metadata

language:
  - en
tags:
  - climate
  - policy
  - legal
size_categories:
  - 1M<n<10M
license: cc

Global Stocktake Open Data

This repo contains the data for the first UNFCCC Global Stocktake. The data consists of document metadata from sources relevant to the Global Stocktake process, as well as full text parsed from the majority of the documents.

The files in this dataset are as follows:

metadata.csv: a CSV containing document metadata for each document we have collected. This metadata may not be the same as what's stored in the source databases – we have cleaned and added metadata where it's corrupted or missing.
full_text.parquet: a parquet file containing the full text of each document we have parsed. Each row is a text block (paragraph) with all the associated text block and document metadata.

A research tool you can use to view this data and the results of some classifiers run on it is at gst1.org.

This data is licensed according to CC BY 4.0, which is a license that represents the terms at the source repositories.

Contents

Sources and data completeness
Field descriptions
Known issues
Usage in Python
- Loading metadata CSV
- Loading text block data

Sources and data completeness

This dataset contains documents from the following sources:

The following Global Stocktake relevant data sources are not yet in this dataset:

Data completeness

The last refresh of the data was on 2023-07-25.

We currently only parse text out of PDFs. Any non-PDF file will only be referenced in metadata.csv, and not be referenced in full_text.parquet.

We have yet to process approximately 150 documents of the 1700 documents due to formatting issues. We are working on resolving this issue as soon as possible. See the document list here.

Data model

This dataset contains individual documents that are grouped into 'document families'.

The way to think of is as follows:

Each row in the dataset is a physical document. A physical document is a single document, in any format.
All physical documents belong to document families. A document family is one or more physical documents, centred around a main document, which jointly contain all relevant information about the main document. For example, where a document has a translation, amendments or annexes, those files are stored together as a family.

Getting unique text blocks

TODO

Field descriptions

author: document author (str)
author_is_party: whether the author is a Party (national government) or not (bool)
block_index: the index of a text block in a document. Starts from 0 (int)
coords: coordinates of the text block on the page
date: publication date of the document
document_content_type: file type. We have only parsed text from PDFs.
document_id: unique identifier for a document
document_family_id: see data model section above
document_family_slug: see data model section above
document_md5_sum: md5sum of the document's content
document_name: document title
document_source_url: URL for document
document_variant: used to identify translations. In [nan, 'Translation', 'Original Language']
has_valid_text: our heuristic about whether text is valid or not in the document based on the parser
language: language of the text block. Either en or nan - see known issues
page_number: page number of text block (0-indexed)
text: text in text block
text_block_id: identifier for a text block which is unique per document
translated: whether we have machine-translated the document to English. Where we have translated documents, both the original and translated exist.
type: type of text block. In ["Text", "Title", "List", "Table", "Figure","Ambiguous"]
type_confidence: confidence from that the text block is of the labelled type
types: list of document types e.g. Nationally Determined Contribution, National Adaptation Plan (list[str])
version: in ['MAIN', 'ANNEX', 'SUMMARY', 'AMENDMENT', 'SUPPORTING DOCUMENTATION', 'PREVIOUS VERSION']

Known issues

Author names are sometimes corrupted
Text block languages are sometimes missing or marked as nan

Usage in Python

The easiest way to access this data via the terminal is to run git clone <this-url>.

Loading metadata CSV

metadata = pd.read_csv("metadata.csv")

Loading text block data

Once loaded into a Huggingface Dataset or Pandas DataFrame object the parquet file can be converted to other formats, e.g. Excel, CSV or JSON.

# Using huggingface (easiest)
dataset = load_dataset("ClimatePolicyRadar/global-stocktake-documents")

# Using pandas
text_blocks = pd.read_parquet("full_text.parquet")