Edit on GitHub
  • John Snow Labs
  • Mar 26, 2023

Legal Arguments Mining in Court Decisions (in German)

  • de
  • licensed
  • classification
  • legal
  • tensorflow

Description

This is a Multiclass classification model in German which classifies arguments in legal discourse. These are the following classes: subsumption, definition, conclusion, other.

Predicted Entities

subsumption, definition, conclusion, other

Copy S3 URI

How to use

documentAssembler = nlp.DocumentAssembler()\
        .setInputCol("text")\
        .setOutputCol("document")

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

embeddings = nlp.RoBertaEmbeddings.pretrained("roberta_large_german_legal", "de")\
        .setInputCols(["document", "token"])\
        .setOutputCol("embeddings")\
        .setMaxSentenceLength(512)

embeddingsSentence = nlp.SentenceEmbeddings()\
        .setInputCols(["document", "embeddings"])\
        .setOutputCol("sentence_embeddings")\
        .setPoolingStrategy("AVERAGE")\


docClassifier = legal.ClassifierDLModel.pretrained("legclf_argument_mining_de", "de", "legal/models")\
        .setInputCols(["sentence_embeddings"])\
        .setOutputCol("category")

nlpPipeline = nlp.Pipeline(stages=[
      documentAssembler, 
      tokenizer,
      embeddings,
      embeddingsSentence,
      docClassifier
])

df = spark.createDataFrame([["Folglich liegt eine Verletzung von Artikel 8 der Konvention vor ."]]).toDF("text")

model = nlpPipeline.fit(df)
result = model.transform(df)

result.select("text", "category.result").show(truncate=False)

Results

+-----------------------------------------------------------------+------------+
|text                                                             |result      |
+-----------------------------------------------------------------+------------+
|Folglich liegt eine Verletzung von Artikel 8 der Konvention vor .|[conclusion]|
+-----------------------------------------------------------------+------------+

Model Information

Model Name: legclf_argument_mining_german
Compatibility: Legal NLP 1.0.0+
License: Licensed
Edition: Official
Input Labels: [sentence_embeddings]
Output Labels: [class]
Language: de
Size: 24.0 MB

References

Train dataset available here

Benchmarking

label         precision  recall    f1-score  support      
conclusion    0.88       0.88      0.88      52  
definition    0.83       0.83      0.83      58  
other         0.86       0.88      0.87      49  
subsumption   0.81       0.80      0.80      64  
accuracy         -          -      0.84      223                     
macro avg     0.85       0.85      0.85      223 
weighted avg  0.84       0.84      0.84      223