ANALISIS SENTIMEN BERBASIS NATURAL LANGUAGE PROCESSING UNTUK ULASAN PENGGUNA TERKAIT PERFORMA APLIKASI PADA PLATFORM E-COMMERCE

Abstract

Pertumbuhan pesat internet di Indonesia mendorong pergeseran belanja masyarakat dari luring ke platform e-commerce. Di tengah persaingan ketat, performa aplikasi menjadi penentu kepuasan dan loyalitas pengguna. Shopee kerap menerima keluhan gangguan teknis yang berpotensi menggerus kepercayaan pengguna jika tidak ditindaklanjuti. Ulasan pengguna di Google Play Store menjadi sumber data yang kaya untuk memahami persoalan tersebut secara langsung dari sudut pandang pengguna. Penelitian ini bertujuan membangun model analisis sentimen berbasis Natural Language Processing dengan menggabungkan IndoBERT dan Bi-LSTM, mengidentifikasi topik dominan beserta distribusinya pada setiap kelas sentimen menggunakan BERTopic, serta menyusun rekomendasi berbasis data bagi pengembang untuk meningkatkan performa aplikasi. Sebanyak 9.395 ulasan berbahasa Indonesia dikumpulkan melalui teknik scraping dengan periode Februari hingga Mei 2026. Setelah melalui tahap praproses (pembersihan teks, case folding, normalisasi kata tidak baku, penghapusan stopword, dan penyaringan), tersisa 6.309 ulasan. Pelabelan sentimen dilakukan dengan pendekatan hybrid: separuh data dilabeli manual sebagai acuan dan separuh lainnya dilabeli otomatis menggunakan model pralatih. Distribusi kelas yaitu 49,2% positif, 40,9% negatif, dan 9,8% netral. Ketidakseimbangan ini ditangani melalui pembobotan kelas (class weight) selama pelatihan model. Model gabungan IndoBERT–Bi-LSTM mencapai akurasi keseluruhan sebesar 85,9% dengan F1-score makro 0,75 pada data uji. Hasil penelitian menunjukkan bahwa performa model sangat baik dalam mengenali sentimen positif (F1-score 0,91) dan negatif (0,88), namun masih lemah pada kelas netral (0,45) akibat jumlah data yang jauh lebih sedikit. Pemodelan topik dengan BERTopic menghasilkan lima topik pada setiap kelas sentimen. Kelima topik pada kelas negatif mencakup 58,8% ulasan negatif, yaitu keterlambatan pengiriman (19,6%), keluhan umum terkait barang dan layanan (17,1%), iklan yang dianggap mengganggu (13,4%), kendala akses akun (5,4%) dan ketidakkonsistenan kinerja aplikasi (3,3%). Pada kelas positif, kelima topik mencakup 74,7% ulasan positif, dengan kepuasan berbelanja (44,4%), kemanfaatan aplikasi (15,4%) dan kemudahan penggunaan (8,1%) sebagai kekuatan yang perlu dipertahankan. Penelitian ini menghasilkan lima rekomendasi bagi stakeholder dan pengembang aplikasi, yaitu memperbaiki akurasi estimasi pengiriman dan keandalan mitra kurir, memperkuat penanganan keluhan terkait barang dan layanan, mengevaluasi ulang strategi penempatan iklan, menyederhanakan mekanisme pemulihan akses akun, serta memperkuat pemantauan konsistensi kinerja aplikasi. Penelitian selanjutnya disarankan menambah data ulasan netral serta menguji arsitektur lain untuk mengevaluasi kemampuan generalisasi model pada data dari periode dan platform yang berbeda.

The rapid growth of internet use in Indonesia has shifted consumer shopping behavior from offline retail to e-commerce platforms. In a highly competitive market, application performance determine user satisfaction and loyalty. Shopee often receives complaints about technical problems that can reduce user trust if they are not addressed. User reviews on the Google Play Store provide a rich data source for understanding these problems directly from the users' point of view. This study aims to develop a sentiment analysis model based on Natural Language Processing by combining IndoBERT and Bi-LSTM, to identify the dominant topics and their distribution in each sentiment class using BERTopic, and to formulate data-driven recommendations for developers to improve application performance. A total of 9,395 Indonesian-language reviews were collected through scraping over the period of February to May 2026. After preprocessing (text cleaning, case folding, normalization of non-standard words, stopword removal, and filtering), 6,309 reviews remained. Sentiment labeling was performed using a hybrid approach, in which half of the data was labeled manually as a reference and the other half was labeled automatically using a validated pre-trained model. This produced an imbalanced class distribution (49.2% positive, 40.9% negative, and 9.8% neutral). The imbalance was handled by applying class weighting during model training. The combined IndoBERT–Bi-LSTM model achieved an overall accuracy of 85.9% with a macro F1-score of 0.75 on the test data. The results show that the model performs well in recognizing positive (F1- score 0.91) and negative sentiment (0.88), but remains weak on the neutral class (0.45) because of the much smaller amount of data in that class. Topic modeling with BERTopic produced five topics in each sentiment class with an uneven distribution. The five topics in the negative class cover 58.8% of negative reviews, namely delivery delays (19.6%), general complaints about goods and services (17.1%), intrusive advertisements (13.4%), account access problems (5.4%), and inconsistent application performance (3.3%). In the positive class, the five topics cover 74.7% of positive reviews, with shopping satisfaction (44.4%), application usefulness (15.4%), and ease of use (8.1%) as strengths that should be maintained. This study produces five recommendations for stakeholders and application developers, namely improving the accuracy of delivery time estimates and the reliability of courier partners, strengthening the handling of complaints related to goods and services, re-evaluating advertisement placement strategies, simplifying account recovery mechanisms, and strengthening the monitoring of application performance consistency. Future research is suggested to add more neutral review data and to test alternative architectures in order to evaluate the generalization capability of the model on data from different periods and platforms.