Table 1 Comparison of different AES methods.
Method category | Feature types | Semantic alignment | Discourse modeling | Cross-prompt adaptability | Interpretability | Improvements of HFC-AES |
|---|---|---|---|---|---|---|
Traditional feature engineering | Lexical, syntactic, length | Weak | Weak | Weak | Strong | Limited in modeling complex semantics |
Single-prompt DL-based models | Word embeddings, CNN, RNN | Moderate | Basic | Weak | Weak | Lacks cross-prompt robustness |
Multilingual pre-trained models (e.g., mBERT) | Cross-lingual semantic representations | Strong | Weak | Moderate | Weak | Lack mechanisms for essay structure modeling |
HFC-AES | Hybrid shallow + deep features | Strong | Strong | Strong | Moderate | Introduces hierarchical modeling and cross-attention |
