zzy99/epidemic-sentence-pair

Implements semantic similarity matching for pandemic-related medical queries using ensemble methods combining BERT-wwm-ext, ERNIE-1.0, and RoBERTa-large-pair with k-fold cross-validation. The approach incorporates adversarial training on embedding layers, symmetric/transitive data augmentation, and sigmoid-space probability averaging to improve discrimination between paraphrased vs. semantically distinct question pairs. Additionally uses pseudo-labeling and threshold tuning (0.47) optimized on domain-specific data from Tianchi competition containing ~10K real clinical question pairs.

435 stars. No commits in the last 6 months.