Large language models have become the dominant paradigm in Natural Language Processing, yet their explainability remains fundamentally limited, making automatic error detection challenging. This study conducts a systematic comparison of explainability signatures between discrete graph-based and continuous token-based architectures for error pattern recognition in text classification. Using text-to-graph conversion, LLM-as-teacher distillation, and unified Shapley-based attribution, we identify four complementary architectural patterns that enable reliable discrimination between correct and incorrect predictions: AUC-based confidence trajectories, feature concentration profiles, margin-preservation patterns, and fidelity-asymmetry signatures.We demonstrate that hierarchical graph-based architectures produce substantially sharper discriminative patterns (99.7--100.0\% error detection accuracy) compared to token-based architectures (88.1--89.6\% accuracy) under fair comparison. This performance gap persists across datasets, graph topologies, and prediction correctness conditions. Integrated logistic regression analysis reveals that all architectures rely on necessity-based features, but hierarchical discrete structures produce significantly stronger necessity signals through binary node participation, while continuous token embeddings produce attenuated signals through distributed compensatory pathways.For high-stakes applications requiring transparent automated error detection, hierarchical graph-based explainability patterns offer clear advantages over token-based approaches. Our four-dimensional evaluation framework provides a reproducible methodology for architectural signature analysis in explainability research.
原文链接
分享
引用
