Penelitian ini membahas analisis sentimen ulasan aplikasi menggunakan kombinasi metode Fasttext dan Long Short-Term Memory (LSTM). Dataset yang digunakan terdiri dari 521.007 data ulasan. Proses preprocessing meliputi Case Folding, filtering karakter non-alfabet, normalisasi slang word, koreksi ejaan (spelling correction), tokenisasi, penghapusan stopwords , dan detokenisasi. Untuk feature engineering, selain data splitting (train, Validation, Test), dilakukan pembuatan Embedding Matrix berbasis Fasttext , perhitungan jumlah kata OOV (out-of-vocabulary), serta ekstraksi fitur jumlah kata yang dikoreksi ejaannya.Model Fasttext digunakan untuk menghasilkan representasi kata (Word embedding ), sedangkan model gabungan Fasttext dan LSTM digunakan untuk klasifikasi sentimen. Hasil evaluasi menunjukkan bahwa model Fasttext dan LSTM mencapai akurasi keseluruhan 94,77%. Dari seluruh data, terdapat 27.251 data (5,23%) yang prediksi sentimennya tidak sesuai (mismatch) dengan label yang diharapkan. Analisis Confusion Matrix menunjukkan model sangat baik dalam mengenali sentimen positif, namun masih terdapat kesalahan pada kelas netral dan negatif.Distribusi hasil prediksi sentimen adalah sebagai berikut: negatif 5,7%, netral 13,8%, dan positif 80,5%. Akurasi prediksi untuk masing-masing sentimen adalah negatif 86,2%, netral 77,4%, dan positif 98,4%. Ketidaksesuaian prediksi paling banyak terjadi pada sentimen netral yang sering diprediksi sebagai positif atau negatif. Temuan ini menunjukkan bahwa kombinasi Fasttext dan LSTM efektif untuk analisis sentimen ulasan aplikasi, namun peningkatan pada penanganan sentimen netral masih diperlukan.

This study discusses Sentiment analysis of application Reviews using a combination of Fasttext and Long Short-Term Memory (LSTM) methods. The dataset used consists of 521,007 Review data. The preprocessing process includes Case Folding, filtering non alphabetic characters, normalizing slang Words, spelling correction, tokenization, Stopword Removal, and detokenization. For feature engineering, in addition to data splitting (train, Validation, Test), a Fasttext -based Embedding Matrix was created, the number of out-of vocabulary (OOV) Words was calculated, and the number of Words with corrected spelling was extracted as a feature. The Fasttext model was used to geneRate Word embedding s, while the combined Fasttext +LSTM model was used for Sentiment classification . The Evaluation Results show that the Fasttext +LSTM model achieved an overall accuracy of 94.77%. Of the entire data set, there were 27,251 data points (5.23%) where the Sentiment Prediction did not match the expected label. The Confusion Matrix analysis shows that the model is very good at recognizing Positive Sentiment , but there are still errors in the neutral and Negative classes. The distribution of Sentiment Prediction Results is as follows: Negative 5.7%, neutral 13.8%, and Positive 80.5%. The Prediction accuracy for each Sentiment is Negative 86.2%, neutral 77.4%, and Positive 98.4%. The most discrepancies in Prediction s occur in neutral Sentiment s, which are often predicted as Positive or Negative . These findings indicate that the combination of Fasttext and LSTM is effective for analyzing app Review Sentiment , but improvements in handling neutral Sentiment s are still needed.