Flare-NER (Financial NER)

Figure 1.

Goal

The Flare-NER dataset is designed for Named Entity Recognition (NER) in the financial domain. Its primary goals are:

  • To identify key financial entities such as Persons (PER), Organizations (ORG), and Locations (LOC) in unstructured financial texts.

  • To facilitate the extraction of structured information from financial agreements and SEC filings.

  • To provide a benchmark for evaluating NER models on specialized financial language.

Description

Flare-NER is a dataset curated by ChanceFocus, consisting of sentences extracted from financial filings.

Dataset Composition

  • Source: Sentences extracted from U.S. SEC filings, specifically financial agreements.

  • Entities: Annotated with three standard entity types relevant to finance:

    • PER (Person)

    • ORG (Organization)

    • LOC (Location)

  • Modality: Text

  • Size: Approximately 600+ rows (based on HuggingFace preview).

  • Format: Parsing-ready format for sequence labeling tasks.

Annotation Process

The dataset focuses on categorizing named entities within the complex sentence structures often found in legal and financial documents.

Example

Below is a representative example of the NER task:

Table 3 Flare-NER Examples

ID

Text

Entities

0

The Borrower shall pay the Bank of New York Mellon, as administrative agent.

ORG: Bank of New York Mellon

1

Mr. John Smith, designated as the authorized signatory for ABC Corp, signed the agreement in London.

PER: John Smith, ORG: ABC Corp, LOC: London

Task Description

Named Entity Recognition

  • Input: A sentence from a financial document (e.g., “Goldman Sachs reported earnings in New York.”)

  • Output: A sequence of tags indicating the start and type of entities (e.g., [B-ORG, I-ORG, O, O, O, B-LOC, I-LOC]).

  • Challenge: Financial texts often contain complex entity names (e.g., “The Bank of Tokyo-Mitsubishi UFJ, Ltd.”) and nested structures.

Evaluation Metrics

  1. F1-Score: The harmonic mean of precision and recall, strictly matching entity boundaries and types.

  2. Precision: The percentage of predicted entities that are correct.

  3. Recall: The percentage of actual entities that were correctly predicted.

Why Use Flare-NER?

  • Domain-Specific: Tailored for financial agreements and SEC filings, distinct from general news datasets (e.g., CoNLL-2003).

  • Structured Information Extraction: Essential for downstream tasks like relation extraction and knowledge graph construction in finance.

  • Open Access: Available on HuggingFace for research and development.

References

For dataset access and more details, visit the HuggingFace dataset page.