Disclaimer

The English datasets for WSD have been previously collected by Raganato et al. EACL 2017 and can be found at http://lcl.uniroma1.it/wsdeval/. This repository contains only for the datasets in SemEval-2013 task 12 and SemEval-2015 task 13 for French, German, Italian and Spanish
The Multilingual datasets can be found at these two links:
- multilingual_wsd_all_v1.0.tar.gz
- multilingual_wsd_wn_v1.0.tar.gz
The code to create the inventories can be found at babelnet_mapping_4.0.jar

Multilingual Datasets for Word Sense Disambiguation

In this repository we release an updated version of SemEval-2013 task 12 and SemEval-2015 task 13 multilingual gold standards. The original data used old versions of BabelNet, i.e., 1.1.1 and 2.5, respectively, which are no longer available. To ease their use and allow a fair comparison among systems, we mapped all the possible gold keys in the original datasets to the latest available version of BabelNet (indices version 4.0 and API version 4.0.1), normalized the Part-of-Speech tags to the Universal POS tags and handled the multiword instances so that they are now associated with a single id and contained within a single XML tag. Furthermore, we provide two standard splits:

all: comprising all the instances that we managed to map,
wn: comprising only the instances tagged with a BabelNet synset which contains a sense in WordNet. This second split is a subset of all.

Datasets and Inventories Creation

To create the new datasets, whenever possible, we used the WordNet sense keys provided with the original data. Thanks to WordNet keys (version 3.0) we could easily map the old BabelNet synset id with the new synset id associated to that sense key in BabelNet 4.0. As for those instances that were not associated with a WordNet sense key, instead, we used the original BabelNet indices to retrieve the Wikipedia page title and the gloss associated with the target BabelNet synset (both in English and in the specific language) and searched them in the new indices. Whenever a match was found (prioritising titles over glosses and English over language-specific information ) we mapped the instance to the found BabelNet id. If no match was found we discarded the instance.

Build the Inventory

To correctly build the inventory, i.e., the association between a lexeme (lemma#pos) and its possible meanings, you shall follow this procedure.

Login or sign up to babelnet.org
Request and download BabelNet indices.
Download BabelNet API.
cd inventory_building
cp /path/to/BabelNet-API-[VERSION]/resources/* .
cp /path/to/BabelNet-API-[VERSION]/config .
Setup the properties files properly this step is essential, if not properly setup, you may end up having wrong inventories (More on this will follow).
bash create_inventories_sem13_15.sh -i /path/to/output_inventories/ -d /path/to/multilingual_datasets -s [wn|all]
Check out the inventories in /path/to/output_inventories/[lang] where lang may be one among [de, es, fr, it]

The multilingual_datasets can be extracted by running tar xvzf multilingual_wsd_*_v1.0.tar.gz

Example:

$ pwd
> /home/user/downloads/
$ cd inventory_building
$ cp -r /home/user/data/BabelNet-API-4.0.1/resources . 
$ cp -r /home/user/data/BabelNet-API-4.0.1/config . 
$ # setting up paths in configs
$ bash create_inventories_sem13_15.sh -i /home/user/downloads/inventory_building/data/multilingual_wsd_wn_v1.0/inventories/ -d /home/user/downloads/inventory_building/data/multilingual_wsd_wn_v1.0/ -s wn
$ ls /home/user/downloads/inventory_building/data/multilingual_wsd_wn_v1.0/inventories/
> de/ es/ fr/ it/
$ ls /home/user/downloads/inventory_building/data/multilingual_wsd_wn_v1.0/inventories/it
> inventory.it.withgold.sorted.txt  inventory.it.withgold.txt

Config Setup

In order to properly setup BabelNet config cd into config parent folder and verify that each path in babelnet.properties, babelnet.var.properties, jlt.properties and jlt.var.properties is reachable, as it is written in the files, from the directory you are in.

For example, if I am in /home/user/multilingual_wsd_wn_v1.0 which has config/ as subfolder. I need to be sure that any path written in the aforementioned files is reachable from /home/user/multilingual_wsd_wn_v1.0.

Most importantly, be sure to set:

babelnet.dir in babelnet.var.properties with the correct path towards BabelNet indices (e.g., BabelNet-4.0.1/ folder).
jlt.wordnetPrefix and jlt.wordnetVersion in jlt.var.properties file with the path to your WordNet installation and WordNet version (this has to be == 3.0).

BabelNet To WordNet mapping

In case you need to have a mapping between BabelNet offsets and WordNet offsets you can compute it by running

java -jar babelnet_mapping_4.0.jar -build_bn2wn_map /path/to/output/file.txt

This wil generate a tab-separated file having in the first column a BabelNet offset ID and in the subsequent columns the corresponding WordNet (version 3.0) offsets.

Data

Each package is structured as follows:

.
├── semeval2013-{lang}
│   ├── semeval2013-{lang}.data.xml
│   ├── semeval2013-{lang}.gold.key.txt
│   └── semeval2013-{lang}.wnids.gold.key.txt
└── semeval2015-{lang}
    ├── semeval2015-{lang}.data.xml
    ├── semeval2015-{lang}.gold.key.txt
    └── semeval2015-{lang}.wnids.gold.key.txt

Where, lang can be replaced by it|es|fr|de in semeval2013 and inventories folders, and by it|es in the semeval2015 folder.

SemEvals

The semeval folders contains the text to tag in form of xml (.data.xml files) and the gold synsets associated to the instances (.gold.keys.txt files). We note that, not all instances within the XML have a gold synset associated with within the gold.key.txt this is because the same XML can be used for both splits (all and wn) which may have different instances.

Statistics

WordNet Split

Dataset	Number of Instances	Word Types	Unique BN Synsets	Unique WN Synsets	Word Average Polysemy	Instance Average Polysemy	Polysemous Words
SemEval2013-it	1490	604	702	702	3.80	5.51	458
SemEval2013-es	1260	541	597	597	4.20	5.52	421
SemEval2013-fr	1449	612	655	655	2.36	3.03	370
SemEval2013-de	1076	484	481	481	1.60	2.17	184
SemEval2015-it	1007	531	688	688	4.38	5.27	420
SemEval2015-es	1043	507	733	733	6.17	6.99	446

ALL Split

Dataset	Number of Instances	Word Types	Unique BN Synsets	Word Average Polysemy	Instance Average Polysemy	Polysemous Words
SemEval2013-it	1665	731	825	4.63	6.46	541
SemEval2013-es	1463	678	730	4.85	6.36	484
SemEval2013-fr	1618	730	779	3.69	4.53	531
SemEval2013-de	1389	692	690	2.52	3.30	362
SemEval2015-it	1063	557	730	5.02	5.87	456
SemEval2015-es	1101	541	774	6.62	7.39	475

Contacts

For any question either open an issue on github or contact pasini[at]di[dot]uniroma1[dot]it

License

All data and codes provided in this repository are subject to the Attribution-Non Commercial-ShareAlike 4.0 International license (CC BY-NC 4.0).

Troubleshotting

If you get the following exception:

2020-07-17 19:27:52,042 [main] [ INFO  ] BabelNet - BabelNet online RESTful API v4.0 written by Francesco Cecconi, Roberto Navigli and Daniele Vannella
Exception in thread "main" java.lang.UnsupportedOperationException: getSynsetIterator: Unsupported online operation

Your babelnet.var.properties in the config/ folder is not properly set and you are attempting to use the online API rather then the indices you should have downloaded.

Solution: edit babelnet.var.properties file and set babelnet.dir to the local path to your BabelNet indeices.

suytingwan / mwsd-datasets