guillaumegenthial/sequence_tagging

Combines character-level and word-level embeddings (GloVe) through parallel BiLSTM pathways, then applies linear-chain CRF decoding for sequence labeling. Achieves F1 scores of 90-91 on standard benchmarks like CoNLL2003, with training optimized for GPU acceleration (110 seconds/epoch on Tesla K80). Designed for token-level tagging tasks beyond NER, accepting CoNLL-format data and integrating pretrained word vectors for transfer learning.

1,953 stars. No commits in the last 6 months.