We created an AMR dataset of 210,000 sentences (203,000 train, 7,000 dev) from sentences drawn mainly from the English Wikipedia. The exact process is described in our paper (see below).
Please note: this data has not been validated manually. If you find any inconsistencies, please let us know or submit a pull request.
Due to a problem when extracting the sentences from Wikipedia and Wikinews, the sentence IDs (which point to the wikidata item of the page) are sometimes incorrect. I.e. a sentences with a Wikipedia ID can be in reality taken from Wikinews.
In a nutshell, we take sentences from Wikipedia pages and filter them in order to avoid processing sentence which are likely too difficult to parse.
We discard sentences:
- which contain non-Latin characters, brackets or braces, and which lack a final punctuation mark.
- which are too short
- which contain long sequences of digits (like ISBN numbers, etc).
- duplicates
We then parse the candidate sentences using four different parsers and compare the resulting graphs (using Smatch). If they are too different, we discard the sentence. Else we check whether the most central graph (the graph which compares best with all the other three graphs) is formally correct:
- whether all verbal and adjectival concepts (i.e. those with numerical suffixes like
-01) are in fact defined in PropBank, and that all their outgoing relations ARGn in the graph are also defined for the given concept in PropBank - whether Named Entities and numbers in the graph also appear in the sentence
The parsers we used are
- AMRlib (trained on AMR 3.0 data)
- X-AMR (model provided on github site)
- Transition AMR parser (to different models available on their github site)
We trained models with this data using a modified version of AMRlib (using Flan-T5 instead of T5) and obtained better results than with a model trained on AMR 3.0:
| train dataset | AMR 3.0 test | QALD-9 test | DRS2AMR test |
|---|---|---|---|
| AMR~3.0 | 82.5 | 87.3 | 83.2 |
| Wiki-AMR | 82.8 (+0.3) | 89.2 (+1.9) | 84.2 (+1.0) |
| both (mixed) | 84.4 (+1.9) | 88.8 (+1.5) | 84.3 (+1.1) |
- The sentences are under the Wikipedia's licence (Attribution-ShareAlike 4.0 International, CC BY-SA 4.0)
- The AMR graphs are also under the CC BY-SA 4.0 licence
If you use this dataset, please cite our article:
@inproceedings{heinecke2026-et-al:wiki-amr,
author = {Heinecke, Johannes and Herledan, Frédéric and Asadullah, Munshi and Damnati, Géraldine},
booktitle = {Proceedings of LREC 2026},
title = {{Creating a high quality Abstract Meaning Representation dataset automatically}},
year = {2026},
address = {Palma de Mallorca, Spain}
}