Long Text Classification Model Based on Transformer Sliding Window and Threshold Optimization

Abstract

In long-text classification tasks, the main issue we face is that traditional text classification methods lack the capability to analyze complex contextual information and implicit semantics, resulting in poor classification performance. To address the shortcomings of existing long-text classification models—such as poor accuracy, low efficiency, inadequate dynamic adjustment, and poor threshold adaptability—we propose a Transformer-based sliding window threshold optimization long-text classification model. We investigate an automatic classification and dynamic partitioning method for long-text semantic analysis. The approach utilizes context-aware fusion of semantic information from long texts to improve the accuracy of automated long-text classification. Finally, we employ an error feedback mechanism to dynamically adjust the classification threshold, achieving optimized threshold settings for long-text content classification. The experiment shows that this model surpasses the baseline model in accuracy and F1 performance on the dataset and exhibits good convergence speed, validating its effectiveness.