Electronic Theses and Dissertation

Penelitian ini mengumpulkan data lima besar partai politik di Indonesia dalam rentang Januari 2020 hingga Februari 2024, menghasilkan 370.053 baris data. Pelabelan emosi dilakukan menggunakan NRC Emotion Lexicon (EmoLex) pada teks untuk mengidentifikasi delapan kategori emosi. Penelitian ini bertujuan untuk melatih, mengevaluasi, dan membandingkan model Recurrent Neural Network (RNN) dengan Long Short Term Memory (LSTM) untuk menganalisis emosi pengguna platform X terhadap lima partai politik tersebut. Hasil menunjukkan bahwa LSTM lebih unggul daripada RNN dalam f1 score, loss, akurasi, presisi, dan recall. Pada data tidak seimbang, LSTM mencapai f1 score sebesar 0.85862, sedangkan RNN hanya 0.67795. Penerapan teknik Synthetic Minority Oversampling Technique (SMOTE) membantu menyeimbangkan data, tetapi tidak berdampak signifikan pada performa model. F1 score untuk model LSTM turun menjadi 0.79626, sementara RNN turun menjadi 0.58277. Visualisasi data mengindikasikan bahwa PKB, NasDem, dan Golkar didominasi emosi trust dan anticipation, sementara Gerindra menunjukkan dominasi emosi trust dengan peningkatan emosi fear yang sedikit. Di sisi lain, PDIP lebih banyak menerima emosi anger meskipun tetap memiliki tingkat trust yang tinggi. Berdasarkan rata-rata engagement, PDIP menempati posisi tertinggi yang menunjukkan bahwa unggahan terkait partai ini menarik perhatian publik. Model terbaik kemudian diimplementasikan dalam website berbasis Streamlit untuk memudahkan pemanfaatan hasil analisis.

Kata kunci : Partai Politik, RNN, LSTM, EmoLex, Analisis Emosi

This research collected data on the top five political parties in Indonesia between January 2020 and February 2024, resulting in 370.053 rows of data. Emotion labelling was performed using the NRC Emotion Lexicon (EmoLex) on the text to identify eight categories of emotions. This research aims to train, evaluate, and compare Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM) to analyse platform X users’ emotions towards the five political parties. Results show that LSTM is better to RNN in f1 score, loss, accuracy, precision, and recall. In case for unbalanced data, LSTM achieves an f1 score of 0.85862, while RNN is only 0.67795. The application of the Synthetic Minority Oversampling Technique (SMOTE) helped to balance the data, but did not have a significant impact on model performance. F1 score for the LSTM model dropped to 0.79626, while the RNN dropped to 0.58277. The data visualisation indicates that PKB, NasDem, and Golkar are dominated by the emotions of trust and anticipation, while Gerindra shows the dominance of trust emotion with a slight increase in fear emotion. On the other hand, PDIP received more anger emotions although it still had a high level of trust. Based on the average engagement, PDIP occupies the highest position which shows that posts related to this party attract public attention. The best model then implemented in a Streamlit-based website to facilitate the utilisation of the analysis results. Kata kunci : Political Parties, RNN, LSTM, EmoLex, Emotion Analysis