multilabel-classification-bert-ace2005

Overview

• This is a BERT-based multi-label token classification model fine tuned on the ACE2005 dataset.
• The entities are one-hot encoded using the BIES (Begin/Inside/End/Single) scheme. As this is a multi-label model, there is no "Outside" label, for clasically outside tokens no class is predicted.
• The model comes with a pipeline to extract named entities from the model predictions
• For a short overview of the adaptions for multi-label token classification, see the non-finetuned parent model jvaquet/multilabel-classification-bert.

Pipeline Usage

Using the NER pipeline is rahter simple:

from transformers import pipeline

pipe = pipeline(model='jvaquet/multilabel-classification-bert-ace2005',
  stride=128,
  threshold=0.5,
  use_hierarchy_heuristic=False,
  trust_remote_code=True)

entities = pipe(my_text)

The parameters are:

• stride - int: Stride for the tokenizer. When the text length exceeds tokenizer.model_max_length, it splits the input accordingly with the specified stride.
• threshold - float: Threshold for entitiy detection. Sigmoid of the logits.
• use_hierarchy_heuristic - bool: Apply heuristic to suppress additional entities when entities of same class overlap hierarchically.