mandarjoshi90/coref

Extends the end-to-end coreference architecture with BERT and SpanBERT encoders, leveraging subword tokenization and span representations for improved mention detection and linking. Provides pretrained models on OntoNotes (up to 79.6 F1 with SpanBERT-large) alongside training infrastructure with configurable segment length and dual learning rates for encoder vs. task-specific parameters. Supports batched inference via JSONL input with speaker and genre metadata, compatible with both TensorFlow and PyTorch checkpoint formats.

455 stars. No commits in the last 6 months.