Addressing imbalanced data problems in corpus linguistics: A comparative evaluation

抄録

Imbalanced data distributions are prevalent in corpus linguistics, where sub-corpora frequently vary in size due to collection constraints or uneven population distributions. This study investigated the impact of class imbalance on text classification and conducted a comparative evaluation of undersampling, oversampling, SMOTE, and a cost-sensitive weighted generalized linear model using the Spam dataset. The experimental results demonstrate that as class imbalance intensifies, the overall accuracy remains misleadingly high, whereas precision and recall decline sharply, underscoring that reliance on accuracy alone yields distorted conclusions. Among the evaluated methods, SMOTE achieved the highest F1 score, and the weighted model provided a robust balance between predictive performance and data integrity. These findings suggest that corpus researchers should evaluate imbalanced classification tasks using precision, recall, and F1 scores and select a compensatory methodology that aligns with their specific analytical objectives.