Aspect-Based Sentiment Analysis of Public Responses to Free Nutritious Meal (MBG) Policy Implementation Using Deep Learning

This research presents an Aspect-Based Sentiment Analysis (ABSA) of public responses to the Free Nutritious Meal (MBG) policy using deep learning. Public comments were collected through web scraping and underwent preprocessing, including case folding, tokenizing, normalization, stemming, and stopword removal. The data was then labeled by aspect and sentiment, followed by Exploratory Data Analysis (EDA) to examine patterns and distributions. The classification models were built using BERT for word embedding and deep learning architectures including GRU, LSTM, and BiLSTM. Hyperparameter tuning was conducted using Bayesian Optimization with the Tree-structured Parzen Estimator (TPE) to optimize model performance. Experimental results showed that the GRU model combined with BERT achieved the highest and most consistent accuracy of $\mathbf{9 2. 5 6 \%}$. While BiLSTM performed well in positive sentiment detection, its performance was lower for negative sentiment. The combination of BERT and deep learning models proved effective in enhancing the performance of aspect-based sentiment classification for public policy analysis.