Description
IMPORTANT: Don’t run this model on the whole legal agreement. Instead:
- Split by paragraphs. You can use notebook 1 in Finance or Legal as inspiration;
- Use the
legclf_cuad_whereas_clauseText Classifier to select only these paragraphs;
This is a Legal NER Model, able to process WHEREAS clauses, to detect the SUBJECT (Who?), the ACTION, the OBJECT (what?) and, in some cases, the INDIRECT OBJECT (to whom?) of the clause.
This is a md (medium version) of the classifier, trained with more data and being more resistent to false positives outside the specific section, which may help to run it at whole document level (although not recommended).
Predicted Entities
WHEREAS_SUBJECT, WHEREAS_OBJECT, WHEREAS_ACTION
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")
embeddings = nlp.RoBertaEmbeddings.pretrained("roberta_embeddings_legal_roberta_base","en") \
.setInputCols(["sentence", "token"]) \
.setOutputCol("embeddings")
ner_model = legal.NerModel.pretrained('legner_whereas_md', 'en', 'legal/models')\
.setInputCols(["sentence", "token", "embeddings"])\
.setOutputCol("ner")
ner_converter = nlp.NerConverter()\
.setInputCols(["sentence","token","ner"])\
.setOutputCol("ner_chunk")
nlpPipeline = nlp.Pipeline(stages=[
documentAssembler,
sentenceDetector,
tokenizer,
embeddings,
ner_model,
ner_converter])
empty_data = spark.createDataFrame([[""]]).toDF("text")
model = nlpPipeline.fit(empty_data)
text = ["""WHEREAS, Seller and Buyer have entered into that certain Stock Purchase Agreement, dated November 14, 2018 (the "Stock Purchase Agreement"); WHEREAS, pursuant to the Stock Purchase Agreement, Seller has agreed to sell and transfer, and Buyer has agreed to purchase and acquire, all of Seller's right, title and interest in and to Armstrong Wood Products, Inc., a Delaware corporation ("AWP") and its Subsidiaries, the Company and HomerWood Hardwood Flooring Company, a Delaware corporation ("HHFC," and together with the Company, the "Company Subsidiaries" and together with AWP, the "Company Entities" and each a "Company Entity") by way of a purchase by Buyer and sale by Seller of the Shares, all upon the terms and condition set forth therein;"""]
res = model.transform(spark.createDataFrame([text]).toDF("text"))
Results
+------------+-----------------+
| token| ner_label|
+------------+-----------------+
| WHEREAS| O|
| ,| O|
| Seller|B-WHEREAS_SUBJECT|
| and| O|
| Buyer|B-WHEREAS_SUBJECT|
| have| B-WHEREAS_ACTION|
| entered| I-WHEREAS_ACTION|
| into| I-WHEREAS_ACTION|
| that| B-WHEREAS_OBJECT|
| certain| I-WHEREAS_OBJECT|
| Stock| I-WHEREAS_OBJECT|
| Purchase| I-WHEREAS_OBJECT|
| Agreement| I-WHEREAS_OBJECT|
| ,| O|
| dated| O|
| November| O|
| 14| O|
| ,| O|
| 2018| O|
| (| O|
| the| O|
| "| O|
| Stock| O|
| Purchase| O|
| Agreement| O|
| ");| O|
| WHEREAS| O|
| ,| O|
| pursuant| O|
| to| O|
| the| O|
| Stock| O|
| Purchase| O|
| Agreement| O|
| ,| O|
| Seller|B-WHEREAS_SUBJECT|
| has| B-WHEREAS_ACTION|
| agreed| I-WHEREAS_ACTION|
| to| I-WHEREAS_ACTION|
| sell| I-WHEREAS_ACTION|
| and| O|
| transfer| O|
| ,| O|
| and| O|
| Buyer|B-WHEREAS_SUBJECT|
| has| B-WHEREAS_ACTION|
| agreed| I-WHEREAS_ACTION|
| to| I-WHEREAS_ACTION|
| purchase| I-WHEREAS_ACTION|
| and| O|
| acquire| O|
| ,| O|
| all| O|
| of| O|
| Seller's| O|
| right| O|
| ,| O|
| title| O|
| and| O|
| interest| O|
| in| O|
| and| O|
| to| O|
| Armstrong| O|
| Wood| O|
| Products| O|
| ,| O|
| Inc| O|
| .,| O|
| a| O|
| Delaware| O|
| corporation| O|
| ("| O|
| AWP| O|
| ")| O|
| and| O|
| its| O|
|Subsidiaries| O|
| ,| O|
| the| O|
| Company| O|
| and| O|
| HomerWood| O|
| Hardwood| O|
| Flooring| O|
| Company| O|
| ,| O|
| a| O|
| Delaware| O|
| corporation| O|
| ("| O|
| HHFC| O|
| ,"| O|
+------------+-----------------+
Model Information
| Model Name: | legner_whereas_md |
| Compatibility: | Legal NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [sentence, token, embeddings] |
| Output Labels: | [ner] |
| Language: | en |
| Size: | 16.1 MB |
References
Manual annotations on CUAD dataset
Benchmarking
label tp fp fn prec rec f1
B-WHEREAS_SUBJECT 95 12 5 0.88785046 0.95 0.9178744
I-WHEREAS_ACTION 112 36 13 0.7567568 0.896 0.82051283
I-WHEREAS_SUBJECT 31 6 6 0.8378378 0.8378378 0.8378378
B-WHEREAS_OBJECT 59 33 30 0.6413044 0.66292137 0.6519337
B-WHEREAS_ACTION 87 12 3 0.8787879 0.96666664 0.9206349
I-WHEREAS_OBJECT 221 108 65 0.67173254 0.77272725 0.71869916
Macro-average 605 207 122 0.77904505 0.8476922 0.81192017
Micro-average 605 207 122 0.7450739 0.83218706 0.78622484