partial_bert_crf

Partial Bert-CRF

In many scenarios, named entity recognition (NER) models severely suffer from unlabeled entity problem, where the entities of a sentence may not be fully annotated. Partial CRF is a parameter estimation method for Conditional Random Fields (CRFs), which enables us to use such incomplete annotations (Tsuboi et al.).

data preprocessing

Partially annotated entity/span should be marked as B-P. You can process the trainset and set the path to dataset: data_files: train of configs/msra.yml.

Training a new model

python -m scripts.train -c examples/partial_bert_crf/configs/resume.yaml

Benchmarks

TODO