Description
This model is a Named Entity Recognition (NER) model focused on financial numeric items. It identifies 5 numeric financial entities from diverse 10-Q and 10-K reports. These entities are annotated using eXtensible Business Reporting Language (XBRL) tags. The annotation process primarily targets numerical tokens, and the context plays a crucial role in accurately assigning the appropriate entity type from the 139 most common financial entities available in the dataset.
Predicted Entities
IncomeLossFromEquityMethodInvestments, DisposalGroupIncludingDiscontinuedOperationConsideration, GainsLossesOnExtinguishmentOfDebt, DerivativeFixedInterestRate, IncomeTaxExpenseBenefit
How to use
documentAssembler = nlp.DocumentAssembler() \
.setInputCol("text") \
.setOutputCol("document")
sentence = nlp.SentenceDetector() \
.setInputCols(["document"]) \
.setOutputCol("sentence")
tokenizer = nlp.Tokenizer()\
.setInputCols(["sentence"])\
.setOutputCol("token")\
.setContextChars(['.', ',', ';', ':', '!', '?', '*', '-', '(', ')', '”', '’', '$','€'])
embeddings = nlp.BertEmbeddings.pretrained("bert_embeddings_sec_bert_base","en") \
.setInputCols(["document", "token"]) \
.setOutputCol("embeddings")\
.setMaxSentenceLength(512)
nerTagger = finance.NerModel.pretrained('finner_10q_xbrl_lg_income', 'en', 'finance/models')\
.setInputCols(["sentence", "token", "embeddings"])\
.setOutputCol("ner")
pipeline = nlp.Pipeline(stages=[documentAssembler,
sentence,
tokenizer,
embeddings,
nerTagger
])
text = "This was partially offset by $ 6.7 million and $ 12.7 million , respectively , of equity in earnings from one of the Company 's unconsolidated entities for the three and six months ended May 31 , 2016 primarily due to sales of 253 homesites and 471 homesites , respectively , to third parties for $ 52.1 million and $ 114.1 million , respectively , that resulted in gross profits of $ 18.3 million and $ 39.0 million , respectively ."
df = spark.createDataFrame([[text]]).toDF("text")
fit = pipeline.fit(df)
result = fit.transform(df)
result_df = result.select(F.explode(F.arrays_zip(result.token.result,result.ner.result, result.ner.metadata)).alias("cols"))\
.select(F.expr("cols['0']").alias("token"),\
F.expr("cols['1']").alias("ner_label"),\
F.expr("cols['2']['confidence']").alias("confidence"))
result_df.show(50, truncate=100)
Results
+--------------+---------------------------------------+----------+
|token |ner_label |confidence|
+--------------+---------------------------------------+----------+
|This |O |1.0 |
|was |O |1.0 |
|partially |O |1.0 |
|offset |O |1.0 |
|by |O |1.0 |
|$ |O |1.0 |
|6.7 |B-IncomeLossFromEquityMethodInvestments|0.9998 |
|million |O |1.0 |
|and |O |1.0 |
|$ |O |1.0 |
|12.7 |B-IncomeLossFromEquityMethodInvestments|0.9975 |
|million |O |1.0 |
|, |O |1.0 |
|respectively |O |1.0 |
|, |O |1.0 |
|of |O |1.0 |
|equity |O |1.0 |
|in |O |1.0 |
|earnings |O |1.0 |
|from |O |1.0 |
|one |O |1.0 |
|of |O |1.0 |
|the |O |1.0 |
|Company |O |1.0 |
|'s |O |1.0 |
|unconsolidated|O |1.0 |
|entities |O |1.0 |
|for |O |1.0 |
|the |O |1.0 |
|three |O |1.0 |
|and |O |1.0 |
|six |O |1.0 |
|months |O |1.0 |
|ended |O |1.0 |
|May |O |1.0 |
|31 |O |1.0 |
|, |O |1.0 |
|2016 |O |1.0 |
|primarily |O |1.0 |
|due |O |1.0 |
|to |O |1.0 |
|sales |O |1.0 |
|of |O |1.0 |
|253 |O |0.9997 |
|homesites |O |1.0 |
|and |O |1.0 |
|471 |O |0.9999 |
|homesites |O |1.0 |
|, |O |1.0 |
|respectively |O |1.0 |
|, |O |1.0 |
|to |O |1.0 |
|third |O |1.0 |
|parties |O |1.0 |
|for |O |1.0 |
|$ |O |1.0 |
|52.1 |O |0.9845 |
|million |O |1.0 |
|and |O |1.0 |
|$ |O |1.0 |
|114.1 |O |0.9984 |
|million |O |1.0 |
|, |O |1.0 |
|respectively |O |1.0 |
|, |O |1.0 |
|that |O |1.0 |
|resulted |O |1.0 |
|in |O |1.0 |
|gross |O |1.0 |
|profits |O |1.0 |
|of |O |1.0 |
|$ |O |1.0 |
|18.3 |O |0.7789 |
|million |O |1.0 |
|and |O |1.0 |
|$ |O |1.0 |
|39.0 |O |0.9363 |
|million |O |1.0 |
|, |O |1.0 |
|respectively |O |1.0 |
|. |O |1.0 |
+--------------+---------------------------------------+----------+
Model Information
| Model Name: | finner_10q_xbrl_lg_income |
| Compatibility: | Finance NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [sentence, token, embeddings] |
| Output Labels: | [ner] |
| Language: | en |
| Size: | 16.3 MB |
References
An in-house modified version of https://huggingface.co/datasets/nlpaueb/finer-139, re-splited and filtered to focus on sentences with bigger density of tags.
Benchmarking
label precision recall f1-score support
B-DerivativeFixedInterestRate 0.9858 1.0000 0.9929 139
B-DisposalGroupIncludingDiscontinuedOperationConsideration 0.9756 0.9479 0.9615 211
B-GainsLossesOnExtinguishmentOfDebt 0.9502 0.9709 0.9604 275
B-IncomeLossFromEquityMethodInvestments 0.9734 0.9846 0.9790 260
B-IncomeTaxExpenseBenefit 0.9901 0.9832 0.9866 713
O 0.9993 0.9993 0.9993 43335
accuracy - - 0.9986 44933
macro-avg 0.9791 0.9810 0.9800 44933
weighted-avg 0.9986 0.9986 0.9986 44933