A Hybrid Interpretation Model Leveraging Improved Grey Wolf Optimization and Graph Attention Networks for Intent Recognition and Dynamic Interaction

Abstract

The study presents an integrated interpreting model that combines intent recognition via Improved Grey Wolf Optimization (IGWO) with independent dynamic interaction via Graph Attention Networks (GAT). It builds an intent-relation graph, uses multi-head attention to capture dynamic links among intents, and leverages IGWO to adapt intent thresholds and attention-head weights. Goal: more reliable multi-intent recognition and better adaptation to changing contexts. On WMT14 (EN–FR) it achieves 94.37% accuracy (DNN & HMM 79.31%, EEMD 75.09%, LLM 64.97%). For 60-s audio it reaches 47.35 dB SNR (EMD 29.74 dB, LLM 26.72 dB); at 160 s it remains highest (47.68 dB). IGWO boosts accuracy via chaotic initialization and Gaussian mutation; a heterogeneous GAT models ternary relations. WMT14 and LibriSpeech are used for translation/ASR, and MixSNIPS/MixATIS for multi-intent understanding. After 500 iterations IGWO hits 92.91% (deep bidirectional pre-trained language model 69.86%, HMM 63.79%); recall exceeds 90% across datasets. Results indicate more accurate, natural translations and stronger handling of technical terminology.