README.md

---
annotations_creators:
- expert-generated
- machine-generated
language:
- en
license: apache-2.0
multilinguality:
- monolingual
pretty_name: BUSTER Expanded NER
size_categories:
- 1K<n<10K
source_datasets:
- extended|expertai/BUSTER
task_categories:
- token-classification
task_ids:
- named-entity-recognition
tags:
- finance
- named-entity-recognition
- sec-filings
- bioes
configs:
- config_name: default
  data_files:
  - split: FOLD_1
    path: data/FOLD_1.jsonl.gz
  - split: FOLD_2
    path: data/FOLD_2.jsonl.gz
  - split: FOLD_3
    path: data/FOLD_3.jsonl.gz
  - split: FOLD_4
    path: data/FOLD_4.jsonl.gz
  - split: FOLD_5
    path: data/FOLD_5.jsonl.gz
---

# BUSTER Expanded NER

BUSTER Expanded NER is a derived annotation layer over the 3,779 manually
annotated English documents in the gold corpus of
[expertai/BUSTER](https://huggingface.co/datasets/expertai/BUSTER). It replaces
BUSTER's transaction-role ontology with four general named-entity labels and
materially expands mention coverage within each document.

The release is designed as a tokenizer-neutral source for training flat NER
models. It stores exact character spans rather than tokenizer-specific tags, so
BIOES labels can be produced after choosing a tokenizer.

> **Annotation scope:** this version annotates safe repeated occurrences of
> already identified entity surfaces and adds a precision-first set of omitted
> entities. It is substantially more complete than the original transaction-role
> layer, but it is **not guaranteed to be exhaustive**. An unlabelled phrase is
> not always a reliable negative.

## What changed from BUSTER

- The five original company roles (buyer, seller, acquired company, legal
  advisor, and generic advisor) are mapped to `organization` by default.
- Reviewed investment funds and investment firms are distinguished from other
  organizations with the `fund` label.
- Reviewed `person`, omitted `organization`/`fund`, and named `event` mentions
  are added.
- Once an entity surface is known in a document, safe non-overlapping
  occurrences of that surface are projected across the document. Matching is
  case-aware, tolerates whitespace variation, and filters URL, domain,
  identifier, overlap, and known-longer-name cases.
- BUSTER's `Generic_Info.ANNUAL_REVENUES` label is outside this release's NER
  ontology and is not included.
- The original manually annotated folds are retained. The automatically
  annotated BUSTER silver corpus is not included.

These changes create a different task from the original BUSTER benchmark.
Scores on this dataset should not be compared directly with scores reported for
BUSTER's transaction-role ontology.

## Labels

| Label | Meaning |
|---|---|
| `person` | A named person. |
| `organization` | A company, university, agency, NGO, accelerator, association, or other organization. |
| `fund` | An investment fund or investment firm, including reviewed venture-capital and private-equity entities. This more specific label takes precedence over `organization`. |
| `event` | A named gathering at which people or organizations attend, meet, pitch, present, or exhibit. |

## Dataset statistics

Counts below are entity **mentions**, not unique names.

| Split | Documents | Organization | Fund | Person | Event | All mentions |
|---|---:|---:|---:|---:|---:|---:|
| `FOLD_1` | 753 | 12,913 | 368 | 879 | 8 | 14,168 |
| `FOLD_2` | 759 | 13,250 | 324 | 808 | 13 | 14,395 |
| `FOLD_3` | 758 | 12,696 | 311 | 876 | 7 | 13,890 |
| `FOLD_4` | 755 | 13,116 | 346 | 955 | 1 | 14,418 |
| `FOLD_5` | 754 | 12,814 | 431 | 910 | 3 | 14,158 |
| **Total** | **3,779** | **64,789** | **1,780** | **4,428** | **32** | **71,029** |

The five folds are inherited cross-validation partitions, not a prescribed
train/validation/test division. For example, use one fold for evaluation and
train on the other four.

## Loading the dataset

Replace the repository name below after uploading:

```python
from datasets import concatenate_datasets, load_dataset

dataset = load_dataset("YOUR_USERNAME/buster-expanded-ner")

held_out = "FOLD_5"
train = concatenate_datasets(
    dataset[fold] for fold in dataset if fold != held_out
)
validation = dataset[held_out]
```

Each compressed file is newline-delimited JSON and can also be read directly:

```python
import gzip
import json

with gzip.open("data/FOLD_1.jsonl.gz", "rt", encoding="utf-8") as handle:
    first_document = json.loads(next(handle))
```

## Data format

Each row contains one complete document:

```json
{
  "schema_version": 1,
  "dataset": "buster",
  "document_id": "...",
  "split": "FOLD_1",
  "language": "en",
  "text": "Alice met Acme Ventures.",
  "entities": [
    {
      "start": 0,
      "end": 5,
      "label": "person",
      "text": "Alice",
      "source_label": "MODEL_REVIEWED_TEXT_MENTION",
      "annotation_source": "review_config:buster_person_additions.json",
      "override": null,
      "override_origin": null,
      "span_addition_origin": null,
      "review_confirmation": null
    },
    {
      "start": 10,
      "end": 23,
      "label": "fund",
      "text": "Acme Ventures",
      "source_label": "Parties.BUYING_COMPANY",
      "annotation_source": null,
      "override": "fund",
      "override_origin": "review_fragment:...",
      "span_addition_origin": null,
      "review_confirmation": null
    }
  ],
  "unresolved_entities": [],
  "training_ready": true,
  "metadata": {
    "source_dataset": "expertai/BUSTER",
    "annotation_scope": "..."
  }
}
```

Offsets use the half-open convention: `text[start:end]`. Entities are sorted by
offset, do not overlap, and satisfy `text[start:end] == entity["text"]`.
Provenance fields on an entity are nullable when they do not apply. The upload
files make those optional keys explicit so the Hugging Face JSON loader infers
a stable `list<struct>` schema; all canonical values are preserved.

## Producing BIOES tags

BIOES tags are deliberately not stored because they depend on tokenization.
For a chosen fast tokenizer:

1. tokenize `text` with `return_offsets_mapping=True`;
2. align each entity's `[start, end)` span to complete model tokens;
3. assign `S-label` to a one-token span or `B-label`, `I-label`, and `E-label`
   to a multi-token span;
4. keep non-entity tokens as `O` and mask special tokens and padding from the
   loss; and
5. split long documents into windows without cutting an entity span.

Reject or explicitly handle any entity that does not align to complete tokens.
Do not silently turn a truncated or partially aligned entity into `O`.

## Annotation process

The original BUSTER BIO annotations were aligned back to the unmodified source
text and converted into character spans. Fund candidates and missing entities
were reviewed in versioned batches. Person additions use strong name, honorific,
or role evidence, and named-event additions use participation, presentation,
speaker, venue, or date evidence. The reviews favor precision over recall.

After reviewed changes are applied, known surfaces are projected only within
their source document. Projection never crosses documents and never overwrites
an existing overlapping span. The canonical build reports 45,800 added
same-document surface occurrences and no unresolved alignments in this release.
All 3,779 rows have `training_ready: true`.

The augmentation includes model-assisted review followed by deterministic
validation. It should not be interpreted as a new, fully human-annotated gold
standard.

## Intended uses

This dataset is suitable for:

- training or cross-validating finance-domain flat NER models;
- deriving tokenizer-specific BIO/BIOES supervision;
- studying organization-versus-fund classification; and
- pretraining before fine-tuning on a smaller, exhaustively annotated corpus.

It should not be used as a drop-in replacement for the original BUSTER task or
as evidence that unannotated text contains no entity.

## Limitations and responsible use

- Omitted people, organizations, funds, and especially events can still remain
  in the text. The 32 event mentions are too sparse for broad event coverage.
- Surface projection improves recall for known names but can propagate an
  incorrect source or review decision to repeated occurrences.
- Fund-versus-organization boundaries can be context-dependent and reflect the
  review policy used for this release.
- Noise in the original BUSTER spans and source documents may remain.
- The source consists of English financial-transaction documents collected
  from SEC EDGAR, so models may not generalize to other languages, periods,
  jurisdictions, industries, or document genres.
- Documents can be long (up to about 9,900 characters in this release), so
  fixed-context models require a span-safe chunking strategy.
- The documents are public filings but can contain names and business contact
  details. Users remain responsible for appropriate handling and downstream
  use.

## License

This derived release follows the upstream BUSTER dataset's Apache License 2.0.
See [LICENSE](LICENSE) and [NOTICE](NOTICE). Users should also review the
[upstream dataset card](https://huggingface.co/datasets/expertai/BUSTER).

## Citation

Please cite the original BUSTER paper when using this dataset:

```bibtex
@inproceedings{zugarini-etal-2023-buster,
  title = {{BUSTER}: a {``}{BUS}iness Transaction Entity Recognition{''} dataset},
  author = {Zugarini, Andrea and Zamai, Andrew and Ernandes, Marco and Rigutini, Leonardo},
  booktitle = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  month = dec,
  year = {2023},
  address = {Singapore},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2023.emnlp-industry.57},
  doi = {10.18653/v1/2023.emnlp-industry.57},
  pages = {605--611}
}
```

When the uploaded dataset has a stable owner, repository URL, and release tag,
add a citation for this derived annotation release as well.