Description
This model is a Deep Learning Portuguese Named Entity Recognition model for the legal domain, trained using Base Bert Embeddings, and is able to predict the following entities:
- ORGANIZACAO (Organizations)
- JURISPRUDENCIA (Jurisprudence)
- PESSOA (Person)
- TEMPO (Time)
- LOCAL (Location)
- LEGISLACAO (Laws)
- O (Other)
You can find different versions of this model in Models Hub:
- With a Deep Learning architecture (non-transformer) and Base Embeddings;
- With a Deep Learning architecture (non-transformer) and Large Embeddings;
- With a Transformers Architecture and Base Embeddings;
- With a Transformers Architecture and Large Embeddings;
Predicted Entities
PESSOA, ORGANIZACAO, LEGISLACAO, JURISPRUDENCIA, TEMPO, LOCAL
How to use
documentAssembler = nlp.DocumentAssembler() \
.setInputCol("text") \
.setOutputCol("document")
sentenceDetector = nlp.SentenceDetectorDLModel.pretrained("sentence_detector_dl", "xx")\
.setInputCols(["document"])\
.setOutputCol("sentence")
tokenizer = nlp.Tokenizer() \
.setInputCols("sentence") \
.setOutputCol("token")
tokenClassifier = nlp.BertForTokenClassification.pretrained("legner_br_bert_base","pt", "legal/models") \
.setInputCols(["sentence", "token"]) \
.setOutputCol("ner")
pipeline = nlp.Pipeline(
stages=[
documentAssembler,
sentenceDetector,
tokenizer,
tokenClassifier])
example = spark.createDataFrame(pd.DataFrame({'text': ["""Mediante do exposto , com fundamento nos artigos 32 , i , e 33 , da lei 8.443/1992 , submetem-se os autos à consideração superior , com posterior encaminhamento ao ministério público junto ao tcu e ao gabinete do relator , propondo : a ) conhecer do recurso e , no mérito , negar-lhe provimento ; b ) comunicar ao recorrente , ao superior tribunal militar e ao tribunal regional federal da 2ª região , a fim de fornecer subsídios para os processos judiciais 2001.34.00.024796-9 e 2003.34.00.044227-3 ; e aos demais interessados a deliberação que vier a ser proferida por esta corte ” ."""]}))
result = pipeline.fit(example).transform(example)
Results
+-------------------+----------------+
| token| ner|
+-------------------+----------------+
| diante| O|
| do| O|
| exposto| O|
| ,| O|
| com| O|
| fundamento| O|
| nos| O|
| artigos| B-LEGISLACAO|
| 32| I-LEGISLACAO|
| ,| I-LEGISLACAO|
| i| I-LEGISLACAO|
| ,| I-LEGISLACAO|
| e| I-LEGISLACAO|
| 33| I-LEGISLACAO|
| ,| I-LEGISLACAO|
| da| I-LEGISLACAO|
| lei| I-LEGISLACAO|
| 8.443/1992| I-LEGISLACAO|
| ,| O|
| submetem-se| O|
| os| O|
| autos| O|
| à| O|
| consideração| O|
| superior| O|
| ,| O|
| com| O|
| posterior| O|
| encaminhamento| O|
| ao| O|
| ministério| B-ORGANIZACAO|
| público| I-ORGANIZACAO|
| junto| O|
| ao| O|
| tcu| B-ORGANIZACAO|
| e| O|
| ao| O|
| gabinete| O|
| do| O|
| relator| O|
| ,| O|
| propondo| O|
| :| O|
| a| O|
| )| O|
| conhecer| O|
| do| O|
| recurso| O|
| e| O|
| ,| O|
| no| O|
| mérito| O|
| ,| O|
| negar-lhe| O|
| provimento| O|
| ;| O|
| b| O|
| )| O|
| comunicar| O|
| ao| O|
| recorrente| O|
| ,| O|
| ao| O|
| superior| B-ORGANIZACAO|
| tribunal| I-ORGANIZACAO|
| militar| I-ORGANIZACAO|
| e| O|
| ao| O|
| tribunal| B-ORGANIZACAO|
| regional| I-ORGANIZACAO|
| federal| I-ORGANIZACAO|
| da| I-ORGANIZACAO|
| 2ª| I-ORGANIZACAO|
| região| I-ORGANIZACAO|
| ,| O|
| a| O|
| fim| O|
| de| O|
| fornecer| O|
| subsídios| O|
| para| O|
| os| O|
| processos| O|
| judiciais| O|
|2001.34.00.024796-9|B-JURISPRUDENCIA|
| e| O|
|2003.34.00.044227-3|B-JURISPRUDENCIA|
| ;| O|
| e| O|
| aos| O|
| demais| O|
| interessados| O|
| a| O|
| deliberação| O|
| que| O|
| vier| O|
| a| O|
| ser| O|
| proferida| O|
| por| O|
| esta| O|
| corte| O|
| ”| O|
| .| O|
+-------------------+----------------+
Model Information
| Model Name: | legner_br_bert_base |
| Compatibility: | Legal NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [document, token] |
| Output Labels: | [ner] |
| Language: | pt |
| Size: | 406.5 MB |
| Case sensitive: | true |
| Max sentence length: | 128 |
References
Original texts available in https://paperswithcode.com/sota?task=Token+Classification&dataset=lener_br and in-house data augmentation with weak labelling
Benchmarking
label precision recall f1-score support
B-ORGANIZACAO 0.86 0.86 0.86 499
I-ORGANIZACAO 0.89 0.89 0.89 859
B-LEGISLACAO 0.94 0.94 0.94 373
I-LEGISLACAO 0.96 0.98 0.97 2235
B-JURISPRUDENCIA 0.76 0.54 0.63 183
I-JURISPRUDENCIA 0.87 0.79 0.83 475
B-TEMPO 0.92 0.61 0.74 192
I-TEMPO 0.90 0.93 0.91 68
B-PESSOA 0.93 0.96 0.95 231
I-PESSOA 0.96 0.99 0.97 494
B-LOCAL 0.78 0.81 0.79 47
I-LOCAL 0.59 0.74 0.66 85
micro-avg 0.91 0.91 0.91 5741
macro-avg 0.86 0.84 0.84 5741
weighted-avg 0.91 0.91 0.91 5741