Feature Selection for Deep Text Classification With Limited Labeled Data: A Cross‐Architecture Empirical Study

ABSTRACT Deep learning‐based text classification has become an important research topic in natural language processing, but deep learning models may suffer from unstable performance and overfitting when only limited labeled data are available. Feature selection has shown effectiveness in traditional machine learning‐based text classification by reducing input dimensionality and removing irrelevant features, but its role in deep learning‐based text classification under limited‐data conditions remains insufficiently understood. This study investigates whether traditional feature selection methods can improve the performance and computational efficiency of deep text classifiers when labeled training data are limited. Three representative feature selection methods—document frequency (DF), information gain (IG), and distinguishing feature selector (DFS)—are integrated with 11 classifiers covering word‐averaging, convolutional, recurrent, BERT‐based, and graph‐based architectures. Comparative experiments are conducted on 16 publicly available datasets using five‐fold cross‐validation. Within each fold, feature selection is performed exclusively on the corresponding training partition to prevent information from the held‐out test data from influencing the selected vocabulary. The results reveal substantial architecture‐dependent effects. For the classic neural classifiers, IG and DFS significantly improve mean accuracy by 2.44 and 3.66 percentage points, respectively. DFS also improves the inductive GCN by an average of 2.91 percentage points and produces gains on 13 of the 16 datasets, although this effect is not significant after multiple‐comparison correction. In contrast, all three word‐level feature selection methods reduce the average performance of the BERT‐based classifiers. The sensitivity analysis shows that the benefits of DFS for the classic and graph‐based models persist across vocabulary sizes from 500 to 2000 words. Feature scoring requires less than 0.07 s per fold on average, while vocabulary reduction substantially decreases the computational cost of the GCN. These findings indicate that conventional feature selection can serve as an effective and computationally inexpensive complementary strategy for classic and graph‐based classifiers under limited labeled data, but its benefits depend on the selector, dataset, and model architecture and do not extend to BERT under the evaluated setting.