Evaluating the Generalizability of Semantic Term Synthesis with Chi-Square Guided Feature Selection Technique across Multiple Classifiers
Feature selection is essential in text classification because it reduces dimensionality and improves the quality of document representation. However, many traditional feature selection methods rely mainly on statistical term distributions and often overlook semantic relationships among terms. This study evaluates the effectiveness of the Semantic Term Synthesis with Chi-Square (T-Sync) framework, which enhances feature representation by synthesizing semantically related terms before feature selection. Experiments were conducted on the 20 Newsgroups and Reuters-21578 benchmark datasets using four machine learning classifiers: Naïve Bayes, Support Vector Machine, K-Nearest Neighbors, and Random Forest. Their performance was evaluated using Accuracy, F1-score, and execution time across different feature dimensions. On the 20 Newsgroups dataset, T-Sync achieved average accuracies of 0.7941 (NB), 0.7917 (SVM), 0.5799 (KNN), and 0.7809 (RF). On the Reuters dataset, the corresponding average accuracies were 0.9261 (NB), 0.9391 (SVM), 0.8150 (KNN), and 0.9318 (RF). In addition to improving classification performance over Var-CV-CHI in most experimental settings, T-Sync consistently required less execution time, demonstrating its computational efficiency. These findings shows that incorporating semantic term synthesis into feature selection can improve both the effectiveness and efficiency of text classification systems.
Download Paper