Description
IMPORTANT: Don’t run this model on the whole financial report. Instead:
- Split by paragraphs. You can use notebook 1 in Finance or Legal as inspiration;
- Use the
finclf_work_experience_itemText Classifier to select only these paragraphs;
This model allows you to analyzed present and past job positions of people, extracting relations between PERSON, ORG, ROLE and DATE. This model requires an NER with the mentioned entities, as finner_org_per_role and can also be combined with finassertiondl_past_roles to detect if the entities are mentioned to have happened in the PAST or not (although you can also infer that from the relations as had_role_until).
This model is a sm model without meaningful directions in the relations (the model was not trained to understand if the direction of the relation is from left to right or right to left). There are bigger models in Models Hub trained also with directed relationships.
Predicted Entities
has_role, had_role_until, has_role_from, works_for, has_role_in_company
How to use
document_assembler = nlp.DocumentAssembler()\
.setInputCol("text")\
.setOutputCol("document")
sentence_detector = nlp.SentenceDetectorDLModel.pretrained("sentence_detector_dl","en")\
.setInputCols(["document"])\
.setOutputCol("sentence")\
tokenizer = nlp.Tokenizer()\
.setInputCols(["sentence"])\
.setOutputCol("token")
embeddings = nlp.BertEmbeddings.pretrained("bert_embeddings_sec_bert_base","en") \
.setInputCols(["sentence", "token"]) \
.setOutputCol("embeddings")
ner_model = finance.NerModel.pretrained('finner_org_per_role_date', 'en', 'finance/models')\
.setInputCols(["sentence", "token", "embeddings"])\
.setOutputCol("ner")
ner_converter = nlp.NerConverter()\
.setInputCols(["sentence","token","ner"])\
.setOutputCol("ner_chunk")
pos = nlp.PerceptronModel.pretrained()\
.setInputCols(["sentence", "token"])\
.setOutputCol("pos")
dependency_parser = nlp.DependencyParserModel().pretrained("dependency_conllu", "en")\
.setInputCols(["sentence", "pos", "token"])\
.setOutputCol("dependencies")
re_ner_chunk_filter = finance.RENerChunksFilter()\
.setInputCols(["ner_chunk", "dependencies"])\
.setOutputCol("re_ner_chunk")\
.setRelationPairs(["PERSON-ROLE, ORG-ROLE, DATE-ROLE, PERSON-ORG"])\
.setMaxSyntacticDistance(5)
re_Model = finance.RelationExtractionDLModel.pretrained("finre_work_experience", "en", "finance/models")\
.setInputCols(["re_ner_chunk", "sentence"])\
.setOutputCol("relations")\
.setPredictionThreshold(0.5)
pipeline = nlp.Pipeline(stages=[
document_assembler,
sentence_detector,
tokenizer,
embeddings,
ner_model,
ner_converter,
pos,
dependency_parser,
re_ner_chunk_filter,
re_Model
])
empty_df = spark.createDataFrame([['']]).toDF("text")
re_model = pipeline.fit(empty_df)
light_model = LightPipeline(re_model)
text_list = ["""We have experienced significant changes in our senior management team over the past several years, including the appointments of Mark Schmitz as our Executive Vice President and Chief Operating Officer in 2019.""",
"""In January 2019, Jose Cil was assigned the CEO of Restaurant Brands International, and Daniel Schwartz was assigned the Executive Chairman of the company.""",
]
results = light_model.fullAnnotate(text_list)
Results
has_role PERSON 129 140 Mark Schmitz ROLE 149 172 Executive Vice President 0.8707728
has_role PERSON 129 140 Mark Schmitz ROLE 178 200 Chief Operating Officer 0.97559035
has_role_from ROLE 149 172 Executive Vice President DATE 205 208 2019 0.9327241
has_role_from ROLE 178 200 Chief Operating Officer DATE 205 208 2019 0.90718126
has_role_from DATE 3 14 January 2019 ROLE 43 45 CEO 0.996639
has_role_from DATE 3 14 January 2019 ROLE 120 137 Executive Chairman 0.9964874
has_role PERSON 17 24 Jose Cil ROLE 43 45 CEO 0.8917691
has_role PERSON 17 24 Jose Cil ROLE 120 137 Executive Chairman 0.8527716
has_role ROLE 43 45 CEO PERSON 87 101 Daniel Schwartz 0.5765097
has_role PERSON 87 101 Daniel Schwartz ROLE 120 137 Executive Chairman 0.79235893
Model Information
| Model Name: | finre_work_experience |
| Compatibility: | Finance NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Language: | en |
| Size: | 409.9 MB |
References
Manual annotations on CUAD dataset, 10K filings and Wikidata
Benchmarking
label Recall Precision F1 Support
had_role_until 0.972 0.972 0.972 36
has_role 0.986 0.980 0.983 146
has_role_from 0.983 0.983 0.983 58
has_role_in_company 0.954 0.969 0.961 65
works_for 0.933 0.933 0.933 15
Avg. 0.966 0.967 0.966 -
Weighted-Avg. 0.975 0.975 0.975 -