SENTIMENT ANALYSIS OF POLITICAL PARTY NEWS ON THE ONLINE NEWS PORTAL DETIK.COM USING LSTM AND CNN

• Main Author: Tuah, Epata
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73587
• Institution: Institut Teknologi Bandung

The online news portal Detik.com contains a number of articles discussing specific political parties. The coverage of these political parties, from the 2019 general elections to the upcoming general elections, can evoke certain sentiments towards the affiliated political parties. Therefore, this research aims to analyze the sentiments arising from the news texts about political parties on Detik.com. The objective of this study is to develop a sentiment analysis model that can be used to analyze political party news data on Detik.com and identify the sentiments expressed in that data. This research commenced by collecting data using the web scraping method to obtain a corpus of textual data from political news articles on Detik.com. The employed method gathered a total of 19,463 rows of data. The data labeling process was conducted in four stages by multiple annotators to mitigate bias and enhance the accuracy of the assigned labels. Subsequently, sentiment analysis on the political news data from Detik.com was conducted utilizing a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models, along with Word2Vec and FastText word embedding vectors. These models were then evaluated to determine their quantitative and qualitative performance. Based on the conducted experiments, the sentiment analysis using the Word2Vec-CNN- LSTM method achieved the highest average accuracy of 69.39%. The same model also attained the highest maximum accuracy, reaching 70.46%. The results of the sentiment analysis indicated that Detik.com tends to exhibit a neutral sentiment towards almost all political parties, except for PSI and Partai Garuda, which had a predominantly negative sentiment. The evaluation of the model revealed that the combination of LSTM and CNN models performed exceptionally well in news articles with a large number of relevant sentiment words, while it exhibited lower accuracy on articles with fewer relevant sentiment words. FastText word embedding vectors demonstrated better performance in predicting non-neutral labeled text data, while Word2Vec exhibited superior performance in predicting neutral text data.