---
license: other
language:
- en
pretty_name: Database of Data Use
task_categories:
- token-classification
tags:
- data-use
- dataset-mentions
- information-extraction
configs:
- config_name: documents
  data_files: documents/*.parquet
- config_name: mentions
  data_files: mentions/*.parquet
- config_name: datasets
  data_files: datasets/*.parquet
- config_name: mention_family
  data_files: mention_family/*.parquet
- config_name: families
  data_files: registry/families.parquet
- config_name: aliases
  data_files: registry/aliases.parquet
---
# Database of Data Use

Mentions of datasets in documents, extracted with the `datause_pipeline`
(`rafmacalaba/gliner_agents_v3` finds mentions, `rafmacalaba/lfm2.5-350M-focus-metadata` gives usage and attributes;
code: https://github.com/rafmacalaba/database-of-data-use).

The dataset is the data-use extractions and the **origin** of each document, for traceability. It carries no
bibliographic metadata (title, authors, abstract); follow `source_url` / `pdf_url` / `doi` to the origin.

The data grows by **source**. Every table has a `source` column, and each source is one parquet file per table
(`documents/<source>.parquet`, ...), so adding a source adds files and does not change existing ones.
All sources share the same schema; a field a source does not have is null.

| table | one row per | main columns |
|---|---|---|
| `documents` | document | `corpus_id`, `source`, `source_doc_id`, `source_url`, `pdf_url`, `doi`, `sha256`, `n_pages`, `extraction_status`, plus `document_type` and `publication_year` as analysis context |
| `mentions` | extracted mention | `mention_id`, `corpus_id`, `page_number` (1-based), `mention_text`, `specificity`, `score`, `role`, `usage_action`, `usage_summary`, `theme_tags`, `dataset_geography`, `dataset_year`, `dataset_version`, `dataset_producer`, `dataset_acronym`, `dataset_data_type` `document_type`, `publication_year`, `source_url` |
| `datasets` | distinct dataset in a document | `corpus_id`, `dataset_name`, `specificity`, `n_occurrences`, `pages`, `max_score`, `role`, `usage_action`, `theme_tags`, `dataset_*` |

| `mention_family` | mention | `mention_id`, `specificity`, `family_id`, `match_method`, `match_score`, `mention_country` (ISO-3), `round_year` |
| `families` | dataset family | `family_id`, `canonical_name`, `acronym`, `kind`, `producer`, `n_mentions` |
| `aliases` | normalised name form | `alias`, `family_id`, `method`, `score` |

Join the tables on `corpus_id` (`<source>:<sha256[:16]>` of the document bytes).

## Reading the data

- `specificity`: `named` (a named dataset), `descriptive` (described, not named), `vague`. Only named mentions carry
  `role` (primary, supporting, background), `usage_action` (source, analyze, curate, validate, inform) and the
  `dataset_*` attributes; for the others they are null.
- `mentions` are per text chunk: the same mention in overlapping chunks can appear twice. `datasets` is the
  deduplicated view.
- `extraction_status != ok` means the document was **not processed** (for example the GPU ran out of memory). It does
  not mean the document mentions no data. Such documents have no mentions.
- The extraction is automatic and has errors: scores and attributes are model output, not curated.
- Byte-identical files are merged into one document; the other origin ids are in `duplicate_source_doc_ids`.

## Dataset families (`family_id`)

"Ghana Living Standards Survey", "Nepal Living Standards Survey IV" and "LSMS 2017" are different strings for one
family of surveys. `mention_family` maps each *named* mention to a `family_id` automatically (no manual curation) and
splits off the country and the round/year. `match_method` says how: `alias` (a form seen before), `acronym`, `fuzzy`
(overlap with a family's name), `extends` (a variant that contains a family's whole name), `new` (first form of a new
family) or `generic` (nothing distinguishing, like "household survey" or "census data"; `family_id` is null).
Named mentions create and join families. Descriptive mentions (if included) only join an *existing* family, with a stricter
rule (a known form, an acronym, or a close match); most stay unresolved (`unmatched`), because descriptions like "passenger
data" are not dataset names. Named results do not depend on whether descriptive mentions are included.
`match_score` is the confidence; low `fuzzy`/`extends` scores are the least reliable. `family_id` is assigned once and
never changes, so adding sources only extends the registry. `canonical_name`, `kind` (majority of the extracted data
type) and `producer` are automatic and may be imperfect.

## Sources

See `summary-<source>.json` in the repository for counts per source.