TheFinAI/JF-TE-private

Dataset Card for JF-TE

Dataset Summary

JF-TE (Japanese Financial Term Extraction) is a benchmark dataset for evaluating hierarchical extraction and ranking of nested financial terminology from Japanese professional disclosures. The dataset consists of 202 note-level instances extracted from 10 professional disclosures, containing 2,412 expert-curated term mentions covering 777 unique finance terms after normalization. The task addresses the challenge of boundary-sensitive grounding of finance terminology in mixed-script text, where nested compounds and script-variant loanwords (kanji, hiragana, and katakana) make term boundaries and semantic scope ambiguous. This dataset provides a linguistically grounded evaluation of financial term extraction that reflects the complexity of real Japanese financial disclosures.

Supported Tasks and Leaderboards

• Evaluation Metrics:
  • Maximal Financial Term F1
  • HitRate@1
  • HitRate@5
  • HitRate@10

Languages

• Japanese (jp)

Dataset Structure

Data Instances

Each instance in the dataset contains:

• text: A note-level excerpt from a Japanese financial disclosure (有価証券報告書 - Annual Securities Report).
• terms: A list of annotated financial term spans, including both maximal (longest) terms and nested terms.
• task_id: A unique identifier for the instance.

Data Fields

Field	Type	Description
text	string	Note-level excerpt from financial disclosure
terms	list	List of annotated financial term spans with start/end positions
task_id	string	Unique identifier for the instance

Term Annotation Structure:

• Each term annotation includes:
  • span: The text span of the financial term
  • start: Character start position
  • end: Character end position
  • is_nested: Boolean indicating if the term is nested within a longer term

Data Splits

Split	# Examples	# Term Mentions	Unique Terms
test	202	2,412	777

Note: The dataset consists of 202 note-level instances from 10 professional disclosures. Each disclosure spans multiple pages, with substantive financial information embedded in fine-grained appendix notes.

Dataset Creation

Curation Rationale

JF-TE was curated to evaluate LLMs' ability to extract and rank financial terminology from Japanese professional disclosures, addressing the unique challenges of Japanese financial term formation. Unlike English financial terminology, Japanese financial terms often appear as nested nominal compounds with mixed script variants (kanji, hiragana, and katakana), making term boundary identification and semantic scope resolution particularly challenging. The dataset focuses on note-level instances because substantive updates and domain-specific information in Japanese financial disclosures are often embedded in fine-grained appendix notes rather than main narrative sections.

Source Data

Initial Data Collection

The source data was collected from Japanese Annual Securities Reports (有価証券報告書 - Yūka shōken hōkokusho) disclosed through EDINET (Electronic Disclosure for Investors' NETwork), the official electronic disclosure system operated by Japan's Financial Services Agency. The corpus covers professional disclosures from publicly listed companies, with each report spanning multiple pages.

To capture information-dense units, individual note entries were manually identified and extracted from each financial disclosure, treating each note as a standalone instance for expert annotation and financial terminology extraction. This process yields 202 curated note-level data instances.

Source Producers

Data was collected from publicly available professional disclosures published by Japanese publicly listed companies through EDINET, the official electronic disclosure system operated by Japan's Financial Services Agency. All data is in the public domain and accessible through official channels.

Annotations

Annotation Process

Annotations were conducted following a rigorous protocol:

1. Guideline Development: Task-specific guidelines were developed tailored to hierarchical financial terminology, with explicit rules governing term boundary decisions and the distinction between financial terminology and general expressions.

2. Iterative Refinement: Annotation guidelines were iteratively refined through pre-annotation rounds, with particular attention to:

   • Term boundary decisions
   • Distinction between financial terminology and general expressions
   • Handling of nested and overlapping terms
   • Mixed-script variant identification

3. Span-Level Annotation: Annotation was conducted at the span level, allowing multi-word expressions, compound nouns, and nested structures to be marked explicitly.

4. Double Annotation: Each instance was independently annotated by 2 native-level Japanese financial experts.

5. Adjudication: Disagreements were resolved through adjudication by a senior expert.

6. Quality Validation: Inter-annotator agreement was measured using Span-level F1, Cohen's κ, and Krippendorff's α to ensure reliability.

Annotation Rules:

• Both nested and non-nested financial terms should be annotated if they independently represent meaningful financial or accounting concepts.
• When a longer term contains a shorter term, each should be annotated separately, provided that each span has standalone financial meaning.
• Annotators should always select the minimal span for each term, even when multiple annotated spans overlap.
• Nested terms should not be merged unless the shorter term lacks independent financial meaning outside the longer expression.

Annotators

All annotations were performed by the same native-level Japanese financial experts as in JF-ICR:

• Annotator 1: A doctoral student in Japan with a solid academic background in financial mathematics and several years of study and residence in Japan. Their research focuses on AI applications in finance, and they have previously contributed to financial benchmark annotation. Prior to doctoral studies, they accumulated professional experience as a quantitative researcher in the financial industry.

• Annotator 2: A researcher at a Japanese fintech company with nearly two decades of experience in the Japanese financial market. They hold a Ph.D. in Economics and possess deep expertise in Japanese corporate finance, financial reporting, and investor communications. As a native-level Japanese speaker, they are familiar with the linguistic, cultural, and institutional conventions of Japanese financial disclosures.

The annotation process was conducted using the Label Studio platform, ensuring consistency and reproducibility.

Inter-Annotator Agreement

Inter-annotator agreement results demonstrate strong consistency in identifying financial term spans:

• Span-level F1: [Reported in paper]
• Cohen's κ: [Reported in paper]
• Krippendorff's α: [Reported in paper]

These results indicate that the annotation rules governing term boundaries and financial termhood are operational and reproducible, supporting the reliability of the resulting high-quality JF-TE dataset.

Personal and Sensitive Information

The dataset contains no personal or sensitive information. All data is derived from publicly available professional disclosures (有価証券報告書) that are required to be published by publicly listed companies in Japan. Company names and financial details are preserved as they are essential for understanding the context and meaning of the financial terminology, but no individual personal information is included.

Considerations for Using the Data

Social Impact of Dataset

JF-TE contributes to research in Japanese financial NLP by enabling evaluation of financial term extraction capabilities in linguistically complex settings. Accurate extraction of financial terminology is essential for downstream tasks such as financial analysis, regulatory compliance, and automated financial document processing. The dataset supports development of systems that can better understand Japanese financial disclosures, potentially improving accessibility and analysis capabilities for both domestic and international stakeholders.

Discussion of Biases

• Document Type: The dataset focuses on Annual Securities Reports (有価証券報告書), which may not represent all types of Japanese financial disclosures.

• Company Representation: The dataset covers 10 professional disclosures, which may not represent the full diversity of Japanese corporate communication styles across different industries and company sizes.

• Term Coverage: The dataset emphasizes financial and accounting terminology, which may underrepresent other types of domain-specific language that appear in financial disclosures.

• Script Representation: While the dataset includes mixed-script terms, the distribution may reflect patterns specific to the selected disclosures.

Other Known Limitations

• Limited Size: The dataset contains 202 note-level instances, which may limit coverage of all possible financial term patterns.

• Note-Level Focus: The dataset focuses on note-level excerpts rather than full documents, which may miss some contextual information relevant for term extraction.

• Nested Term Complexity: The hierarchical nature of nested terms may make evaluation challenging, as different models may prioritize different levels of the hierarchy.

• Mixed-Script Handling: The complexity of mixed-script Japanese text (kanji, hiragana, katakana) may pose challenges for models not specifically trained on Japanese text.

Additional Information

Dataset Curators

• The Ebisu Benchmark Team

Licensing Information

• License: Apache License 2.0

Citation Information

If you use this dataset, please cite:

@misc{ebisu2025,
      title={EBISU: Benchmarking Large Language Models in Japanese Finance}, 
      author={[Authors]},
      year={2025},
      eprint={[arXiv number]},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/[number]}, 
}

Note: Please check the official paper for the complete citation once published.