akurasi validasi sebesar 94,33%. Ketiga model dilatih menggunakan teknik fine-tuning dengan konfigurasi hyperparameter yang sama menggunakan framework HuggingFace Transformers dan PyTorch. Hasil evaluasi menunjukkan bahwa model RoBERTa memperoleh performa terbaik dengan nilai Accuracy sebesar 91,87% dan F1-Score sebesar 86,07%, diikuti oleh BERT dengan Accuracy sebesar 89,28% dan F1-Score sebesar 83,05%, serta DistilBERT dengan Accuracy sebesar 86,86% dan F1-Score sebesar 79,91%. Selain itu, model hasil fine-tuning diimplementasikan ke dalam aplikasi Streamlit yang memungkinkan pengguna melakukan prediksi sentimen ulasan film secara interaktif.

Manual sentiment analysis of movie reviews has become inefficient due to the increasing volume of data on platforms such as IMDb. This study aims to apply and compare the performance of Transformer models, namely BERT, RoBERTa, and DistilBERT, in classifying movie review sentiment into three categories: positive, neutral, and negative, as well as implementing the models into a Streamlit-based web application. Data were collected using web scraping techniques on the IMDb website, resulting in 11,380 movie review records. Data labeling was performed automatically using the LLaMA 3.1 8B Instant LLM model via the Groq API, achieving a validation accuracy of 94.33%. All three models were trained using fine-tuning with identical hyperparameter configurations using the HuggingFace Transformers framework and PyTorch. Evaluation results show that RoBERTa achieved the best performance with an Accuracy of 91.87% and an F1-Score of 86.07%, followed by BERT with an Accuracy of 89.28% and an F1-Score of 83.05%, and DistilBERT with an Accuracy of 86.86% and an F1-Score of 79.91%. The fine-tuned models were subsequently deployed into a Streamlit web application that allows users to perform movie review sentiment prediction interactively.