v0.5.0

• Spans are now at the token level, making the model more efficient.
• tibert.run_test now correctly supports all possible document splits configuration.
• Fix a bug in mention pruning that would cause successive mention to always be discarded.
• Fix a bug when selecting antecedents that would greatly impact model's performance, specifically in terms of MUC Recall.

v0.4.0

• Add a first version of hierarchical merging
• Fix a bug where the distance feature was not returning coherent results
• Add the tibert/run_test.py script for testing performance
• Add a new utility function, tibert.utils.limit_doc_size_tokens_
• tibert.utils.pprint_coreference_document now has useful colors!

v0.3.2

• critical fix for score_lea

v0.3.1

• support for BLANC and LEA
• critical fix for score_muc

v0.2.4

• Support for Democrat
• New prediction interface : stream_predict_coref

v0.2.3

• small enhancements
• dependencies compat with Grimbert

v0.2.2

• Fix a crash when restoring training checkpoint

v0.2.1

• Avoid truncating predictions for long sequences
• Better checkpoint saving for training

v0.2.0

Bugfixs and support for Camembert (pre-trained model coming soon)

v0.1.0

https://pypi.org/project/tibert/0.1.0/