AAU-NLP/HiFi-KPI

HiFi-KPI: Hierarchical Financial KPI Extraction

Paper: HiFi-KPI: A Dataset for Hierarchical KPI Extraction from Earnings Filings

Dataset Summary

HiFi-KPI is a large-scale dataset designed for financial numerical key performance indicator (KPI) extraction from earnings filings. It is derived from iXBRL filings mandated by the SEC, featuring hierarchical labels structured from the XBRL taxonomy. The dataset consists of ∼1.8M paragraphs and ∼5M entities, each linked to labels in the iXBRL calculation and presentation taxonomies.

Languages

The dataset is in English, extracted from SEC 10-K and 10-Q filings.

Dataset Structure

Data Fields

Each entry in HiFi-KPI includes the following fields:

• —form_type: "10-K" or "10-Q"
• —accession_number: Unique filing identifier
• —filing_date: Timestamp of the filing
• —quarter_ending: Fiscal quarter end date
• —company_name: Name of the reporting entity
• —text: Extracted paragraph from the filing
• —entities (list of extracted entities):
• —start_character / end_character: Position of the entity in the text
• —label: iXBRL-based tag (e.g., us-gaap:Revenues)
• —start_date_for_period / end_date_for_period: Time period of the financial figure
• —currency/unit: Currency (e.g., USD, EUR)
• —value: Extracted numerical figure

Dataset Statistics

Data Splits

• —HiFi-KPI (full dataset): Contains all extracted entities. See the GitHub Repository for an example of how to obtain granular labels.
• —The dataset includes calculationMasterTaxonomy.json and presentationMasterTaxonomy.json, which define the master hierarchical structure for the calculation and presentation layers.

Baselines and Benchmarks

We establish baselines using:

• —Text Classification: fine-tuning all-MiniLM-L6-v2 to classify entity labels from text snippets.
• —Sequence Labeling: fine-tuning BERT (bert-base-uncased) with a token classification head.

Text Classification Performance Across Hierarchical Collapsing Levels

Macro F1 Performance on 1000 Most Common Labels

Uses and Applications

HiFi-KPI can be used for:

• —Financial Information Extraction: Extracting key financial metrics for downstream applications.
• —XBRL-based Entity Mapping: Linking textual content to structured financial labels.
• —Document Understanding: Training models to interpret financial data.

Citation

If you use HiFi-KPI in your research, please cite:

@inproceedings{aavang-etal-2026-hifi,
    title = "{H}i{F}i-{KPI}: A Dataset for Hierarchical {KPI} Extraction from Earnings Filings",
    author = "Aavang, Rasmus T.  and
      Rizzi, Giovanni  and
      Tjalk-B{\o}ggild, Rasmus  and
      Iolov, Alexandre  and
      Zhang, Mike  and
      Bjerva, Johannes",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://aclanthology.org/2026.lrec-1.30/",
    doi = "10.63317/2nbsp7zzfb3g",
    pages = "441--455",
    abstract = "Accurate tagging of earnings reports can yield significant short-term returns for stakeholders. The machine-readable inline eXtensible Business Reporting Language (iXBRL) is mandated for public financial filings. Yet, its complex, fine-grained taxonomy limits the cross-company transferability of tagged Key Performance Indicators (KPIs). To address this, we introduce the Hierarchical Financial Key Performance Indicator (HiFi-KPI) dataset, a large-scale corpus of 1.65M paragraphs and 198k unique, hierarchically organized labels linked to iXBRL taxonomies. HiFi-KPI supports multiple tasks and we evaluate three: KPI classification, KPI extraction, and structured KPI extraction. For rapid evaluation, we also release HiFi-KPI-Lite, a manually curated 2.5K-instance subset. Baselines on HiFi-KPI-Lite show that encoder-based models achieve over 0.906 macro-F1 on classification, while Large Language Models (LLMs) reach 0.440 F1 on structured extraction. Finally, a qualitative analysis reveals that extraction errors primarily relate to dates. We open-source all code and data at Anonymous."
}

Access

More info is avalible at GitHub Repository