JF-ICR / README.md
xiangr's picture
Update README.md
0ec1422 verified
metadata
license: apache-2.0
language:
  - ja

Dataset Card for JF-ICR

Dataset Summary

JF-ICR (Japanese Financial Implicit Commitment Recognition) is a benchmark dataset for evaluating implicit commitment and refusal recognition in Japanese investor-facing Q&A. The dataset consists of 94 single-turn question-answer pairs from 4 companies spanning 3 years, annotated into 5 implicit stance categories that distinguish agreement, hedging, and refusal under high-context discourse. This dataset addresses the challenge of understanding indirect expression and implicit commitment in Japanese financial communication, where refusals are frequently conveyed through pragmatic cues rather than explicit statements.

Supported Tasks and Leaderboards

  • Evaluation Metrics:
    • Accuracy

Languages

  • Japanese (jp)

Dataset Structure

Data Instances

Each instance in the dataset contains:

  • question: A financial question posed by an investor or analyst in Japanese.
  • response: The corresponding company response in Japanese.
  • label: The annotated intent label from the set {+2, +1, 0, -1, -2}.
  • task_id: A unique identifier for the instance.

Data Fields

Field Type Description
question string Financial question in Japanese
response string Company response in Japanese
label string Intent label: +2, +1, 0, -1, or -2
task_id string Unique identifier for the instance

Label Definitions:

  • +2: Strong Commitment
  • +1: Weak or Qualified Commitment
  • 0: Neutral or Hedged Intent
  • -1: Weak Refusal
  • -2: Strong Refusal

Data Splits

Split # Examples
test 94

Note: The dataset consists of 94 instances from 4 companies spanning 3 years (2023-2026).

Dataset Creation

Curation Rationale

JF-ICR was curated to evaluate LLMs' ability to recognize implicit commitment and refusal in Japanese financial communication. The dataset addresses a critical gap in existing benchmarks, which typically focus on explicit sentiment or exam-style QA rather than pragmatic intent inference. Japanese financial communication relies heavily on indirect expression and high-context discourse, making this a linguistically and culturally grounded evaluation task that reflects real-world corporate communication practices.

Source Data

Initial Data Collection

The source data was collected from real-world, publicly available Japanese corporate disclosures, including:

  1. Earnings call / investor briefing Q&A transcripts from major Japanese companies
  2. Shareholder meeting Q&A materials
  3. Financial results briefings (決算説明会) Q&A sessions

The corpus comprises 8 source documents from 5 companies, with an average of multiple Q&A exchanges per document. To ensure unambiguous intent annotation, careful manual curation was performed: only single-turn question-answer pairs focusing on a single topic were retained, excluding multi-part questions, intertwined follow-ups, or exchanges requiring broader conversational context.

Source Producers

Data was collected from publicly available corporate disclosures published by Japanese publicly listed companies through official channels, including company investor relations websites and EDINET (Electronic Disclosure for Investors' NETwork), the official electronic disclosure system operated by Japan's Financial Services Agency.

Annotations

Annotation Process

Annotations were conducted following a rigorous, multi-step process:

  1. Guideline Development: A detailed annotation guideline was developed in collaboration with Japanese financial experts, specifying criteria for each of the five intent labels.

  2. Pre-annotation Rounds: Guidelines were iteratively refined through pre-annotation rounds, with particular attention to incorporating rules for ambiguous cases.

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

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

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

Annotation Philosophy: The task focuses on the degree of commitment expressed in management responses, rather than the surface intent of the question or factual correctness. Annotations are determined solely by whether and how the answer conveys a forward-looking action, decision, or policy stance. Only responses expressing strong, clear, explicit, and definitive agreement or refusal were assigned the ±2 labels.

Annotators

All annotations were performed by 2 native-level Japanese financial experts with extensive industry experience:

  • 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 a streamlined and reproducible workflow.

Inter-Annotator Agreement

Inter-annotator agreement results demonstrate high consistency:

  • Macro-F1: 0.9215
  • Cohen's κ: 0.8769
  • Krippendorff's α: 0.8768

These results indicate that the annotation guidelines are clearly defined and consistently applied, supporting the construction of a high-quality dataset.

Personal and Sensitive Information

The dataset contains no personal or sensitive information. All data is derived from publicly available corporate disclosures and investor communications that are already in the public domain. Company names and specific financial details are preserved as they are essential for understanding the context of the Q&A exchanges, but no individual personal information is included.

Considerations for Using the Data

Social Impact of Dataset

JF-ICR contributes to research in Japanese financial NLP by enabling evaluation of pragmatic understanding and implicit intent recognition in financial communication. This capability is important for cross-border finance, as foreign investors held 32.4% of listed share market value in Japan in FY2024. The dataset supports development of systems that can better understand Japanese corporate communication, potentially improving accessibility for international investors and analysts.

Discussion of Biases

  • Company Representation: The dataset covers 4 companies over 3 years, which may not represent the full diversity of Japanese corporate communication styles across different industries and company sizes.

  • Temporal Bias: Data spans 2023-2026, which may reflect communication patterns specific to this time period.

  • Question Types: The dataset focuses on investor-facing Q&A, which may emphasize certain types of questions and responses over others.

  • Annotation Subjectivity: Despite high inter-annotator agreement, annotation remains inherently subjective, particularly for edge cases involving ambiguous hedging or indirect refusals.

Other Known Limitations

  • Limited Size: The dataset contains 94 examples, which may limit statistical power for some evaluation scenarios.

  • Single-Turn Focus: Only single-turn Q&A pairs are included, excluding multi-turn conversational context that may be relevant in some scenarios.

  • Company-Specific: Limited to 4 companies, which may not generalize to all Japanese corporate communication styles.

  • High-Context Dependency: The dataset requires understanding of Japanese high-context communication norms, which may be challenging for models trained primarily on low-context languages.

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.