roberta-large-ner-english: model fine-tuned from roberta-large for NER task

Introduction

[roberta-large-ner-english] is an english NER model that was fine-tuned from roberta-large on conll2003 dataset. Model was validated on emails/chat data and outperformed other models on this type of data specifically. In particular the model seems to work better on entity that don't start with an upper case.

Training data

Training data was classified as follow:

Abbreviation	Description
O	Outside of a named entity
MISC	Miscellaneous entity
PER	Person’s name
ORG	Organization
LOC	Location

In order to simplify, the prefix B- or I- from original conll2003 was removed. I used the train and test dataset from original conll2003 for training and the "validation" dataset for validation. This resulted in a dataset of size:

Train	Validation
17494	3250

How to use roberta-large-ner-english with HuggingFace

Load roberta-large-ner-english and its sub-word tokenizer :

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("Jean-Baptiste/roberta-large-ner-english")
model = AutoModelForTokenClassification.from_pretrained("Jean-Baptiste/roberta-large-ner-english")


##### Process text sample (from wikipedia)

from transformers import pipeline

nlp = pipeline('ner', model=model, tokenizer=tokenizer, aggregation_strategy="simple")
nlp("Apple was founded in 1976 by Steve Jobs, Steve Wozniak and Ronald Wayne to develop and sell Wozniak's Apple I personal computer")


[{'entity_group': 'ORG',
  'score': 0.99381506,
  'word': ' Apple',
  'start': 0,
  'end': 5},
 {'entity_group': 'PER',
  'score': 0.99970853,
  'word': ' Steve Jobs',
  'start': 29,
  'end': 39},
 {'entity_group': 'PER',
  'score': 0.99981767,
  'word': ' Steve Wozniak',
  'start': 41,
  'end': 54},
 {'entity_group': 'PER',
  'score': 0.99956465,
  'word': ' Ronald Wayne',
  'start': 59,
  'end': 71},
 {'entity_group': 'PER',
  'score': 0.9997918,
  'word': ' Wozniak',
  'start': 92,
  'end': 99},
 {'entity_group': 'MISC',
  'score': 0.99956393,
  'word': ' Apple I',
  'start': 102,
  'end': 109}]

Model performances

Model performances computed on conll2003 validation dataset (computed on the tokens predictions)

entity	precision	recall	f1
PER	0.9914	0.9927	0.9920
ORG	0.9627	0.9661	0.9644
LOC	0.9795	0.9862	0.9828
MISC	0.9292	0.9262	0.9277
Overall	0.9740	0.9766	0.9753

On private dataset (email, chat, informal discussion), computed on word predictions:

entity	precision	recall	f1
PER	0.8823	0.9116	0.8967
ORG	0.7694	0.7292	0.7487
LOC	0.8619	0.7768	0.8171

By comparison on the same private dataset, Spacy (en_core_web_trf-3.2.0) was giving:

entity	precision	recall	f1
PER	0.9146	0.8287	0.8695
ORG	0.7655	0.6437	0.6993
LOC	0.8727	0.6180	0.7236

For those who could be interested, here is a short article on how I used the results of this model to train a LSTM model for signature detection in emails: https://medium.com/@jean-baptiste.polle/lstm-model-for-email-signature-detection-8e990384fefa