Background Vocabulary serves as the foundation for text comprehension. Therefore, texts at different proficiency levels should incorporate appropriately adapted vocabulary to support second language learners in developing their reading skills. Methods To objectively quantify the grading process of L2 Chinese (Chinese for a Second Language) reading texts, we designed a text classification experiment grounded in lexical fitness evaluation. This study employs a comparative framework, pitting a classical machine learning classification model against a neural network-based classification model incorporating a multi-head attention mechanism. The experimental results demonstrate that the rule-based lexical feature extraction method achieves optimal performance under the random forest (RF) classification model, attaining an F1-score of 0.879. To further optimize and streamline the classification process, we identified 22 optimal feature sets based on mutual information (MI) rankings of lexical features, incorporating novel features introduced in this study. Results A comparative analysis with prior work underscores the superiority of our lexical feature construction framework. Crucially, the assessment of lexical form, lexical meaning, and lexical syntax, as well as their interactions across different textual levels, exhibits significant variability. Optimal evaluation can only be achieved through a holistic integration of all three dimensions.