Wiki-Quantities and Wiki-Measurements: Datasets of Quantities and their Measurement Context from Wikipedia

The task of measurement extraction is typically approached in a pipeline manner, where 1) quantities are identified before 2) their individual measurement context is extracted (see our review paper). To support the development and evaluation of systems for measurement extraction, we present two large datasets that correspond to the two tasks:

Wiki-Quantities, a dataset for identifying quantities, and

Wiki-Measurements, a dataset for extracting measurement context for given quantities.

The datasets are heuristically generated from Wikipedia articles and Wikidata facts. For a detailed description of the datasets, please refer to the upcoming corresponding paper: 

Wiki-Quantities and Wiki-Measurements: Datasets of Quantities and their Measurement Context from Wikipedia. 2025. Jan Göpfert, Patrick Kuckertz, Jann M. Weinand, and Detlef Stolten.

Versions

The datasets are released in different versions:

Processing level: the pre-processed versions can be used directly for training and evaluating models, while the raw versions can be used to create custom pre-processed versions or for other purposes. Wiki-Quantities is pre-processed for IOB sequence labeling, while Wiki-Measurements is pre-processed for SQuAD-style generative question answering.

Filtering level: 

Wiki-Quantities is available in a raw, large, small, and tiny version: The raw version is the original version, which includes all the samples originally obtained. In the large version, all duplicates and near duplicates present in the raw version are removed. The small and tiny versions are subsets of the large version which are additionally filtered to balance the data with respect to units, properties, and topics.

Wiki-Measurements is available in a large`, small, large_strict, small_strict, small_context, and large_strict_context version: The large version contains all examples minus a few duplicates. The small version is a subset of the large version with very similar examples removed. In the context versions, additional sentences are added around the annotated sentence. In the strict versions, the quantitative facts are more strictly aligned with the text.

Quality: all data has been automatically annotated using heuristics. In contrast to the silver data, the gold data has been manually curated.

Format

The datasets are stored in JSON format. The pre-processed versions are formatted for direct use for IOB sequence labeling or SQuAD-style generative question answering in NLP frameworks such as Huggingface Transformers. In the not pre-processed versions of the datasets, annotations are visualized using emojis to facilitate curation. For example:

Wiki-Quantities (only quantities annotated):

"In a 🍏100-gram🍏 reference amount, almonds supply 🍏579 kilocalories🍏 of food energy."

"Extreme heat waves can raise readings to around and slightly above 🍏38 °C🍏, and arctic blasts can drop lows to 🍏−23 °C to 0 °F🍏."

"This sail added another 🍏0.5 kn🍏."

 Wiki-Measurements (measurement context for a single quantity; qualifiers and quantity modifiers are only sparsely annotated):

"The 🔭French national census🔭 of 📆2018📆 estimated the 🍊population🍊 of 🌶️Metz🌶️ to be 🍐116,581🍐, while the population of Metz metropolitan area was about 368,000."

"The 🍊surface temperature🍊 of 🌶️Triton🌶️ was 🔭recorded by Voyager 2🔭 as 🍐-235🍐 🍓°C🍓 (-391 °F)."

"🙋The Babylonians🙋 were able to find that the 🍊value🍊 of 🌶️pi🌶️ was ☎️slightly greater than☎️ 🍐3🍐, by simply 🔭making a big circle and then sticking a piece of rope onto the circumference and the diameter, taking note of their distances, and then dividing the circumference by the diameter🔭."

The mapping of annotation types to emojis is as follows:

Basic quantitative statement:

Entity: 🌶️

Property: 🍊

Quantity: 🍏

Value: 🍐

Unit: 🍓

Quantity modifier: ☎️

Qualifier:

Temporal scope: 📆

Start time: ⏱️

End time: ⏰️

Location: 📍

Reference: 🙋

Determination method: 🔭

Criterion used: 📏

Applies to part: 🦵

Scope: 🔎

Some qualifier: 🛁

Note that for each version of Wiki-Measurements sample IDs are randomly assigned. Therefore, they are not consistent, e.g., between silver small and silver large. The proportions of train, dev, and test sets are unusual because Wiki-Quantities and Wiki-Measurements are intended to be used in conjunction with other non-heuristically generated data.

Evaluation

The evaluation directories contain the manually validated random samples used for evaluation. The evaluation is based on the large versions of the datasets. Manual validation of 100 samples each of Wiki-Quantities and Wiki-Measurements showed that 100% of the Wiki-Quantities samples and 94% (or 84% if strictly scored) of the Wiki-Measurements samples were correct. 

License

In accordance with Wikipedia's and Wikidata's licensing terms, the datasets are released under the CC BY-SA 4.0 license, except for Wikidata facts in ./Wiki-Measurements/raw/additional_data.json, which are released under the CC0 1.0 license (the texts are still CC BY-SA 4.0).

About Us 

We are the Institute of Climate and Energy Systems (ICE) - Jülich Systems Analysis belonging to the Forschungszentrum Jülich. Our interdisciplinary department's research is focusing on energy-related process and systems analyses. Data searches and system simulations are used to determine energy and mass balances, as well as to evaluate performance, emissions and costs of energy systems. The results are used for performing comparative assessment studies between the various systems. Our current priorities include the development of energy strategies, in accordance with the German Federal Government’s greenhouse gas reduction targets, by designing new infrastructures for sustainable and secure energy supply chains and by conducting cost analysis studies for integrating new technologies into future energy market frameworks.

Acknowledgements

The authors would like to thank the German Federal Government, the German State Governments, and the Joint Science Conference (GWK) for their funding and support as part of the NFDI4Ing consortium. Funded by the German Research Foundation (DFG) – project number: 442146713. Furthermore, this work was supported by the Helmholtz Association under the program "Energy System Design".