Perkembangan bank digital di Indonesia semakin pesat. Salah satunya adalah Bank Neo Commerce dengan aplikasi Neobank sebagai platform utama untuk transaksi digital. Hingga tahun 2024, aplikasi ini telah diunduh lebih dari 25 juta kali, tetapi hanya memiliki rating 3,5/5, yang lebih rendah dibandingkan dengan aplikasi bank digital lainnya. Dengan jumlah unduhan yang tinggi tetapi rating yang rendah, analisis sentimen level topik pada ulasan pengguna sangat penting untuk memahami kepuasan dan persepsi pengguna. Analisis sentimen pada umumnya dilakukan menggunakan Long Short-Term Memory (LSTM). Namun, LSTM memiliki tiga limitasi utama, yaitu ketidakmampuan untuk merevisi keputusan penyimpanan, kapasitas penyimpanan terbatas, dan kurangnya paralelisasi karena memory mixing. Untuk mengatasi hal tersebut, xLSTM diusulkan dengan memperkenalkan dua modifikasi utama pada LSTM, yaitu exponential gating dan struktur memori baru. Penelitian ini menerapkan analisis sentimen menggunakan arsitektur Extended Long Short-Term Memory (xLSTM) dan pendeteksian topik menggunakan BERTopic. Analisis dilakukan pada data ulasan aplikasi Neobank serta tiga dataset e-commerce, yaitu Shopee, Tokopedia, dan Lazada. Empat konfigurasi xLSTM (1:0, 0:1, 1:1, dan 7:1) dibandingkan dengan model LSTM dan Attention menggunakan metrik evaluasi accuracy, precision, recall, dan F1 score. Hasil menunjukkan bahwa xLSTM secara konsisten mengungguli model pembanding, dengan konfigurasi xLSTM[7:1] memberikan kinerja terbaik dengan rata-rata evaluasi accuracy 83,34% ± 0,80%, precision 83,56% ± 0,85%, recall 82,54% ± 1,00%, dan F1 score 82,61% ± 0,86%.. Analisis sentimen terhadap 100.000 ulasan Neobank menunjukkan bahwa 55,5% ulasan bersentimen positif. Proses pendeteksian topik menggunakan BERTopic dilakukan melalui embedding SBERT, reduksi dimensi UMAP, clustering HDBSCAN, representasi topik dengan c-TF-IDF, dan interpretasi label topik menggunakan model LLM Gemma 2. Hasil akhir menghasilkan CV coherence score sebesar 0,646 dan 15 topik utama, dengan tujuh topik didominasi sentimen positif dan enam topik negatif, terutama terkait login, verifikasi wajah, dan program referral.

---

The development of digital banking in Indonesia has grown rapidly. One prominent example is Bank Neo Commerce, which provides Neobank as its main platform for digital transactions. By 2024, the application has been downloaded more than 25 million times, yet it holds a relatively low rating of 3.5/5, lower than other digital banking applications. With high download numbers but a low rating, topic-level sentiment analysis on user reviews becomes crucial to understand user satisfaction and perception Sentiment analysis is commonly performed using LSTM. However, LSTM has three main limitations: the inability to revise storage decisions, limited memory capacity, and lack of parallelism due to memory mixing. To overcome these issues, xLSTM has been proposed by introducing two major modifications to LSTM, namely exponential gating and a new memory structure. This study applies sentiment analysis using the Extended Long Short-Term Memory (xLSTM) architecture and topic modeling using BERTopic. The analysis was conducted on reviews from the Neobank application and three e-commerce platforms: Shopee, Tokopedia, and Lazada. Four xLSTM configurations (1:0, 0:1, 1:1, and 7:1) were compared against LSTM and Attention-based models using accuracy, precision, recall, and F1 score as evaluation metrics. The results show that xLSTM consistently outperformed the baseline models, with the xLSTM[7:1] configuration achieving the best performance, averaging 83,34% ± 0,80% accuracy, 83,56% ± 0,85% precision, 82,54% ± 1,00% recall, and 82,61% ± 0,86% F1 score across all datasets. Sentiment analysis on 100,000 Neobank user reviews revealed that 55.5% were classified as positive sentiment. Topic modeling using BERTopic was carried out through SBERT embedding, dimensionality reduction with UMAP, clustering using HDBSCAN, topic representation with c-TF-IDF, and topic label interpretation using the LLM model Gemma 2. The final result yielded a CV coherence score of 0.646 and 15 main topics, with eight topics dominated by positive sentiment and six by negative sentiment, mainly related to login issues, facial verification, and referral programs.