Kleister NDA — Information Extraction (orgrctera/kleister_nda_information_extraction)

Overview

This release packages the Kleister NDA split of the Kleister benchmark as rows suitable for information extraction (IE) evaluation. Each example points at a Non-Disclosure Agreement (NDA) document and specifies which attribute keys should be filled; the target is a JSON object of normalized string values for those keys (with null when a value is absent or not applicable).

Kleister NDA was introduced to stress Key Information Extraction (KIE) on long, layout-rich English documents: a mix of scanned and born-digital PDFs where systems must use text, structure, and (optionally) page layout—not only flat OCR text. The original collection comprises 540 NDAs, 3,229 unique pages, and 2,160 annotated entities at the benchmark level (Stanisławek et al., 2021).

This Hub dataset is aligned with internal Langfuse export metadata: benchmark kleister, sub-benchmark nda, task type information_extraction, 540 total items across three splits.

Task

• Task type: Information extraction (also described as key information extraction / slot filling in the Kleister literature). You are given which fields to predict (input.keys); the goal is to output the field values, not spans or mentions (this is not a classic NER tagging task).
• Inputs: Document identity (input.filename) and the space-separated list of keys to extract for that instance (input.keys).
• Supervision: expected_output is a JSON string whose keys are effective_date, jurisdiction, party, and term. Values use the Kleister-style normalized tokenization (e.g. underscores in multi-word strings; null when unknown or not applicable for that row). Some keys in input.keys may be decoys relative to what is scored for a given row—systems should only commit values they can justify from the document (see the original task description on the upstream repository).

Evaluation note (original benchmark): The Kleister NDA leaderboard setup uses F1 over normalized values (with auxiliary true-cased F1 in the paper’s baselines). Consult the upstream Kleister NDA repository and GEval tooling if you need to reproduce official scores on the raw TSV release.

Background

Why Kleister?

KIE is central to enterprise NLP, but public benchmarks historically under-represent messy, long, real-world documents. Kleister was created to narrow that gap with corpora where both lexical and layout / OCR effects matter.

Kleister NDA documents

NDAs are long formal contracts. The corpus mixes scanned pages and clean digital PDFs; OCR noise, headers/footers, and multi-column layouts make extraction non-trivial. The original work reports strong Transformer and layout-aware baselines (e.g. BERT-family and LayoutLM-style models) yet leaves headroom versus the best published systems.

Splits and labels

Split	Rows	Gold in expected_output
train	254	Yes
dev	83	Yes
test	203	Withheld (empty string in this export; suitable for blind evaluation)

Total: 540 rows.

Data fields

Column	Type	Description
id	string	Stable UUID for the row.
input.filename	string	PDF filename keying into the original Kleister document bundle (documents are not hosted in this parquet-only Hub snapshot).
input.keys	string	Space-separated attribute names to treat as requested keys for this instance (e.g. effective_date jurisdiction party term).
expected_output	string	JSON string with keys effective_date, jurisdiction, party, term. Empty on test in this release.
metadata.item	string	JSON string echoing keys, expected object, and filename for traceability.

Examples

Example 1 — train (all four keys requested)

• input.filename: 7496116e8680dac321f36147b6312411.pdf
• input.keys: effective_date jurisdiction party term
• expected_output (parsed):

{
  "term": "18_months",
  "party": "Open_Text_Corporation",
  "jurisdiction": "California",
  "effective_date": "2014-07-24"
}

Example 2 — dev (partial / nullable fields)

• input.filename: af344c9a1d0fc128bcab1737a6b7d0ec.pdf
• input.keys: effective_date jurisdiction party term
• expected_output (parsed):

{
  "term": null,
  "party": "Bernd_Brust",
  "jurisdiction": "California",
  "effective_date": "2011-11-01"
}

Example 3 — dev (narrow key set)

• input.keys: jurisdiction
• expected_output (parsed) still carries the full schema object; only the requested keys are the focus of the instance:

{
  "term": null,
  "party": "Oceanfinancial_Corp.",
  "jurisdiction": "New_Jersey",
  "effective_date": null
}

References

1. Stanisławek, T., Graliński, F., Wróblewska, A., Lipiński, D., Kaliska, A., Rosalska, P., Topolski, B., & Biecek, P. (2021). Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts. ICDAR 2021.

   • arXiv: 2105.05796 — Abstract (verbatim): “The relevance of the Key Information Extraction (KIE) task is increasingly important in natural language processing problems. But there are still only a few well-defined problems that serve as benchmarks for solutions in this area. To bridge this gap, we introduce two new datasets (Kleister NDA and Kleister Charity). They involve a mix of scanned and born-digital long formal English-language documents. In these datasets, an NLP system is expected to find or infer various types of entities by employing both textual and structural layout features. The Kleister Charity dataset consists of 2,788 annual financial reports of charity organizations, with 61,643 unique pages and 21,612 entities to extract. The Kleister NDA dataset has 540 Non-disclosure Agreements, with 3,229 unique pages and 2,160 entities to extract. We provide several state-of-the-art baseline systems from the KIE domain (Flair, BERT, RoBERTa, LayoutLM, LAMBERT), which show that our datasets pose a strong challenge to existing models. The best model achieved an 81.77% and an 83.57% F1-score on respectively the Kleister NDA and the Kleister Charity datasets. We share the datasets to encourage progress on more in-depth and complex information extraction tasks.”
   • Springer chapter DOI: 10.1007/978-3-030-86549-8_36

2. Original task & data layout (TSV, PDFs, GEval): applicaai/kleister-nda on GitHub.

3. Related benchmark: Kleister Charity (financial reports of charities), released in the same paper, with larger page/entity counts.

Citation

@inproceedings{stanislawek2021kleister,
  title     = {Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts},
  author    = {Stanis{\l}awek, Tomasz and Grali{\'n}ski, Filip and Wr{\'o}blewska, Anna and Lipi{\'n}ski, Dawid and Kaliska, Agnieszka and Rosalska, Paulina and Topolski, Bartosz and Biecek, Przemys{\l}aw},
  booktitle = {International Conference on Document Analysis and Recognition (ICDAR)},
  year      = {2021},
  doi       = {10.1007/978-3-030-86549-8_36}
}

Also cite arXiv 2105.05796 if you rely on the open PDF/preprint version.