XLNet Token Classification Large - NER CoNLL (xlnet_large_token_classifier_conll03)

Description

XLNet 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.

xlnet_large_token_classifier_conll03 is a fine-tuned XLNet 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: location (LOC), organizations (ORG), person (PER), and Miscellaneous (MISC).

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

Predicted Entities

Download

How to use

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

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

tokenClassifier = XlnetForTokenClassification \
      .pretrained('xlnet_large_token_classifier_conll03', '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 = XlnetForTokenClassification.pretrained("xlnet_large_token_classifier_conll03", "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-PER, I-PER, O, O, O, B-LOC, O, O, O, B-LOC, O, O, O, O, B-PER, O, O, O, O, B-LOC]|
 +------------------------------------------------------------------------------------+

Model Information

Model Name: xlnet_large_token_classifier_conll03
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://www.clips.uantwerpen.be/conll2003/ner/

Benchmarking

          precision    recall  f1-score   support

       B-LOC       0.95      0.96      0.96      1837
      B-MISC       0.83      0.93      0.87       922
       B-ORG       0.90      0.93      0.91      1341
       B-PER       0.97      0.97      0.97      1842
       I-LOC       0.93      0.93      0.93       257
      I-MISC       0.86      0.83      0.84       346
       I-ORG       0.90      0.89      0.89       751
       I-PER       0.96      0.98      0.97      1307
           O       1.00      0.99      1.00     42759

    accuracy                           0.99     51362
   macro avg       0.92      0.93      0.93     51362
weighted avg       0.99      0.99      0.99     51362



processed 51362 tokens with 5942 phrases; found: 6168 phrases; correct: 5552.
accuracy:  94.56%; (non-O)
accuracy:  98.57%; precision:  90.01%; recall:  93.44%; FB1:  91.69
              LOC: precision:  94.46%; recall:  95.65%; FB1:  95.05  1860
             MISC: precision:  79.41%; recall:  89.91%; FB1:  84.33  1044
              ORG: precision:  85.47%; recall:  90.38%; FB1:  87.86  1418
              PER: precision:  95.02%; recall:  95.22%; FB1:  95.12  1846