FinRED (Financial Relation Extraction)
Figure 1.
Goal
The FinRED dataset is designed for Relation Extraction (RE) in the financial domain. Its primary goals are:
To identify and classify semantic relationships between financial entities (e.g., “acquired by”, “subsidiary of”, “competitor of”).
To facilitate the construction of financial knowledge graphs.
To enable structured information retrieval from unstructured financial texts.
Description
FinRED is a relation extraction dataset from ChanceFocus, consisting of sentence-level relation annotations.
Dataset Composition
Source: Financial news and filings.
Task: Extract triples in the format
(Head Entity, Tail Entity, Relation).Size: 1K - 10K examples (based on HuggingFace metadata).
Modality: Text
Format: Parquet
Annotation Process
The dataset targets the extraction of specific financial relationships, requiring models to understand the context connecting two entities.
Example
Below is a representative example of the Relation Extraction task:
ID |
Text |
Relations (Head; Tail; Rel) |
|---|---|---|
0 |
Google acquired YouTube in a billion-dollar deal. |
Google; YouTube; acquired |
1 |
Apple, based in Cupertino, released the new iPhone. |
Apple; Cupertino; headquarters_location |
Task Description
Relation Extraction
Input: A sentence containing marked entities.
Output: A list of triplets
head; tail; relrepresenting the relationships found.Challenge: Handling multiple relations per sentence and identifying the directionality of relationships.
Evaluation Metrics
F1-Score: Harmonic mean of precision and recall for extracted triplets.
Precision: Percentage of extracted relations that are correct.
Recall: Percentage of ground truth relations that were extracted.
Why Use FinRED?
Structured Knowledge: Converts unstructured text into structured databases (Knowledge Graphs).
Financial Context: capturing specific relations relevant to markets (e.g., M&A, leadership changes).
Open Resource: Part of the FLARE benchmark for advancing financial NLP.
References
For dataset access and more details, visit the HuggingFace dataset page.