Perbandingan Kinerja Pre-Trained Word Embedding Terhadap Performa Klasifikasi Sentimen Ulasan Produk Tokopedia Dengan Long Short-Term Memory(LSTM)

The product review dataset is a rapidly growing and interesting source of data to explore. The increase in the number of internet users and customer shopping habits through online stores has a significant impact on the growth of product review data, especially for online stores in Indonesia, such as Tokopedia. The sample data used amounted to 1079. This research aims to evaluate the performance of three types of pre-trained word embeddings, namely FastText, GloVe, and Word2Vec, in the Long Short-Term Memory (LSTM) model for sentiment classification of product reviews on Tokopedia. An automated sentiment classification system is needed to process many product reviews, making it easier for sellers to know what consumers think of their products. This research contributes by evaluating the impact of various pre-trained word embeddings on the performance of LSTM models in sentiment classification tasks. In addition, this research also aims to measure the effectiveness of LSTM models combined with multiple pre-trained word embeddings. By implementing a deep learning architecture, computers can learn and recognize contextual data stored in review sentences. The research was conducted in three stages: model selection, layer setup, and hyperparameter optimization, to feature in-depth testing of the deep learning architecture used and the appropriate combination of layers and parameters to obtain high sentiment classification performance. The experimental results show that FastText with LSTM provides the best performance with 85.08% accuracy, Word2Vec with 84.62% accuracy, and GloVe with 83.04% accuracy. The main contribution of this research is to present an in-depth test of the product review dataset and provide a deep learning architecture along with a combination of layers and parameters that has the best performance in recognizing sentiment on the product review dataset. This architecture achieves higher performance than the BERT method with CNN and BiLSTM layers.