FSRL (Financial Semantic Role Labeling)

Figure 1.

Goal

The FSRL dataset is designed for Semantic Role Labeling (SRL) or Textual Analogy Parsing (TAP) in the financial domain. Its primary goals are:

  • To parse financial sentences into semantic arguments like TIME, VALUE, QUANT, AGENT, and THEME.

  • To identify “who did what to whom, when, and how much” in financial contexts.

  • To structure complex financial narratives into actionable data points.

Description

FSRL is a dataset from ChanceFocus, applying semantic role labeling frameworks to financial news and reports.

Dataset Composition

  • Task: Sequence Labeling (Token Classification).

  • Labels: A rich set of semantic roles tailored for finance:

    • TIME, REF_TIME: Temporal expressions.

    • VALUE, QUANT: Monetary amounts and quantities.

    • AGENT, THEME: Actors and subjects of actions.

    • MANNER, CAUSE, CONDITION: Modifiers explaining how/why.

  • Source: Financial news articles.

  • Modality: Text

  • Language: English

Annotation Process

Annotators tag every token in a sentence with its semantic role, providing a comprehensive understanding of the sentence structure and meanining.

Example

Below is a representative example of the FSRL task:

Table 9 FSRL Examples

ID

Text

Roles

0

Earnings totaled $36.6 million in the nine months.
THEME: Earnings
VALUE: $36.6 million
TIME: in the nine months

1

Shares fell 46% from a year earlier.
THEME: Shares
VALUE: 46%
TIME: a year earlier

Task Description

Semantic Parsing

  • Input: A financial sentence.

  • Output: A sequence of BIO tags assigning semantic roles to tokens (e.g., B-TIME, I-VALUE).

  • Challenge: Capturing the intricate relationships between numbers, dates, and entities in dense financial text.

Evaluation Metrics

  1. Span-based F1-Score: Measures the correctness of identified semantic arguments.

  2. Role Accuracy: Ensuring the correct label is assigned to the correct span.

Why Use FSRL?

  • Deep Understanding: Moves beyond entity recognition to understanding the relationships and roles of entities.

  • Question Answering: Enables systems to answer complex questions like “How much did earnings increase in Q3?”.

  • Information Extraction: Facilitates the extraction of structured event data from news feeds.

References

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