Legal Whereas Clause Binary Classifier

Description

This model is a Binary Classifier (True, False) for the whereas clause type. To use this model, make sure you provide enough context as an input. Adding Sentence Splitters to the pipeline will make the model see only sentences, not the whole text, so it’s better to skip it, unless you want to do Binary Classification as sentence level.

If you have big legal documents, and you want to look for clauses, we recommend you to split the documents using any of the techniques available in our Legal NLP Workshop Tokenization & Splitting Tutorial (link here), namely:

• Paragraph splitting (by multiline);
• Splitting by headers / subheaders;
• etc.

Take into consideration the embeddings of this model allows up to 512 tokens. If you have more than that, consider splitting in smaller pieces (you can also check the same tutorial link provided above).

This model can be combined with any of the other “hundreds” of Legal Clauses Classifiers you will find in Models Hub, getting as an output a series of True/False values for each of the legal clause model you have added.

Predicted Entities

other, whereas

Live Demo Open in Colab Download Copy S3 URI

How to use

documentAssembler = nlp.DocumentAssembler() \
     .setInputCol("clause_text") \
     .setOutputCol("document")
  
embeddings = nlp.BertSentenceEmbeddings.pretrained("sent_bert_base_cased", "en") \
      .setInputCols("document") \
      .setOutputCol("sentence_embeddings")

docClassifier = nlp.ClassifierDLModel.pretrained("legclf_whereas_clause", "en", "legal/models")\
    .setInputCols(["sentence_embeddings"])\
    .setOutputCol("category")
    
nlpPipeline = nlp.Pipeline(stages=[
    documentAssembler, 
    embeddings,
    docClassifier])
 
df = spark.createDataFrame([["YOUR TEXT HERE"]]).toDF("clause_text")
model = nlpPipeline.fit(df)
result = model.transform(df)

Results

+-------+
| result|
+-------+
|[whereas]|
|[other]|
|[other]|
|[whereas]|

Model Information

References

In-house annotations on CUAD dataset

Benchmarking

       label  precision    recall  f1-score   support
       other       0.98      0.94      0.96        67
     whereas       0.91      0.98      0.94        41
    accuracy          -         -      0.95       108
   macro-avg       0.95      0.96      0.95       108
weighted-avg       0.96      0.95      0.95       108