Long Text Matching Model with Multi-granularity Feature-Hierarchical Filtering

For the problem of significant noise interference and easy disappearance of key information in long text matching tasks, a multi-granularity feature hierarchical filtering model is proposed. An innovative design of multi-granularity feature extraction and dynamic length reduction strategy is adopted to improve the feature representation and the effect of redundant word deletion. A joint text sentence graph is constructed using the TextRank algorithm to extract key sentences for compressed input; in the word filtering layer, parallel implementation of word-level global attention, phrase-level sliding window attention, and sentence-level gated attention multi-granularity feature extraction is achieved, and through learnable weights, the three-level features are dynamically fused to optimize the feature representation; finally, driven by four-dimensional indicators including L2 norm, semantic difference, entropy and gradient, the PageRank graph is iteratively compressed to achieve adaptive redundant word filtering. Experiments show that this model leads by 0.25 and 2.24 percentage points in accuracy on the Chinese News Same Event and Chinese News Same Story datasets, effectively improving the accuracy of long text matching tasks.