Description
This model is a Named Entity Recognition (NER) model focused on financial numeric items. It identifies 6 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
TreasuryStockValueAcquiredCostMethod, StockRepurchasedDuringPeriodShares, TreasuryStockAcquiredAverageCostPerShare, UnrecognizedTaxBenefits, SupplementalInformationForPropertyCasualtyInsuranceUnderwritersPriorYearClaimsAndClaimsAdjustmentExpense, TreasuryStockSharesAcquired
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_md_subset12', 'en', 'finance/models')\
.setInputCols(["sentence", "token", "embeddings"])\
.setOutputCol("ner")
pipeline = nlp.Pipeline(stages=[documentAssembler,
sentence,
tokenizer,
embeddings,
nerTagger
])
text = "At both May 31 , 2016 and November 30 , 2015 , the Company had $ 12.3 million of gross unrecognized tax benefits . "
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|
+------------+-------------------------+----------+
|token |ner_label |confidence|
+------------+-------------------------+----------+
|At |O |1.0 |
|both |O |1.0 |
|May |O |1.0 |
|31 |O |1.0 |
|, |O |1.0 |
|2016 |O |1.0 |
|and |O |1.0 |
|November |O |1.0 |
|30 |O |1.0 |
|, |O |1.0 |
|2015 |O |1.0 |
|, |O |1.0 |
|the |O |1.0 |
|Company |O |1.0 |
|had |O |1.0 |
|$ |O |1.0 |
|12.3 |B-UnrecognizedTaxBenefits|1.0 |
|million |O |1.0 |
|of |O |1.0 |
|gross |O |1.0 |
|unrecognized|O |1.0 |
|tax |O |1.0 |
|benefits |O |1.0 |
|. |O |1.0 |
+------------+-------------------------+----------+
Model Information
| Model Name: | finner_10q_xbrl_md_subset12 |
| Compatibility: | Finance NLP 1.0.0+ |
| License: | Licensed |
| Edition: | Official |
| Input Labels: | [sentence, token, embeddings] |
| Output Labels: | [ner] |
| Language: | en |
| Size: | 16.4 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-StockRepurchasedDuringPeriodShares 0.5798 0.4523 0.5082 241
B-SupplementalInformationForPropertyCasualtyInsuranceUnderwritersPriorYearClaimsAndClaimsAdjustmentExpense 0.9008 1.0000 0.9478 236
B-TreasuryStockAcquiredAverageCostPerShare 0.8359 0.9819 0.9030 166
B-TreasuryStockSharesAcquired 0.6911 0.8193 0.7497 415
B-TreasuryStockValueAcquiredCostMethod 0.7214 0.5153 0.6012 196
B-UnrecognizedTaxBenefits 0.9897 0.9897 0.9897 291
I-TreasuryStockSharesAcquired 0.0000 0.0000 0.0000 1
O 0.9969 0.9962 0.9965 32427
accuracy - - 0.9873 33973
macro-avg 0.7144 0.7193 0.7120 33973
weighted-avg 0.9870 0.9873 0.9869 33973