DistilRoBERTa Token Classification - NER OntoNotes (distilroberta_base_token_classifier_ontonotes)

Description

RoBERTa Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks.

distilroberta_base_token_classifier_ontonotes is a fine-tuned RoBERTa model that is ready to use for Named Entity Recognition and achieves state-of-the-art performance for the NER task. This model has been trained to recognize four types of entities: CARDINAL, DATE, EVENT, FAC, GPE, LANGUAGE, LAW, LOC, MONEY, NORP, ORDINAL, ORG, PERCENT, PERSON, PRODUCT, QUANTITY, TIME, and WORK_OF_ART.

We used TFRobertaForTokenClassification to train this model and used RoBertaForTokenClassification annotator in Spark NLP 🚀 for prediction at scale!

Predicted Entities

CARDINAL, DATE, EVENT, FAC, GPE, LANGUAGE, LAW, LOC, MONEY, NORP, ORDINAL, ORG, PERCENT, PERSON, PRODUCT, QUANTITY, TIME, WORK_OF_ART

Download

How to use

document_assembler = DocumentAssembler() \
    .setInputCol('text') \
    .setOutputCol('document')

tokenizer = Tokenizer() \
    .setInputCols(['document']) \
    .setOutputCol('token')

tokenClassifier = RoBertaForTokenClassification \
      .pretrained('distilroberta_base_token_classifier_ontonotes', 'en') \
      .setInputCols(['token', 'document']) \
      .setOutputCol('ner') \
      .setCaseSensitive(True) \
      .setMaxSentenceLength(512)

# since output column is IOB/IOB2 style, NerConverter can extract entities
ner_converter = NerConverter() \
    .setInputCols(['document', 'token', 'ner']) \
    .setOutputCol('entities')

pipeline = Pipeline(stages=[
    document_assembler, 
    tokenizer,
    tokenClassifier,
    ner_converter
])

example = spark.createDataFrame([['My name is John!']]).toDF("text")
result = pipeline.fit(example).transform(example)
val document_assembler = DocumentAssembler() 
    .setInputCol("text") 
    .setOutputCol("document")

val tokenizer = Tokenizer() 
    .setInputCols("document") 
    .setOutputCol("token")

val tokenClassifier = RoBertaForTokenClassification.pretrained("distilroberta_base_token_classifier_ontonotes", "en")
      .setInputCols("document", "token")
      .setOutputCol("ner")
      .setCaseSensitive(true)
      .setMaxSentenceLength(512)

// since output column is IOB/IOB2 style, NerConverter can extract entities
val ner_converter = NerConverter() 
    .setInputCols("document", "token", "ner") 
    .setOutputCol("entities")

val pipeline = new Pipeline().setStages(Array(document_assembler, tokenizer, tokenClassifier, ner_converter))

val example = Seq.empty["My name is John!"].toDS.toDF("text")

val result = pipeline.fit(example).transform(example)

Results

+------------------------------------------------------------------------------------+
 |result                                                                              |
 +------------------------------------------------------------------------------------+
 |[B-PERSON, I-PERSON, O, O, O, B-LOC, O, O, O, B-LOC, O, O, O, O, B-PERSON, O, O, O, O, B-LOC]|
 +------------------------------------------------------------------------------------+

Model Information

Model Name: distilroberta_base_token_classifier_ontonotes
Compatibility: Spark NLP 3.3.0+
License: Open Source
Edition: Official
Input Labels: [token, document]
Output Labels: [ner]
Language: en
Case sensitive: true
Max sentense length: 512

Data Source

https://catalog.ldc.upenn.edu/LDC2013T19

Benchmarking

 B-CARDINAL       0.82      0.86      0.84       935
       B-DATE       0.86      0.85      0.85      1602
      B-EVENT       0.69      0.56      0.61        63
        B-FAC       0.66      0.59      0.62       135
        B-GPE       0.94      0.89      0.91      2240
   B-LANGUAGE       0.91      0.45      0.61        22
        B-LAW       0.91      0.53      0.67        40
        B-LOC       0.69      0.71      0.70       179
      B-MONEY       0.86      0.89      0.87       314
       B-NORP       0.84      0.87      0.85       841
    B-ORDINAL       0.81      0.89      0.85       195
        B-ORG       0.85      0.83      0.84      1795
    B-PERCENT       0.92      0.92      0.92       349
     B-PERSON       0.92      0.93      0.93      1988
    B-PRODUCT       0.64      0.64      0.64        76
   B-QUANTITY       0.73      0.76      0.74       105
       B-TIME       0.71      0.54      0.61       212
B-WORK_OF_ART       0.72      0.52      0.61       166
   I-CARDINAL       0.82      0.77      0.80       331
       I-DATE       0.87      0.88      0.88      2011
      I-EVENT       0.69      0.79      0.74       130
        I-FAC       0.72      0.73      0.73       213
        I-GPE       0.90      0.77      0.83       628
        I-LAW       0.98      0.60      0.75       106
        I-LOC       0.79      0.68      0.73       180
      I-MONEY       0.92      0.96      0.94       685
       I-NORP       0.86      0.57      0.68       160
    I-ORDINAL       0.00      0.00      0.00         4
        I-ORG       0.89      0.92      0.90      2406
    I-PERCENT       0.93      0.96      0.94       523
     I-PERSON       0.94      0.92      0.93      1412
    I-PRODUCT       0.69      0.71      0.70        69
   I-QUANTITY       0.79      0.87      0.82       206
       I-TIME       0.73      0.78      0.75       255
I-WORK_OF_ART       0.70      0.57      0.63       337
            O       0.99      0.99      0.99    131815

     accuracy                           0.97    152728
    macro avg       0.80      0.74      0.76    152728
 weighted avg       0.97      0.97      0.97    152728



processed 152728 tokens with 11257 phrases; found: 11382 phrases; correct: 9305.
accuracy:  85.75%; (non-O)
accuracy:  97.36%; precision:  81.75%; recall:  82.66%; FB1:  82.20
         CARDINAL: precision:  79.09%; recall:  83.74%; FB1:  81.35  990
             DATE: precision:  78.48%; recall:  81.02%; FB1:  79.73  1654
            EVENT: precision:  57.14%; recall:  57.14%; FB1:  57.14  63
              FAC: precision:  58.52%; recall:  58.52%; FB1:  58.52  135
              GPE: precision:  90.96%; recall:  86.21%; FB1:  88.52  2123
         LANGUAGE: precision:  90.91%; recall:  45.45%; FB1:  60.61  11
              LAW: precision:  68.97%; recall:  50.00%; FB1:  57.97  29
              LOC: precision:  63.30%; recall:  66.48%; FB1:  64.85  188
            MONEY: precision:  78.90%; recall:  86.94%; FB1:  82.73  346
             NORP: precision:  82.29%; recall:  86.21%; FB1:  84.20  881
          ORDINAL: precision:  80.84%; recall:  88.72%; FB1:  84.60  214
              ORG: precision:  78.98%; recall:  79.55%; FB1:  79.27  1808
          PERCENT: precision:  85.99%; recall:  87.97%; FB1:  86.97  357
           PERSON: precision:  88.43%; recall:  90.74%; FB1:  89.57  2040
          PRODUCT: precision:  58.97%; recall:  60.53%; FB1:  59.74  78
         QUANTITY: precision:  57.14%; recall:  72.38%; FB1:  63.87  133
             TIME: precision:  59.44%; recall:  50.47%; FB1:  54.59  180
      WORK_OF_ART: precision:  59.21%; recall:  54.22%; FB1:  56.60  152