Pendekatan Aspect Based Sentiment Analysis Berbasis BERT untuk Ulasan Game dalam Perspektif Business Intelligence

Bert-Based Aspect-Based Sentiment Analysis Approach for Game Reviews from a Business Intelligence Perspective

Video game awalnya dikembangkan di lingkungan militer sebagai sarana hiburan dan kemudian berkembang menjadi industri global dengan miliaran pemain. Salah satu game populer bergenre RPG adalah Elden Ring, yang menghasilkan ulasan dalam jumlah besar sehingga menyulitkan pengembang dalam mengidentifikasi pujian dan keluhan secara efektif. Penelitian ini bertujuan untuk menganalisis sentimen berbasis aspek pada ulasan pemain, meliputi gameplay, graphics, performance, dan story, serta membandingkan performa algoritma LR, BERT, dan RF, serta memvisualisasikan hasilnya melalui dashboard BI. Hasil penelitian menunjukkan bahwa seluruh aspek didominasi oleh sentimen positif. BERT mencapai akurasi tertinggi sebesar 90%, diikuti LR (89%) dan RF (83%), meskipun LR menunjukkan kestabilan confidence yang lebih baik. Visualisasi juga menunjukkan bahwa sentimen positif cenderung konsisten, sedangkan sentimen negatif lebih kompleks dan beragam. Untuk penelitian selanjutnya, disarankan penerapan teknik kalibrasi model, termasuk kalibrasi output softmax, integrasi uncertainty dalam proses pelatihan, serta evaluasi menggunakan Expected Calibration Error guna meningkatkan kesesuaian antara confidence dan performa model.

Video games were initially developed in military environments as a form of entertainment and have since evolved into a global industry with billions of players. One of the popular RPG games is Elden Ring, which generates a large volume of user reviews, making it difficult for developers to effectively identify praise and complaints. This study aims to perform ABSA on player reviews, covering gameplay, graphics, performance, and story, while also comparing the performance of LR, BERT, and RF, and visualizing the results through a BI dashboard. The results indicate that all aspects are dominated by positive sentiment. BERT achieved the highest accuracy at 90%, followed by LR (89%) and RF (83%), although LR demonstrated more stable confidence values. The visualization further reveals that positive sentiments are more consistent, whereas negative sentiments tend to be more complex and diverse. For future research, it is recommended to apply model calibration techniques, including softmax output calibration, incorporating uncertainty into the training process, and evaluating calibration using Expected Calibration Error to improve alignment between confidence and actual model performance.