<![CDATA[Sentiment analysis is one of the applications of Natural Language Processing (NLP) for identifying public opinion on social media. Although previous studies on Tapera sentiment analysis have applied conventional machine learning and transformer-based approaches, limited attention has been given to integrating automatic lexicon-based sentiment labeling with semantic word representation and deep learning classification within a unified framework. This study proposes an approach that integrates the Indonesian Sentiment Lexicon (InSet), Word2Vec, and Convolutional Neural Network (CNN) to classify public sentiment toward the Tabungan Perumahan Rakyat (Tapera) policy on social media X. A total of 1,291 tweets were processed through the preprocessing stage and automatically labeled using InSet. Word2Vec was employed to construct semantic word representations that were used as the initial weights of the CNN embedding layer. The proposed model was evaluated using four hold-out data split scenarios and compared with BiLSTM and Random Forest. The experimental results showed that the 80:20 configuration achieved the best performance with an accuracy of 74.52%, outperforming BiLSTM (71.43%) and Random Forest (69.88%). The novelty of this study lies in integrating automatic sentiment labeling using the Indonesian Sentiment Lexicon (InSet), semantic representation through Word2Vec, and CNN-based classification within a unified framework, together with an empirical evaluation of different train-test split ratios for Tapera policy sentiment analysis. These findings demonstrate that the proposed approach provides competitive sentiment classification performance while reducing dependence on manual sentiment annotation.]]>
