Long-Context Contrastive Adaptation for Automated Essay Scoring: Mitigating Information Blindness with BGE-M3

A long-context Automated Essay Scoring system built on BGE-M3 — score-aware contrastive fine-tuning and imbalance-aware oversampling on ASAP 2.0, where expanding the context window from 128 to 1024 tokens cut MAE from 0.78 to 0.47 and reached a QWK of 0.77.