Description
This is a Financial nlp.BertForTokenClassification NER model aimed to extract Job Titles / Roles of people in Companies, and was trained using Resumes, Wikipedia Articles, Financial and Legal documents, annotated in-house.
Predicted Entities
ROLE
How to use
from johnsnowlabs import *
documentAssembler = nlp.DocumentAssembler()\
.setInputCol("text")\
.setOutputCol("document")
tokenizer = nlp.Tokenizer()\
.setInputCols("document")\
.setOutputCol("token")
tokenClassifier = finance.BertForTokenClassification.pretrained("finner_bert_roles", "en", "finance/models")\
.setInputCols("token", "document")\
.setOutputCol("label")\
.setCaseSensitive(True)
ner_converter = nlp.NerConverter()\
.setInputCols(["document","token","label"])\
.setOutputCol("ner_chunk")
pipeline = nlp.Pipeline(stages=[
documentAssembler,
tokenizer,
tokenClassifier,
ner_converter
]
)
import pandas as pd
p_model = pipeline.fit(spark.createDataFrame(pd.DataFrame({'text': ['']})))
text = 'Jeffrey Preston Bezos is an American entrepreneur, founder and CEO of Amazon'
res = p_model.transform(spark.createDataFrame([[text]]).toDF("text"))
result_df = res.select(F.explode(F.arrays_zip(res.token.result,res.label.result, res.label.metadata)).alias("cols"))\
.select(F.expr("cols['0']").alias("token"),
F.expr("cols['1']").alias("label"),
F.expr("cols['2']['confidence']").alias("confidence"))
result_df.show(50, truncate=100)
Results
+------------+---------+----------+
| token|ner_label|confidence|
+------------+---------+----------+
| Jeffrey| O| 0.9984|
| Preston| O| 0.9878|
| Bezos| O| 0.9939|
| is| O| 0.999|
| an| O| 0.9988|
| American| B-ROLE| 0.8294|
|entrepreneur| I-ROLE| 0.9358|
| ,| O| 0.9979|
| founder| B-ROLE| 0.8645|
| and| O| 0.857|
| CEO| B-ROLE| 0.72|
| of| O| 0.995|
| Amazon| O| 0.9428|
+------------+---------+----------+
Model Information
| Model Name: | finner_bert_roles |
| Type: | finance |
| Compatibility: | Finance NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [sentence, token] |
| Output Labels: | [ner] |
| Language: | en |
| Size: | 402.8 MB |
| Case sensitive: | true |
| Max sentence length: | 128 |
References
In-house annotations on Wikidata, CUAD dataset, Financial 10-K documents and Resumes
Benchmarking
label tp fp fn prec rec f1
B-ROLE 3553 174 262 0.95331365 0.9313237 0.9421904
I-ROLE 4868 250 243 0.9511528 0.95245546 0.9518037
Macro-average 8421 424 505 0.9522332 0.9418896 0.9470331
Micro-average 8421 424 505 0.9520633 0.9434237 0.94772375