Description
Pretrained named entity recognition deep learning model for diseases. The SparkNLP deep learning model (NerDL) is inspired by a former state of the art model for NER: Chiu & Nicols, Named Entity Recognition with Bidirectional LSTM-CNN.
Predicted Entities
Disease.
Live Demo Open in Colab Copy S3 URI
How to use
Use as part of an nlp pipeline with the following stages: DocumentAssembler, SentenceDetector, Tokenizer, WordEmbeddingsModel, NerDLModel. Add the NerConverter to the end of the pipeline to convert entity tokens into full entity chunks.
...
embeddings_clinical = WordEmbeddingsModel.pretrained("embeddings_clinical", "en", "clinical/models")\
.setInputCols(["sentence", "token"])\
.setOutputCol("embeddings")
clinical_ner = NerDLModel.pretrained("ner_diseases", "en", "clinical/models") \
.setInputCols(["sentence", "token", "embeddings"]) \
.setOutputCol("ner")
...
nlpPipeline = Pipeline(stages=[document_assembler, sentence_detector, tokenizer, embeddings_clinical, clinical_ner, ner_converter])
model = nlpPipeline.fit(spark.createDataFrame([[""]]).toDF("text"))
results = model.transform(spark.createDataFrame([['Detection of various other intracellular signaling proteins is also described. Genetic characterization of transactivation of the human T-cell leukemia virus type 1 promoter: Binding of Tax to Tax-responsive element 1 is mediated by the cyclic AMP-responsive members of the CREB/ATF family of transcription factors. To achieve a better understanding of the mechanism of transactivation by Tax of human T-cell leukemia virus type 1 Tax-responsive element 1 (TRE-1), we developed a genetic approach with Saccharomyces cerevisiae. We constructed a yeast reporter strain containing the lacZ gene under the control of the CYC1 promoter associated with three copies of TRE-1. Expression of either the cyclic AMP response element-binding protein (CREB) or CREB fused to the GAL4 activation domain (GAD) in this strain did not modify the expression of the reporter gene. Tax alone was also inactive. ']], ["text"]))
...
val embeddings_clinical = WordEmbeddingsModel.pretrained("embeddings_clinical", "en", "clinical/models")
.setInputCols(Array("sentence", "token"))
.setOutputCol("embeddings")
val ner = NerDLModel.pretrained("ner_diseases", "en", "clinical/models")
.setInputCols(Array("sentence", "token", "embeddings"))
.setOutputCol("ner")
...
val pipeline = new Pipeline().setStages(Array(document_assembler, sentence_detector, tokenizer, embeddings_clinical, ner, ner_converter))
val data = Seq("Detection of various other intracellular signaling proteins is also described. Genetic characterization of transactivation of the human T-cell leukemia virus type 1 promoter: Binding of Tax to Tax-responsive element 1 is mediated by the cyclic AMP-responsive members of the CREB/ATF family of transcription factors. To achieve a better understanding of the mechanism of transactivation by Tax of human T-cell leukemia virus type 1 Tax-responsive element 1 (TRE-1), we developed a genetic approach with Saccharomyces cerevisiae. We constructed a yeast reporter strain containing the lacZ gene under the control of the CYC1 promoter associated with three copies of TRE-1. Expression of either the cyclic AMP response element-binding protein (CREB) or CREB fused to the GAL4 activation domain (GAD) in this strain did not modify the expression of the reporter gene. Tax alone was also inactive. ").toDF("text")
val result = pipeline.fit(data).transform(data)
}
Results
The output is a dataframe with a sentence per row and a "ner" column containing all of the entity labels in the sentence, entity character indices, and other metadata. To get only the tokens and entity labels, without the metadata, select "token.result" and "ner.result" from your output dataframe, or add the "Finisher" to the end of your pipeline.
+------------------------------+---------+
|chunk |ner |
+------------------------------+---------+
|the cyst |Disease |
|a large Prolene suture |Disease |
|a very small incisional hernia|Disease |
|the hernia cavity |Disease |
|omentum |Disease |
|the hernia |Disease |
|the wound lesion |Disease |
|The lesion |Disease |
|the existing scar |Disease |
|the cyst |Disease |
|the wound |Disease |
|this cyst down to its base |Disease |
|a small incisional hernia |Disease |
|The cyst |Disease |
|The wound |Disease |
+------------------------------+---------+
Model Information
| Model Name: | ner_diseases_en_2.4.4_2.4 |
| Type: | ner |
| Compatibility: | Spark NLP 2.4.4+ |
| Edition: | Official |
| License: | Licensed |
| Input Labels: | [sentence,token, embeddings] |
| Output Labels: | [ner] |
| Language: | [en] |
| Case sensitive: | false |
Data Source
Trained with augmented version of i2b2 dataset with embeddings_clinical.
https://portal.dbmi.hms.harvard.edu/projects/n2c2-nlp/
Benchmarking
| | label | tp | fp | fn | prec | rec | f1 |
|---:|:--------------|-------:|-----:|-----:|---------:|---------:|---------:|
| 0 | I-Disease | 5014 | 222 | 171 | 0.957601 | 0.96702 | 0.962288 |
| 1 | B-Disease | 6004 | 213 | 159 | 0.965739 | 0.974201 | 0.969952 |
| 2 | Macro-average | 11018 | 435 | 330 | 0.96167 | 0.970611 | 0.96612 |
| 3 | Micro-average | 11018 | 435 | 330 | 0.962019 | 0.97092 | 0.966449 |
PREVIOUSClinical Analysis