Note We are continuing to update and refine our methodology as we develop our tool and its capabilities. This is the methodology for our approach to detecting concepts in the text of documents. For information about the research methods and taxonomies used to identify references to specific concepts, please refer to our concept-specific methodologies.
The text labels and filter tools in our Global Stocktake search tool are based on data collected using Explorer, an internal text analysis tools developed by Climate Policy Radar. Explorer enables search for references to a number of concepts simultaneously across a large number of documents, using keywords. This section documents how Explorer works, and the methodology we follow to develop keyword lists to identify references to concepts within documents related to the Global Stocktake.
To search a set of documents for references to a concept using Explorer, we create bespoke CSV input files for each concept, containing keywords organised across a three-tier hierarchy. For example, for fossil fuels:
|Tier 1 concept||Tier 2 concept||Tier 3 concept|
|Fossil fuels||Oil||Crude oil|
|Fossil fuels||Oil||Petroleum oil|
|Fossil fuels||Gas||Natural gas|
|Fossil fuels||Gas||Petroleum gas|
In our input files, we indicate where a certain keyword may have multiple synonyms or different linguistic expressions, such a present/past tense and singular/plural. These variations are then searched for automatically. This enables us to perform a precise and comprehensive search for each concept without exhaustively listing all its possible variations or linguistic expressions.
There are two concepts for which our method is different: sectors and policy instruments. For these, as a mention of each doesn’t necessary correspond to a mention of a single keyword or phrase within a paragraph, we train machine learning models to classify paragraphs as referring to zero or more sectors and/or policy instruments (see Section 4).
Wherever possible, we use existing taxonomies developed by expert third-party sources as the basis for our keyword lists. For example, our keyword list for technologies is based on the UN Climate Technology Centre and Network’s Taxonomy of Climate Technologies, while our keyword list for climate-related hazards is based on the UN Office for Disaster Risk Reduction’s Hazard Definition and Classification Review.
Where such resources are available, we take the following steps to develop a keyword list:
Where we are not able to identify an existing credible source providing a taxonomy for a given concept, we create keyword lists based on desk research.
To do this, we review literature from sources such as Wikidata, as well as academic journals and reports and publications by international agencies, to identify the relevant key terms for a given concept.
We then create keyword lists following the process described in Section 2 above.
For two concepts – sectors and policy instruments – we use machine learning classifiers to assign concepts to paragraphs. This is because a mention of each doesn’t necessary correspond to a mention of a single keyword or phrase within a paragraph.
Each of these classifiers is trained on examples of paragraphs that are assigned to each concept. Similarly to the process for developing keyword lists, we label these examples according to existing taxonomies, details of which can be found in the concept-specific methodologies.
The machine learning approach we use, SetFit, achieves high accuracy with little labelled data. We test the performance of our classifiers and publish performance statistics in each of the concept-specific methodologies that use a classifier.
We are looking to open source our data, including the labelled examples used to train our classifiers, in the near future. Until then, please contact us for access to this data.
When you search using the filters on the left side of the page, phrases in the text of a document that represent or express a concept will be tagged and labelled. When a machine learning classifier has been used, the tags appear below each paragraph instead.
Use these labels to find instances of your chosen concept, and read the surrounding text to understand its context in the document.