PENDEKATAN SENTIMEN BERBASIS ASPEK PADA ULASAN SIRKUIT MANDALIKA MENGGUNAKAN CNN DAN REPRESENTASI FASTTEXT

Aspect-Based Sentiment Approach to Mandalika Circuit Reviews Using CNN and FastText Representation

Abstract

Reviews are texts that contain an assessment or comment on something and can be used to provide more in-depth information. This research aims to analyze community reviews of the Mandalika Circuit using the aspect-based sentiment analysis technique CNN method. The CNN model is trained using two types of word embedding, namely Keras and FastText, and supported by the Multilabel Stratified K-Fold Cross Validation method to ensure an even distribution of data on each label and produce a stable accuracy evaluation. The results show that CNN with FastText word embedding has a higher average accuracy than CNN with Keras word embedding for both aspect and sentiment classification tasks. However, the model had difficulty in classifying the positive class in the sentiment label, which was due to the smaller amount of review data with positive sentiment than neutral and negative. Therefore, for future research, it is recommended to use data augmentation techniques on the imbalanced classes to improve the accuracy of the model.