luopeixiang/named_entity_recognition

Implements four sequential architectures—HMM with Viterbi decoding, CRF with hand-crafted feature functions, BiLSTM, and BiLSTM+CRF—trained on resume NER data using BIOES tagging, with ensemble voting combining all models. Built in PyTorch, the project provides complete implementations including OOV handling, log-space probability computation to prevent underflow, and dynamic programming loss calculation for CRF layers. Achieves 95.7% F1 with BiLSTM+CRF alone on the ResumeNER dataset included in the repository.

2,275 stars. No commits in the last 6 months.