FNXL (Financial Numeric Extraction)

Figure 1.

Goal

The FNXL dataset is designed for Numeric Entity Recognition and Linking in the financial domain. Its primary goals are:

  • To extract numeric values from financial texts and link them to rigorous financial concepts (taxonomy tags).

  • To handle extremely fine-grained semantic labels found in financial reporting standards (like XBRL).

  • To bridge the gap between unstructured financial text and structured financial data.

Description

FNXL is a dataset from ChanceFocus, dealing with the identification and categorization of numeric entities in SEC filings.

Dataset Composition

  • Task: Sequence Labeling (Token Classification) for numeric entities.

  • Labels: High-cardinality label set corresponding to financial taxonomy concepts (e.g., DeferredCompensationArrangementWithIndividualCompensationExpense).

  • Source: Financial reports and 10-K filings.

  • Modality: Text

  • Language: English

Annotation Process

The dataset involves tagging numbers with specific, standardized financial tags, often requiring deep domain knowledge and context from the surrounding text.

Example

Below is a representative example of the FNXL task:

Table 8 FNXL Examples

ID

Text

Tagged Entities

0

Compensation expense was $0.6 million in 2019.
$0.6: DeferredCompensationArrangement…

1

The company recorded a tax benefit of $1.2 billion.
$1.2: TaxCutsAndJobsActOf2017…

Task Description

Numeric Entity Labeling

  • Input: A financial sentence containing numerical data.

  • Output: A sequence of tags assigning specific financial concepts to the numbers.

  • Challenge: The sheer number of unique labels (thousands of potential XBRL tags) and the need to disambiguate similar numbers based on context.

Evaluation Metrics

  1. Micro-F1: Weighted average of F1 scores across all classes.

  2. Key-F1: Performance on the most frequent or important financial tags.

Why Use FNXL?

  • Automated Auditing: Verification of reported numbers against textual descriptions.

  • Financial Analysis: Extracting structured data series from unstructured notes.

  • Taxonomy Alignment: Direct mapping to standardized accounting principles.

References

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