SAER : Comparison of Rule Prediction Algorithms on Constructing a Corpus for Taxation Related Tweet Aspect-Based Sentiment Analysis

Abstract

Twitter is a popular social media in Indonesia, and sentiment analysis on Twitter has an important role in measuring public trust, especially in taxation issues. Aspect extraction is an important task in sentiment analysis. In this research, we propose SAER, a Syntactic Aspect-opinion Extraction and Rule prediction, that used language rule-based approach using syntactic features for aspect and opinion extraction, and we compare several algorithm for rule prediction such as Random Forest Regression, Decision Tree Regression, K-Nearest Neighbor Regression (KNN), Linear Regression, Support Vector Regression (SVR), and Extreme Gradient Boosting Regression (XGBoost) that can generate rules with a tree-based approach. By employing syntactic features and rule prediction, it has been able to explore important features in a sentence. In rule prediction, comparison results show that Support Vector Regression (SVR) was identified as the most effective model for aspects rule prediction, providing the best results with a Mean Squared Error (MSE) of 0.022, Root Mean Squared Error (RMSE) of 0.150, and Mean Absolute Error (MAE) of 0.123. While XGBoost was identified as the most effective model for opinions rule prediction, with MSE of 0.013, RMSE of 0.117, and MAE of 0.075. Since we used syntactic feature-based approaches and rule prediction in this work, it is expected to be implemented for other cases, with other domain datasets.

Keywords

• taxation
• aspect-opinion extraction
• rule prediction
• saer
• syntactic feature