Optimizing Arabic Document Representation via Minimal-Term TF-IDF and Co-Occurrence Integration: Empirical Evaluation Across Multiple Classifiers

Abstract

This study outlines a methodical investigation into optimizing Arabic document representation for classification. It concludes that a balanced selection of minimal terms can achieve high performance while reducing computational load. We conducted experiments on a corpus of 10,000 Arabic news sections from 111,728 documents across five categories and compared feature-extraction methods with classifiers. In Experiment 1, we evaluated the impact of selecting the six most common TF-IDF words per document, varying the number from 1 to 10 to determine the optimal setting. The use of these chosen words, together with similar document vectors, significantly increased model efficacy, achieving a test accuracy of up to 0.99 and an F1-score of 0.99 with a low computational latency across the used classifiers, such as ‘MLPClassifier’, ‘LinearSVC’, ‘RandomForest’, and ‘LogisticRegression’. As further elaborated on in the next step, Experiment 2 added the highest associated co-occurring TF-IDF words (two to ten words per document) to assess semantic enrichment. The findings revealed that, despite a slight increase in feature dimensionality with the addition of associated words, the most potent combination, with the six common words repeatedly replicated, achieved an accuracy of 0.97-0.98. Adding more than 10 connected terms had a minor detrimental effect on classification performance, with a marginal increase in noise. Compared with earlier work, our method demonstrates that semantic integrity, improved classification performance, and reduced processing time can be maintained with minimal-term augmentation, providing a lightweight yet powerful solution for real-time, resource-restricted Arabic NLP applications. These observations explain the value of a carefully chosen, strategically representative, and related subset of terms and offer a reproducible scheme for high-performance Arabic document classification and Arabic NLP applications, which could be extended to other languages.