rafmacalaba's picture
Add dataset_mentions (gliner_agents_v3) and refresh tables
8db66b4 verified
|
Raw History Blame Contribute Delete
6.69 kB
---
pretty_name: SCD Indicator Inventory
license: other
language: [en]
tags: [world-bank, systematic-country-diagnostic, indicators, ner, gliner, development-economics]
size_categories: [10K<n<100K]
configs:
- config_name: documents
data_files: data/documents.parquet
- config_name: exhibits
data_files: data/exhibits.parquet
default: true
- config_name: indicators
data_files: data/indicators.parquet
- config_name: concepts
data_files: data/concepts.parquet
- config_name: data_gaps
data_files: data/data_gaps.parquet
- config_name: paragraphs
data_files: data/paragraphs.parquet
- config_name: data360_catalog
data_files: data/data360_catalog.parquet
- config_name: dataset_mentions
data_files: data/dataset_mentions.parquet
---
# SCD Indicator Inventory
Which indicators do World Bank **Systematic Country Diagnostics (SCDs)** use, where do the data
come from, how are they shown, and what data gaps do SCD teams report? This dataset holds the
corpus metadata and every intermediate and final output of the extraction pipeline, so each
number in the analysis can be traced back to a page of an SCD.
Built 09 November 2026. **Preliminary**: no hand-labelled validation. Reliability is assessed by
agreement between three independent extractors (see `concepts`, `extractor` column).
## Tables (configs)
| config | rows | one row per |
|---|---|---|
| `documents` | 206 | SCD document record (API); `is_primary` marks the 165 documents analysed |
| `exhibits` | 10,185 | figure, table, map or box (located by its "Source:" note) |
| `indicators` | 42,174 | normalized indicator mention × extractor (`gliner2`, `gliner1`, `llm_direct`) |
| `concepts` | 13,087 | indicator concept × extractor, with SCD counts and tier |
| `data_gaps` | 3,874 | candidate data-gap sentence, with model scores and LLM verdict |
| `paragraphs` | 68,327 | narrative text block (for search and context) |
| `dataset_mentions` | 11,568 | named data source found in the text or exhibit sources (model span, canonical `dataset`, `group`, `list_db`, score, passage) |
| `data360_catalog` | 9,850 | Data360 indicator (id, name, database, definition, unit, topics) from the public Data360 API |
Join on `doc_id`, and on (`doc_id`, `exhibit_seq`) between exhibits and indicators.
## What is readily available vs. what is extracted
| Information | Where it comes from | Column prefix / columns |
|---|---|---|
| Title, document type, report no., date, disclosure status, author, country, region, language, report & PDF links | **WBG Documents & Reports API** (ready-made) | `documents`: `title`… `pdf_url` |
| Abstract, topics/subtopics, keywords, owning unit, project ID | **API** (ready-made, partial coverage: 59–79%) | `documents.api_*` |
| PDF title/author/subject/keywords properties | **PDF file** (ready-made but mostly generic: "World Bank Document", "World Bank Group"; subject/keywords <2%) | `documents.pdf_*` |
| Bookmarks / outline | **PDF file** (41% of SCDs) | `documents.pdf_has_outline` |
| Exhibit captions, labels, source & note lines, in-exhibit text, page | **PDF layout parsing** | `exhibits`: `caption`, `source_text`, `note_text`, `exhibit_text` |
| Presentation (chart / table / map / box) | PDF layout (caption label) | `exhibits.presentation` |
| Source / database, SCD-recommended database cited, WDI cited | **rules** on the source line | `exhibits.rules_source_*`, `rules_scd_list_databases`, `rules_cites_wdi` |
| Time dimension (trend / two points / latest value) | **rules** (years in caption & axis labels) | `exhibits.rules_time_dimension` |
| Benchmarking (peers, region, income group, within-country) | **rules** + WB country/aggregate gazetteer | `exhibits.rules_benchmark`, `rules_comparators` |
| Indicator names | **GLiNER2** NER (primary) and **GLiNER v1** NER (comparison): verbatim spans | `exhibits.gliner2_spans`, `gliner1_spans` |
| Theme | **GLiNER2** classification | `exhibits.gliner2_theme` |
| Standardized indicator, concept, WDI series | **LLM normalization** of the NER spans (never adds indicators) + WDI catalog matching | `indicators.canonical`, `concept`, `wdi_code` |
| Datasets named in the text and exhibit sources | **`rafmacalaba/gliner_agents_v3`** (GLiNER fine-tuned for data-use mentions; NAMED_DATA / DESCRIPTIVE_DATA at score >= 0.6) + rule-based spelling merge | `dataset_mentions` |
| Sector of each indicator | **LLM** classification of the normalized indicator name into 17 themes | `concepts.sector` |
| Availability in WDI / Data360 | TF-IDF retrieval over the Data360 catalog + **LLM** choice of the same-concept series (or none) | `concepts.available_in`, `d360_*` |
| Comparators named in exhibits | **rules** + country/aggregate list (a bare "World"/"Africa" is no longer matched) | `exhibits.rules_comparators`, `rules_benchmark` |
| Data gaps | GLiNER2 candidates → **LLM verification**, type, topic, data type | `data_gaps` |
| LLM reading exhibits directly | **comparison only**, not used in the inventory | `*.llm_direct_*`, `indicators[extractor=llm_direct]` |
## Method in brief
1. Corpus: all records with document type "Systematic Country Diagnostic" in the WBG Documents & Reports API
(283) → English full SCDs, SCD Updates and Streamlined CDs; duplicates, volumes, summaries, background notes
and information notices removed → one primary document per SCD edition.
2. Parsing (PyMuPDF): exhibits anchored on "Source:" notes; caption, notes and in-exhibit text attached by layout.
3. Extraction: rules for source/time/benchmarking; GLiNER2 for indicator spans and theme.
4. Normalization: an LLM (WBG mAIFactory, `claude-sonnet-5-5`) maps each exhibit's spans to standardized names,
then to WDI series (1,498-series catalog) and concepts; synonyms consolidated.
5. Counting: a concept is counted once per SCD. Tiers: core ≥50% of SCDs and ≥5 of 7 regions; frequent ≥25%;
occasional ≥10%.
## Limitations
- No hand-labelled validation; extractor agreement is the reliability evidence.
- Rule-based time dimension is undetectable for many exhibits ("unspecified"); within-country breakdowns are
under-detected by rules.
- Only indicators shown in exhibits are inventoried; narrative-only mentions are in `paragraphs` but not counted.
- `concept_aligned` (cross-extractor alignment) is coarser than `concept` and has known over-merges
(e.g. "Employment by type" → "Employment by sector").
## Source documents
SCDs are published by the World Bank on documents.worldbank.org. This dataset contains links, extracted
snippets (captions, source lines, short sentences, paragraphs) and derived annotations, not the PDFs.
Check each document's terms of use before reuse.