HYBRID TEXT CLASSIFICATION BASED ON TF-IDF AND ADAPTIVE ALSHE ENSEMBLE

Abstract

This article investigates the task of multi-class classification of technical texts. The experiments utilized the LocalDocs-10 corpus, compiled from software package descriptions and partitioned into 10 thematic classes. Texts were represented via word-level and character-level TF-IDF  -grams, alongside compact SVD-derived features. A comparative evaluation was conducted between classical machine learning algorithms and several hybrid approaches. Special emphasis was placed on the adaptive ALSHE-Gated model, which integrates Complement Naive Bayes and LinearSVC through a confidence-driven switching mechanism. The Passive-Aggressive Classifier achieved the highest performance among baseline models, attaining an Accuracy of   and a Macro F1-score of  . These findings affirm that lightweight TF-IDF models constitute a viable alternative to computationally intensive neural networks for small- to medium-sized corpora.